feat(mcp): add dynamic MCP server loading and unloading

- Add AddServer/RemoveServer to MCPToolManager for runtime server management - Add RemoveConnection to MCPConnectionPool for per-server teardown - Add AddMCPServer/RemoveMCPServer/ListMCPServers to Agent and SDK Kit - Lazily create connection pool so AddServer works without prior LoadTools - Wire onToolsChanged callback to trigger agent tool list rebuild - Make MCPToolManager.Close nil-safe when pool was never initialized Tests: - Integration tests with real stdio MCP server (Python echo server) - Agent-level tests using mock LLM model (no API key needed) - Unit tests for error paths, callbacks, idempotency, nil safety - SDK type surface tests
feat(sdk): add MCPTokenStoreFactory for custom OAuth token storage
2026-06-21 06:30:09 +00:00 · 2026-04-09 13:54:11 +03:00 · 2026-04-09 13:27:40 +03:00 · 2026-04-09 13:00:23 +03:00 · 2026-04-09 12:35:00 +03:00 · 2026-04-09 12:07:42 +03:00
21 changed files with 2236 additions and 116 deletions
@@ -30,6 +30,11 @@ type AgentConfig struct {
 	// If nil, remote MCP servers that require OAuth will fail to connect.
 	AuthHandler tools.MCPAuthHandler

+	// TokenStoreFactory, if non-nil, creates a custom token store for each
+	// remote MCP server's OAuth tokens. When nil, the default file-based
+	// token store is used.
+	TokenStoreFactory tools.TokenStoreFactory
+
 	// CoreTools overrides the default core tool set. If empty, core.AllTools()
 	// is used. This allows SDK users to provide a custom tool set (e.g.
 	// CodingTools or tools with a custom WorkDir).
@@ -236,6 +241,9 @@ func NewAgent(ctx context.Context, agentConfig *AgentConfig) (*Agent, error) {
 		if agentConfig.AuthHandler != nil {
 			toolManager.SetAuthHandler(agentConfig.AuthHandler)
 		}
+		if agentConfig.TokenStoreFactory != nil {
+			toolManager.SetTokenStoreFactory(agentConfig.TokenStoreFactory)
+		}
 		if agentConfig.DebugLogger != nil {
 			toolManager.SetDebugLogger(agentConfig.DebugLogger)
 		}
@@ -826,6 +834,59 @@ func (a *Agent) SetExtraTools(extraTools []fantasy.AgentTool) {
 	a.rebuildFantasyAgent()
 }

+// AddMCPServer connects to a new MCP server at runtime and makes its tools
+// available to the agent. Returns the number of tools loaded.
+// If the agent has no tool manager (no MCP servers were configured at init),
+// one is created automatically.
+func (a *Agent) AddMCPServer(ctx context.Context, name string, cfg config.MCPServerConfig) (int, error) {
+	// Ensure MCP tools from initial load are settled first.
+	a.ensureMCPTools()
+
+	if a.toolManager == nil {
+		a.toolManager = tools.NewMCPToolManager()
+		a.toolManager.SetModel(a.model)
+		a.toolManager.SetOnToolsChanged(func() {
+			a.rebuildFantasyAgent()
+		})
+	}
+
+	count, err := a.toolManager.AddServer(ctx, name, cfg)
+	if err != nil {
+		return 0, err
+	}
+
+	// AddServer's onToolsChanged callback triggers rebuildFantasyAgent,
+	// but only if it was wired. Ensure rebuild happens regardless.
+	a.rebuildFantasyAgent()
+	return count, nil
+}
+
+// RemoveMCPServer disconnects an MCP server and removes its tools from the agent.
+func (a *Agent) RemoveMCPServer(name string) error {
+	if a.toolManager == nil {
+		return fmt.Errorf("no MCP servers loaded")
+	}
+
+	// Ensure MCP tools from initial load are settled first.
+	a.ensureMCPTools()
+
+	err := a.toolManager.RemoveServer(name)
+	if err != nil {
+		return err
+	}
+
+	// RemoveServer's onToolsChanged callback triggers rebuildFantasyAgent,
+	// but ensure rebuild happens regardless.
+	a.rebuildFantasyAgent()
+	return nil
+}
+
+// GetMCPToolManager returns the underlying MCP tool manager.
+// Returns nil if no MCP servers have been configured.
+func (a *Agent) GetMCPToolManager() *tools.MCPToolManager {
+	return a.toolManager
+}
+
 // GetLoadingMessage returns the loading message from provider creation.
 func (a *Agent) GetLoadingMessage() string {
 	return a.loadingMessage
@@ -839,9 +900,11 @@ func (a *Agent) GetLoadedServerNames() []string {
 	return a.toolManager.GetLoadedServerNames()
 }

-// SetModel swaps the agent's LLM provider to a new model. The existing tools,
-// system prompt, and configuration are preserved. The old provider is closed
-// if it has a closer. Returns the previous model string for notification.
+// SetModel swaps the agent's LLM provider to a new model. The existing tools
+// and configuration are preserved. When the new model's ProviderConfig carries
+// a system prompt (from per-model settings), it replaces the agent's stored
+// prompt so the rebuilt fantasy agent uses it. The old provider is closed if
+// it has a closer.
 func (a *Agent) SetModel(ctx context.Context, config *models.ProviderConfig) error {
 	// Ensure MCP tools are loaded before rebuilding (SetModel may be called
 	// before the first LLM call).
@@ -868,6 +931,13 @@ func (a *Agent) SetModel(ctx context.Context, config *models.ProviderConfig) err
 	a.skipMaxOutputTokens = providerResult.SkipMaxOutputTokens
 	a.modelConfig = config

+	// Update system prompt when the config carries one (from per-model
+	// settings or the global config). This allows model-specific system
+	// prompts to take effect on model switch.
+	if config.SystemPrompt != "" {
+		a.systemPrompt = config.SystemPrompt
+	}
+
 	// Update provider type.
 	if config.ModelString != "" {
 		if p, _, err := models.ParseModelString(config.ModelString); err == nil {
@@ -0,0 +1,242 @@
+package agent
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+
+	"charm.land/fantasy"
+
+	"github.com/mark3labs/kit/internal/config"
+)
+
+// mockModel is a minimal LanguageModel that satisfies the interface
+// without making real API calls. Used to test tool management wiring.
+type mockModel struct{}
+
+func (m *mockModel) Generate(_ context.Context, _ fantasy.Call) (*fantasy.Response, error) {
+	return &fantasy.Response{}, nil
+}
+func (m *mockModel) Stream(_ context.Context, _ fantasy.Call) (fantasy.StreamResponse, error) {
+	return nil, nil
+}
+func (m *mockModel) GenerateObject(_ context.Context, _ fantasy.ObjectCall) (*fantasy.ObjectResponse, error) {
+	return &fantasy.ObjectResponse{}, nil
+}
+func (m *mockModel) StreamObject(_ context.Context, _ fantasy.ObjectCall) (fantasy.ObjectStreamResponse, error) {
+	return nil, nil
+}
+func (m *mockModel) Provider() string { return "mock" }
+func (m *mockModel) Model() string    { return "mock-model" }
+
+// testdataDir returns the absolute path to the tools testdata directory.
+func testdataDir(t *testing.T) string {
+	t.Helper()
+	_, file, _, ok := runtime.Caller(0)
+	if !ok {
+		t.Fatal("cannot determine test file path")
+	}
+	return filepath.Join(filepath.Dir(file), "..", "tools", "testdata")
+}
+
+// echoServerConfig returns an MCPServerConfig for the test echo MCP server.
+func echoServerConfig(t *testing.T) config.MCPServerConfig {
+	t.Helper()
+	script := filepath.Join(testdataDir(t), "echo_server.py")
+	if _, err := os.Stat(script); err != nil {
+		t.Skipf("echo_server.py not found: %v", err)
+	}
+	return config.MCPServerConfig{
+		Command: []string{"python3", script},
+	}
+}
+
+// newTestAgent creates a minimal Agent with a mock model and no core tools,
+// suitable for testing MCP server management without an API key.
+func newTestAgent() *Agent {
+	model := &mockModel{}
+	a := &Agent{
+		model:        model,
+		coreTools:    nil,
+		extraTools:   nil,
+		maxSteps:     10,
+		systemPrompt: "test",
+		fantasyAgent: fantasy.NewAgent(model),
+	}
+	return a
+}
+
+func TestAgent_AddMCPServer(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	a := newTestAgent()
+	defer func() { _ = a.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Initially no MCP tools.
+	if a.GetMCPToolCount() != 0 {
+		t.Fatalf("Expected 0 MCP tools initially, got %d", a.GetMCPToolCount())
+	}
+
+	// Add a server.
+	count, err := a.AddMCPServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("AddMCPServer failed: %v", err)
+	}
+	if count != 2 {
+		t.Errorf("Expected 2 tools, got %d", count)
+	}
+
+	// Verify tools are in the agent's tool list.
+	if a.GetMCPToolCount() != 2 {
+		t.Errorf("Expected 2 MCP tools, got %d", a.GetMCPToolCount())
+	}
+
+	allTools := a.GetTools()
+	toolNames := make(map[string]bool)
+	for _, tool := range allTools {
+		toolNames[tool.Info().Name] = true
+	}
+	if !toolNames["echo__echo"] {
+		t.Error("Expected tool 'echo__echo' in agent tools")
+	}
+	if !toolNames["echo__greet"] {
+		t.Error("Expected tool 'echo__greet' in agent tools")
+	}
+
+	// Verify loaded server names.
+	names := a.GetLoadedServerNames()
+	found := false
+	for _, n := range names {
+		if n == "echo" {
+			found = true
+		}
+	}
+	if !found {
+		t.Errorf("Expected 'echo' in loaded server names: %v", names)
+	}
+}
+
+func TestAgent_RemoveMCPServer(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	a := newTestAgent()
+	defer func() { _ = a.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add then remove.
+	_, err := a.AddMCPServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("AddMCPServer failed: %v", err)
+	}
+
+	err = a.RemoveMCPServer("echo")
+	if err != nil {
+		t.Fatalf("RemoveMCPServer failed: %v", err)
+	}
+
+	// Verify tools removed.
+	if a.GetMCPToolCount() != 0 {
+		t.Errorf("Expected 0 MCP tools after removal, got %d", a.GetMCPToolCount())
+	}
+
+	// Verify agent's tool list has no MCP tools.
+	for _, tool := range a.GetTools() {
+		if strings.Contains(tool.Info().Name, "echo__") {
+			t.Errorf("Found leftover tool after removal: %s", tool.Info().Name)
+		}
+	}
+}
+
+func TestAgent_RemoveMCPServer_NoToolManager(t *testing.T) {
+	a := newTestAgent()
+	defer func() { _ = a.Close() }()
+
+	err := a.RemoveMCPServer("nonexistent")
+	if err == nil {
+		t.Fatal("Expected error when no tool manager exists")
+	}
+	if !strings.Contains(err.Error(), "no MCP servers loaded") {
+		t.Errorf("Expected 'no MCP servers loaded' error, got: %v", err)
+	}
+}
+
+func TestAgent_AddMCPServer_CreatesToolManager(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	a := newTestAgent()
+	defer func() { _ = a.Close() }()
+
+	// Initially no tool manager.
+	if a.GetMCPToolManager() != nil {
+		t.Fatal("Expected nil tool manager initially")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+	_, err := a.AddMCPServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("AddMCPServer failed: %v", err)
+	}
+
+	// Tool manager should now exist.
+	if a.GetMCPToolManager() == nil {
+		t.Fatal("Expected tool manager to be created by AddMCPServer")
+	}
+}
+
+func TestAgent_AddRemoveAdd_MCP(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	a := newTestAgent()
+	defer func() { _ = a.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add → Remove → Add cycle.
+	_, err := a.AddMCPServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("First add failed: %v", err)
+	}
+
+	err = a.RemoveMCPServer("echo")
+	if err != nil {
+		t.Fatalf("Remove failed: %v", err)
+	}
+
+	count, err := a.AddMCPServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("Re-add failed: %v", err)
+	}
+	if count != 2 {
+		t.Errorf("Expected 2 tools on re-add, got %d", count)
+	}
+	if a.GetMCPToolCount() != 2 {
+		t.Errorf("Expected 2 MCP tools after re-add, got %d", a.GetMCPToolCount())
+	}
+}
@@ -38,6 +38,10 @@ type AgentCreationOptions struct {
 	DebugLogger tools.DebugLogger // Optional debug logger
 	// AuthHandler handles OAuth authorization for remote MCP servers
 	AuthHandler tools.MCPAuthHandler
+	// TokenStoreFactory, if non-nil, creates a custom token store for each
+	// remote MCP server's OAuth tokens. When nil, the default file-based
+	// token store is used.
+	TokenStoreFactory tools.TokenStoreFactory
 	// CoreTools overrides the default core tool set. If empty, core.AllTools()
 	// is used.
 	CoreTools []fantasy.AgentTool
@@ -66,6 +70,7 @@ func CreateAgent(ctx context.Context, opts *AgentCreationOptions) (*Agent, error
 		StreamingEnabled:  opts.StreamingEnabled,
 		DebugLogger:       opts.DebugLogger,
 		AuthHandler:       opts.AuthHandler,
+		TokenStoreFactory: opts.TokenStoreFactory,
 		CoreTools:         opts.CoreTools,
 		DisableCoreTools:  opts.DisableCoreTools,
 		ToolWrapper:       opts.ToolWrapper,
@@ -930,7 +930,8 @@ func (a *App) QuitFromExtension() {
 // controls styling: "" for plain text, "info" for a system message block,
 // "error" for an error block. In interactive mode it sends an
 // ExtensionPrintEvent through the program so the TUI can render it with the
-// appropriate renderer. In non-interactive mode it falls back to stdout.
+// appropriate renderer. In non-interactive mode it falls back to stderr with
+// a level prefix so errors are distinguishable from plain output.
 func (a *App) PrintFromExtension(level, text string) {
 	a.mu.Lock()
 	prog := a.program
@@ -939,8 +940,16 @@ func (a *App) PrintFromExtension(level, text string) {
 		prog.Send(ExtensionPrintEvent{Text: text, Level: level})
 		return
 	}
-	// Non-interactive fallback: write directly to stdout.
-	fmt.Println(text)
+	// Non-interactive fallback: write to stderr with a level prefix so that
+	// errors and info messages are distinguishable from plain output.
+	switch level {
+	case "error":
+		fmt.Fprintf(os.Stderr, "[ERROR] %s\n", text)
+	case "info":
+		fmt.Fprintf(os.Stderr, "[INFO] %s\n", text)
+	default:
+		fmt.Println(text)
+	}
 }

 // SetEditorTextFromExtension sends an EditorTextSetEvent to the TUI to
@@ -1122,11 +1131,12 @@ func (a *App) PrintBlockFromExtension(opts extensions.PrintBlockOpts) {
 		})
 		return
 	}
-	// Non-interactive fallback.
+	// Non-interactive fallback: render a simple framed block to stderr so
+	// it is visually distinct from plain stdout output.
 	if opts.Subtitle != "" {
-		fmt.Printf("%s\n  — %s\n", opts.Text, opts.Subtitle)
+		fmt.Fprintf(os.Stderr, "--- %s ---\n%s\n", opts.Subtitle, opts.Text)
 	} else {
-		fmt.Println(opts.Text)
+		fmt.Fprintf(os.Stderr, "---\n%s\n---\n", opts.Text)
 	}
 }

@@ -157,6 +157,21 @@ type Theme struct {
 	Markdown MarkdownThemeConfig `json:"markdown,omitzero" yaml:"markdown,omitempty"`
 }

+// GenerationParams defines generation parameter defaults that can be attached
+// to individual models. These act as model-level defaults — CLI flags and
+// global config values take precedence when explicitly set.
+type GenerationParams struct {
+	MaxTokens        *int     `json:"maxTokens,omitempty" yaml:"maxTokens,omitempty"`
+	Temperature      *float32 `json:"temperature,omitempty" yaml:"temperature,omitempty"`
+	TopP             *float32 `json:"topP,omitempty" yaml:"topP,omitempty"`
+	TopK             *int32   `json:"topK,omitempty" yaml:"topK,omitempty"`
+	FrequencyPenalty *float32 `json:"frequencyPenalty,omitempty" yaml:"frequencyPenalty,omitempty"`
+	PresencePenalty  *float32 `json:"presencePenalty,omitempty" yaml:"presencePenalty,omitempty"`
+	StopSequences    []string `json:"stopSequences,omitempty" yaml:"stopSequences,omitempty"`
+	ThinkingLevel    string   `json:"thinkingLevel,omitempty" yaml:"thinkingLevel,omitempty"`
+	SystemPrompt     string   `json:"systemPrompt,omitempty" yaml:"systemPrompt,omitempty"`
+}
+
 // CustomModelConfig defines a custom model that can be used with custom/custom
 // or other custom/ prefixed models. These models are loaded from the config file
 // and merged into the custom provider in the model registry.
@@ -171,6 +186,11 @@ type CustomModelConfig struct {
 	Knowledge   string      `json:"knowledge,omitempty" yaml:"knowledge,omitempty"`
 	Cost        CostConfig  `json:"cost" yaml:"cost"`
 	Limit       LimitConfig `json:"limit" yaml:"limit"`
+
+	// Generation parameter defaults for this model.
+	// These are applied when the user hasn't explicitly set the corresponding
+	// CLI flag or global config value.
+	Params GenerationParams `json:"params,omitzero" yaml:"params,omitempty"`
 }

 // CostConfig defines the pricing for a custom model.
@@ -219,6 +239,12 @@ type Config struct {

 	// Custom model definitions (under custom/ provider)
 	CustomModels map[string]CustomModelConfig `json:"customModels,omitempty" yaml:"customModels,omitempty"`
+
+	// Per-model generation parameter overrides. Keys are "provider/model" strings
+	// (e.g. "anthropic/claude-sonnet-4-5-20250929", "openai/gpt-4o"). These
+	// settings act as model-level defaults — CLI flags and global config values
+	// take precedence when explicitly set.
+	ModelSettings map[string]GenerationParams `json:"modelSettings,omitempty" yaml:"modelSettings,omitempty"`
 }

 // GetTransportType returns the transport type for the server config, mapping
@@ -367,7 +393,7 @@ mcpServers:
 # debug: false                                 # Enable debug logging
 # system-prompt: "/path/to/system-prompt.txt" # System prompt text file

-# Model generation parameters (all optional)
+# Model generation parameters (all optional, apply globally to all models)
 # max-tokens: 4096                             # Maximum tokens in response
 # temperature: 0.7                             # Randomness (0.0-1.0)
 # top-p: 0.95                                  # Nucleus sampling (0.0-1.0)
@@ -376,9 +402,46 @@ mcpServers:
 # presence-penalty: 0.0                         # Penalize present tokens (0.0-2.0)
 # stop-sequences: ["Human:", "Assistant:"]     # Custom stop sequences

+# Per-model generation parameter overrides (apply to specific models)
+# These act as model-level defaults — CLI flags and global settings above take precedence.
+# Keys are "provider/model" strings matching the model you use.
+# modelSettings:
+#   anthropic/claude-sonnet-4-5-20250929:
+#     temperature: 0.3
+#     maxTokens: 8192
+#   openai/gpt-4o:
+#     temperature: 0.7
+#     topP: 0.95
+#     topK: 40
+#     frequencyPenalty: 0.1
+#     presencePenalty: 0.1
+#   anthropic/claude-opus-4-6:
+#     thinkingLevel: "high"
+#     maxTokens: 16384
+#     systemPrompt: "You are a deep reasoning assistant."  # or a file path
+
 # API Configuration (can also use environment variables)
 # provider-api-key: "your-api-key"         # API key for OpenAI, Anthropic, or Google
 # provider-url: "https://api.openai.com/v1" # Base URL for OpenAI, Anthropic, or Ollama
+
+# Custom model definitions (under custom/ provider)
+# customModels:
+#   my-local-llama:
+#     name: "Local Llama 3"
+#     baseUrl: "http://localhost:8080/v1"
+#     family: "llama"
+#     temperature: true
+#     cost:
+#       input: 0.0
+#       output: 0.0
+#     limit:
+#       context: 131072
+#       output: 8192
+#     params:                              # Generation parameter defaults for this model
+#       temperature: 0.8
+#       topP: 0.95
+#       topK: 40
+#       systemPrompt: "You are a helpful local assistant."
 `

 	_, err = file.WriteString(content)
@@ -65,6 +65,10 @@ type AgentSetupOptions struct {
 	// AuthHandler handles OAuth authorization for remote MCP servers.
 	// When set, remote transports are configured with OAuth support.
 	AuthHandler tools.MCPAuthHandler
+	// TokenStoreFactory, if non-nil, creates a custom token store for each
+	// remote MCP server's OAuth tokens. When nil, the default file-based
+	// token store is used.
+	TokenStoreFactory tools.TokenStoreFactory
 	// OnMCPServerLoaded, if non-nil, is called when each MCP server finishes
 	// loading (successfully or with error). Called from the background goroutine.
 	OnMCPServerLoaded func(serverName string, toolCount int, err error)
@@ -82,36 +86,55 @@ type AgentSetupResult struct {

 // BuildProviderConfig creates a *models.ProviderConfig from the current viper
 // state. All entry points (root, script, SDK) converge through this function.
+//
+// Generation parameter pointers (Temperature, TopP, etc.) are only set when
+// the user has explicitly configured them via CLI flag, environment variable,
+// or global config file. This allows per-model defaults from modelSettings
+// and customModels to fill in unset parameters downstream.
 func BuildProviderConfig() (*models.ProviderConfig, string, error) {
 	systemPrompt, err := config.LoadSystemPrompt(viper.GetString("system-prompt"))
 	if err != nil {
 		return nil, "", fmt.Errorf("failed to load system prompt: %w", err)
 	}

-	temperature := float32(viper.GetFloat64("temperature"))
-	topP := float32(viper.GetFloat64("top-p"))
-	topK := int32(viper.GetInt("top-k"))
-	frequencyPenalty := float32(viper.GetFloat64("frequency-penalty"))
-	presencePenalty := float32(viper.GetFloat64("presence-penalty"))
 	numGPU := int32(viper.GetInt("num-gpu-layers"))
 	mainGPU := int32(viper.GetInt("main-gpu"))

 	cfg := &models.ProviderConfig{
-		ModelString:      viper.GetString("model"),
-		SystemPrompt:     systemPrompt,
-		ProviderAPIKey:   viper.GetString("provider-api-key"),
-		ProviderURL:      viper.GetString("provider-url"),
-		MaxTokens:        viper.GetInt("max-tokens"),
-		Temperature:      &temperature,
-		TopP:             &topP,
-		TopK:             &topK,
-		FrequencyPenalty: &frequencyPenalty,
-		PresencePenalty:  &presencePenalty,
-		StopSequences:    viper.GetStringSlice("stop-sequences"),
-		NumGPU:           &numGPU,
-		MainGPU:          &mainGPU,
-		TLSSkipVerify:    viper.GetBool("tls-skip-verify"),
-		ThinkingLevel:    models.ParseThinkingLevel(viper.GetString("thinking-level")),
+		ModelString:    viper.GetString("model"),
+		SystemPrompt:   systemPrompt,
+		ProviderAPIKey: viper.GetString("provider-api-key"),
+		ProviderURL:    viper.GetString("provider-url"),
+		MaxTokens:      viper.GetInt("max-tokens"),
+		StopSequences:  viper.GetStringSlice("stop-sequences"),
+		NumGPU:         &numGPU,
+		MainGPU:        &mainGPU,
+		TLSSkipVerify:  viper.GetBool("tls-skip-verify"),
+		ThinkingLevel:  models.ParseThinkingLevel(viper.GetString("thinking-level")),
+	}
+
+	// Only set generation parameter pointers when the user has explicitly
+	// provided a value. This leaves nil pointers for unset params, allowing
+	// per-model defaults (modelSettings / customModels params) to apply.
+	if viper.IsSet("temperature") {
+		v := float32(viper.GetFloat64("temperature"))
+		cfg.Temperature = &v
+	}
+	if viper.IsSet("top-p") {
+		v := float32(viper.GetFloat64("top-p"))
+		cfg.TopP = &v
+	}
+	if viper.IsSet("top-k") {
+		v := int32(viper.GetInt("top-k"))
+		cfg.TopK = &v
+	}
+	if viper.IsSet("frequency-penalty") {
+		v := float32(viper.GetFloat64("frequency-penalty"))
+		cfg.FrequencyPenalty = &v
+	}
+	if viper.IsSet("presence-penalty") {
+		v := float32(viper.GetFloat64("presence-penalty"))
+		cfg.PresencePenalty = &v
 	}

 	return cfg, systemPrompt, nil
@@ -200,6 +223,7 @@ func SetupAgent(ctx context.Context, opts AgentSetupOptions) (*AgentSetupResult,
 		SpinnerFunc:       opts.SpinnerFunc,
 		DebugLogger:       debugLogger,
 		AuthHandler:       opts.AuthHandler,
+		TokenStoreFactory: opts.TokenStoreFactory,
 		CoreTools:         opts.CoreTools,
 		DisableCoreTools:  opts.DisableCoreTools,
 		ToolWrapper:       toolWrapper,
@@ -2,6 +2,8 @@ package models

 import (
 	"log"
+	"os"
+	"strings"

 	"github.com/spf13/viper"
 )
@@ -31,7 +33,7 @@ func loadCustomModelsFromConfig() map[string]ModelInfo {

 // modelConfigToModelInfo converts a CustomModelConfig to a ModelInfo.
 func modelConfigToModelInfo(modelID string, cfg CustomModelConfig) ModelInfo {
-	return ModelInfo{
+	info := ModelInfo{
 		ID:          modelID,
 		Name:        cfg.Name,
 		Attachment:  cfg.Attachment,
@@ -48,21 +50,242 @@ func modelConfigToModelInfo(modelID string, cfg CustomModelConfig) ModelInfo {
 			Output:  cfg.Limit.Output,
 		},
 	}
+
+	// Convert custom model generation params if any are set.
+	if p := convertGenerationParams(cfg.Params); p != nil {
+		info.Params = p
+	}
+
+	return info
+}
+
+// LoadModelSettingsFromConfig loads per-model generation parameter overrides
+// from the config file. Keys are "provider/model" strings. Returns nil if
+// no model settings are configured.
+func LoadModelSettingsFromConfig() map[string]*GenerationParams {
+	if !viper.IsSet("modelSettings") {
+		return nil
+	}
+
+	var settings map[string]GenerationParamsConfig
+	if err := viper.UnmarshalKey("modelSettings", &settings); err != nil {
+		log.Printf("Warning: Failed to parse modelSettings: %v", err)
+		return nil
+	}
+
+	result := make(map[string]*GenerationParams, len(settings))
+	for modelKey, cfg := range settings {
+		if p := convertGenerationParams(cfg); p != nil {
+			result[modelKey] = p
+		}
+	}
+
+	return result
+}
+
+// convertGenerationParams converts a GenerationParamsConfig to a GenerationParams.
+// Returns nil if no parameters are set.
+func convertGenerationParams(cfg GenerationParamsConfig) *GenerationParams {
+	p := &GenerationParams{}
+	any := false
+
+	if cfg.MaxTokens != nil {
+		p.MaxTokens = cfg.MaxTokens
+		any = true
+	}
+	if cfg.Temperature != nil {
+		p.Temperature = cfg.Temperature
+		any = true
+	}
+	if cfg.TopP != nil {
+		p.TopP = cfg.TopP
+		any = true
+	}
+	if cfg.TopK != nil {
+		p.TopK = cfg.TopK
+		any = true
+	}
+	if cfg.FrequencyPenalty != nil {
+		p.FrequencyPenalty = cfg.FrequencyPenalty
+		any = true
+	}
+	if cfg.PresencePenalty != nil {
+		p.PresencePenalty = cfg.PresencePenalty
+		any = true
+	}
+	if len(cfg.StopSequences) > 0 {
+		p.StopSequences = cfg.StopSequences
+		any = true
+	}
+	if cfg.ThinkingLevel != "" {
+		p.ThinkingLevel = ParseThinkingLevel(cfg.ThinkingLevel)
+		any = true
+	}
+	if cfg.SystemPrompt != "" {
+		p.SystemPrompt = cfg.SystemPrompt
+		any = true
+	}
+
+	if !any {
+		return nil
+	}
+	return p
+}
+
+// ApplyModelSettings merges per-model generation parameter defaults from the
+// registry into a ProviderConfig. Model-level params are only applied for
+// fields where the user has not explicitly set a value (i.e., the
+// corresponding viper key is not set via CLI flag or global config).
+//
+// The lookup order is:
+//  1. modelSettings["provider/model"] from config (highest model-level priority)
+//  2. ModelInfo.Params from custom model definitions
+//
+// Both are overridden by explicit CLI flags / global config values.
+func ApplyModelSettings(config *ProviderConfig, modelInfo *ModelInfo) {
+	provider, modelName, err := ParseModelString(config.ModelString)
+	if err != nil {
+		return
+	}
+
+	// Collect model-level params: modelSettings override > custom model params.
+	// modelSettings takes priority because it's the more specific/intentional config.
+	var params *GenerationParams
+
+	// First check modelSettings from config.
+	if settings := LoadModelSettingsFromConfig(); settings != nil {
+		modelKey := provider + "/" + modelName
+		if p, ok := settings[modelKey]; ok {
+			params = p
+		}
+	}
+
+	// Fall back to ModelInfo.Params (from custom model definitions).
+	if params == nil && modelInfo != nil && modelInfo.Params != nil {
+		params = modelInfo.Params
+	}
+
+	if params == nil {
+		return
+	}
+
+	// Apply each parameter only when the user hasn't explicitly set it.
+	// We check viper.IsSet() which returns true only when the key was
+	// set via CLI flag, environment variable, or config file global section.
+
+	if params.MaxTokens != nil && !isExplicitlySet("max-tokens") {
+		config.MaxTokens = *params.MaxTokens
+	}
+	if params.Temperature != nil && !isExplicitlySet("temperature") {
+		config.Temperature = params.Temperature
+	}
+	if params.TopP != nil && !isExplicitlySet("top-p") {
+		config.TopP = params.TopP
+	}
+	if params.TopK != nil && !isExplicitlySet("top-k") {
+		config.TopK = params.TopK
+	}
+	if params.FrequencyPenalty != nil && !isExplicitlySet("frequency-penalty") {
+		config.FrequencyPenalty = params.FrequencyPenalty
+	}
+	if params.PresencePenalty != nil && !isExplicitlySet("presence-penalty") {
+		config.PresencePenalty = params.PresencePenalty
+	}
+	if len(params.StopSequences) > 0 && !isExplicitlySet("stop-sequences") {
+		config.StopSequences = params.StopSequences
+	}
+	if params.ThinkingLevel != "" && !isExplicitlySet("thinking-level") {
+		config.ThinkingLevel = params.ThinkingLevel
+	}
+	if params.SystemPrompt != "" && config.SystemPrompt == "" {
+		// Resolve file paths: if the value points to an existing file, read it.
+		// We check config.SystemPrompt == "" rather than isExplicitlySet because
+		// viper.BindPFlag causes IsSet to return true even for unset flags.
+		config.SystemPrompt = LoadSystemPromptValue(params.SystemPrompt)
+	}
+}
+
+// LoadSystemPromptValue resolves a system prompt value that may be either
+// inline text or a file path. If the value is a path to an existing file,
+// its contents are read and returned. Otherwise the string is returned as-is.
+// This mirrors config.LoadSystemPrompt but lives in the models package to
+// avoid circular dependencies.
+func LoadSystemPromptValue(input string) string {
+	if input == "" {
+		return ""
+	}
+	if info, err := os.Stat(input); err == nil && !info.IsDir() {
+		content, err := os.ReadFile(input)
+		if err != nil {
+			log.Printf("Warning: failed to read system prompt file %q: %v", input, err)
+			return input
+		}
+		return strings.TrimSpace(string(content))
+	}
+	return input
+}
+
+// isExplicitlySet returns true when the user has explicitly set a config key
+// via CLI flag, environment variable, or the global section of the config file.
+// Model-level defaults should not override explicitly set values.
+func isExplicitlySet(key string) bool {
+	// viper.IsSet returns true if the key has been set in any of the
+	// data stores (flag, env, config file, default). We need to check
+	// whether the value was set at the global config level (not just
+	// as a default). For generation params, the global config keys use
+	// hyphenated names (e.g. "max-tokens", "top-p").
+	//
+	// Since viper merges all sources, IsSet returns true even for config
+	// file values. This means global config file values (e.g.
+	// temperature: 0.7 at the top level) will correctly take precedence
+	// over model-level defaults, which is the desired behavior.
+	return viper.IsSet(key)
+}
+
+// GenerationParams holds per-model generation parameter defaults.
+// These are stored on ModelInfo and applied during provider creation.
+// Nil pointer fields mean "no model-level default" — the global config
+// or CLI flag value (if any) will be used instead.
+type GenerationParams struct {
+	MaxTokens        *int
+	Temperature      *float32
+	TopP             *float32
+	TopK             *int32
+	FrequencyPenalty *float32
+	PresencePenalty  *float32
+	StopSequences    []string
+	ThinkingLevel    ThinkingLevel
+	SystemPrompt     string // Per-model system prompt (inline text or file path)
 }

 // CustomModelConfig defines a custom model configuration loaded from the config file.
 // This is a duplicate here to avoid circular dependencies with internal/config.
 type CustomModelConfig struct {
-	Name        string      `json:"name" yaml:"name"`
-	BaseURL     string      `json:"baseUrl,omitempty" yaml:"baseUrl,omitempty"`
-	APIKey      string      `json:"apiKey,omitempty" yaml:"apiKey,omitempty"`
-	Family      string      `json:"family,omitempty" yaml:"family,omitempty"`
-	Attachment  bool        `json:"attachment,omitempty" yaml:"attachment,omitempty"`
-	Reasoning   bool        `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
-	Temperature bool        `json:"temperature,omitempty" yaml:"temperature,omitempty"`
-	Knowledge   string      `json:"knowledge,omitempty" yaml:"knowledge,omitempty"`
-	Cost        CostConfig  `json:"cost" yaml:"cost"`
-	Limit       LimitConfig `json:"limit" yaml:"limit"`
+	Name        string                 `json:"name" yaml:"name"`
+	BaseURL     string                 `json:"baseUrl,omitempty" yaml:"baseUrl,omitempty"`
+	APIKey      string                 `json:"apiKey,omitempty" yaml:"apiKey,omitempty"`
+	Family      string                 `json:"family,omitempty" yaml:"family,omitempty"`
+	Attachment  bool                   `json:"attachment,omitempty" yaml:"attachment,omitempty"`
+	Reasoning   bool                   `json:"reasoning,omitempty" yaml:"reasoning,omitempty"`
+	Temperature bool                   `json:"temperature,omitempty" yaml:"temperature,omitempty"`
+	Knowledge   string                 `json:"knowledge,omitempty" yaml:"knowledge,omitempty"`
+	Cost        CostConfig             `json:"cost" yaml:"cost"`
+	Limit       LimitConfig            `json:"limit" yaml:"limit"`
+	Params      GenerationParamsConfig `json:"params,omitzero" yaml:"params,omitempty"`
+}
+
+// GenerationParamsConfig is the JSON/YAML-serializable form of generation
+// parameter defaults. Used in both customModels[].params and modelSettings[].
+type GenerationParamsConfig struct {
+	MaxTokens        *int     `json:"maxTokens,omitempty" yaml:"maxTokens,omitempty"`
+	Temperature      *float32 `json:"temperature,omitempty" yaml:"temperature,omitempty"`
+	TopP             *float32 `json:"topP,omitempty" yaml:"topP,omitempty"`
+	TopK             *int32   `json:"topK,omitempty" yaml:"topK,omitempty"`
+	FrequencyPenalty *float32 `json:"frequencyPenalty,omitempty" yaml:"frequencyPenalty,omitempty"`
+	PresencePenalty  *float32 `json:"presencePenalty,omitempty" yaml:"presencePenalty,omitempty"`
+	StopSequences    []string `json:"stopSequences,omitempty" yaml:"stopSequences,omitempty"`
+	ThinkingLevel    string   `json:"thinkingLevel,omitempty" yaml:"thinkingLevel,omitempty"`
+	SystemPrompt     string   `json:"systemPrompt,omitempty" yaml:"systemPrompt,omitempty"`
 }

 // CostConfig defines the pricing for a custom model.
@@ -0,0 +1,422 @@
+package models
+
+import (
+	"os"
+	"testing"
+
+	"github.com/spf13/viper"
+)
+
+func TestConvertGenerationParams(t *testing.T) {
+	t.Run("empty config returns nil", func(t *testing.T) {
+		cfg := GenerationParamsConfig{}
+		p := convertGenerationParams(cfg)
+		if p != nil {
+			t.Errorf("expected nil, got %+v", p)
+		}
+	})
+
+	t.Run("temperature only", func(t *testing.T) {
+		temp := float32(0.7)
+		cfg := GenerationParamsConfig{Temperature: &temp}
+		p := convertGenerationParams(cfg)
+		if p == nil {
+			t.Fatal("expected non-nil")
+		}
+		if p.Temperature == nil || *p.Temperature != 0.7 {
+			t.Errorf("expected temperature 0.7, got %v", p.Temperature)
+		}
+		if p.TopP != nil {
+			t.Errorf("expected nil TopP, got %v", p.TopP)
+		}
+	})
+
+	t.Run("all params set", func(t *testing.T) {
+		maxTokens := 8192
+		temp := float32(0.5)
+		topP := float32(0.9)
+		topK := int32(50)
+		freqPenalty := float32(0.1)
+		presPenalty := float32(0.2)
+		cfg := GenerationParamsConfig{
+			MaxTokens:        &maxTokens,
+			Temperature:      &temp,
+			TopP:             &topP,
+			TopK:             &topK,
+			FrequencyPenalty: &freqPenalty,
+			PresencePenalty:  &presPenalty,
+			StopSequences:    []string{"STOP"},
+			ThinkingLevel:    "high",
+		}
+		p := convertGenerationParams(cfg)
+		if p == nil {
+			t.Fatal("expected non-nil")
+		}
+		if p.MaxTokens == nil || *p.MaxTokens != 8192 {
+			t.Errorf("expected maxTokens 8192, got %v", p.MaxTokens)
+		}
+		if p.Temperature == nil || *p.Temperature != 0.5 {
+			t.Errorf("expected temperature 0.5, got %v", p.Temperature)
+		}
+		if p.TopP == nil || *p.TopP != 0.9 {
+			t.Errorf("expected topP 0.9, got %v", p.TopP)
+		}
+		if p.TopK == nil || *p.TopK != 50 {
+			t.Errorf("expected topK 50, got %v", p.TopK)
+		}
+		if p.FrequencyPenalty == nil || *p.FrequencyPenalty != 0.1 {
+			t.Errorf("expected frequencyPenalty 0.1, got %v", p.FrequencyPenalty)
+		}
+		if p.PresencePenalty == nil || *p.PresencePenalty != 0.2 {
+			t.Errorf("expected presencePenalty 0.2, got %v", p.PresencePenalty)
+		}
+		if len(p.StopSequences) != 1 || p.StopSequences[0] != "STOP" {
+			t.Errorf("expected stop sequences [STOP], got %v", p.StopSequences)
+		}
+		if p.ThinkingLevel != ThinkingHigh {
+			t.Errorf("expected thinking level high, got %v", p.ThinkingLevel)
+		}
+	})
+
+	t.Run("thinking level parsing", func(t *testing.T) {
+		cfg := GenerationParamsConfig{ThinkingLevel: "medium"}
+		p := convertGenerationParams(cfg)
+		if p == nil {
+			t.Fatal("expected non-nil")
+		}
+		if p.ThinkingLevel != ThinkingMedium {
+			t.Errorf("expected thinking level medium, got %v", p.ThinkingLevel)
+		}
+	})
+	t.Run("system prompt only", func(t *testing.T) {
+		cfg := GenerationParamsConfig{SystemPrompt: "You are helpful."}
+		p := convertGenerationParams(cfg)
+		if p == nil {
+			t.Fatal("expected non-nil")
+		}
+		if p.SystemPrompt != "You are helpful." {
+			t.Errorf("expected system prompt, got %q", p.SystemPrompt)
+		}
+	})
+}
+
+func TestModelConfigToModelInfoWithParams(t *testing.T) {
+	temp := float32(0.8)
+	topP := float32(0.95)
+	cfg := CustomModelConfig{
+		Name:        "Test Model",
+		BaseURL:     "http://localhost:8080/v1",
+		Temperature: true,
+		Params: GenerationParamsConfig{
+			Temperature: &temp,
+			TopP:        &topP,
+		},
+	}
+
+	info := modelConfigToModelInfo("test-model", cfg)
+
+	if info.Params == nil {
+		t.Fatal("expected non-nil Params")
+	}
+	if info.Params.Temperature == nil || *info.Params.Temperature != 0.8 {
+		t.Errorf("expected temperature 0.8, got %v", info.Params.Temperature)
+	}
+	if info.Params.TopP == nil || *info.Params.TopP != 0.95 {
+		t.Errorf("expected topP 0.95, got %v", info.Params.TopP)
+	}
+}
+
+func TestModelConfigToModelInfoWithoutParams(t *testing.T) {
+	cfg := CustomModelConfig{
+		Name:    "Test Model",
+		BaseURL: "http://localhost:8080/v1",
+	}
+
+	info := modelConfigToModelInfo("test-model", cfg)
+
+	if info.Params != nil {
+		t.Errorf("expected nil Params, got %+v", info.Params)
+	}
+}
+
+func TestApplyModelSettings(t *testing.T) {
+	// Save and restore viper state.
+	originalViper := viper.AllSettings()
+	defer func() {
+		viper.Reset()
+		for k, v := range originalViper {
+			viper.Set(k, v)
+		}
+	}()
+
+	t.Run("applies model params when not explicitly set", func(t *testing.T) {
+		viper.Reset()
+
+		temp := float32(0.8)
+		topK := int32(50)
+		maxTokens := 4096
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				Temperature: &temp,
+				TopK:        &topK,
+				MaxTokens:   &maxTokens,
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.Temperature == nil || *config.Temperature != 0.8 {
+			t.Errorf("expected temperature 0.8, got %v", config.Temperature)
+		}
+		if config.TopK == nil || *config.TopK != 50 {
+			t.Errorf("expected topK 50, got %v", config.TopK)
+		}
+		if config.MaxTokens != 4096 {
+			t.Errorf("expected maxTokens 4096, got %d", config.MaxTokens)
+		}
+	})
+
+	t.Run("explicit viper values take precedence", func(t *testing.T) {
+		viper.Reset()
+		viper.Set("temperature", 0.3)
+
+		temp := float32(0.8)
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				Temperature: &temp,
+			},
+		}
+
+		explicitTemp := float32(0.3)
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+			Temperature: &explicitTemp,
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		// Temperature should NOT be overridden because it's explicitly set in viper
+		if config.Temperature == nil || *config.Temperature != 0.3 {
+			t.Errorf("expected temperature 0.3 (explicit), got %v", config.Temperature)
+		}
+	})
+
+	t.Run("nil model info is safe", func(t *testing.T) {
+		viper.Reset()
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		// Should not panic
+		ApplyModelSettings(config, nil)
+
+		if config.Temperature != nil {
+			t.Errorf("expected nil temperature, got %v", config.Temperature)
+		}
+	})
+
+	t.Run("model info without params is safe", func(t *testing.T) {
+		viper.Reset()
+
+		modelInfo := &ModelInfo{ID: "test-model"}
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.Temperature != nil {
+			t.Errorf("expected nil temperature, got %v", config.Temperature)
+		}
+	})
+
+	t.Run("modelSettings from viper takes priority over ModelInfo.Params", func(t *testing.T) {
+		viper.Reset()
+
+		// Set up modelSettings in viper (simulating config file)
+		viper.Set("modelSettings", map[string]any{
+			"custom/test-model": map[string]any{
+				"temperature": 0.5,
+				"topK":        30,
+			},
+		})
+
+		// ModelInfo has different params
+		temp := float32(0.8)
+		topK := int32(50)
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				Temperature: &temp,
+				TopK:        &topK,
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		// modelSettings should win over ModelInfo.Params
+		if config.Temperature == nil || *config.Temperature != 0.5 {
+			t.Errorf("expected temperature 0.5 (from modelSettings), got %v", config.Temperature)
+		}
+		if config.TopK == nil || *config.TopK != 30 {
+			t.Errorf("expected topK 30 (from modelSettings), got %v", config.TopK)
+		}
+	})
+
+	t.Run("stop sequences applied from model params", func(t *testing.T) {
+		viper.Reset()
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				StopSequences: []string{"STOP", "END"},
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if len(config.StopSequences) != 2 || config.StopSequences[0] != "STOP" {
+			t.Errorf("expected stop sequences [STOP END], got %v", config.StopSequences)
+		}
+	})
+
+	t.Run("thinking level applied from model params", func(t *testing.T) {
+		viper.Reset()
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				ThinkingLevel: ThinkingHigh,
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.ThinkingLevel != ThinkingHigh {
+			t.Errorf("expected thinking level high, got %v", config.ThinkingLevel)
+		}
+	})
+
+	t.Run("system prompt applied from model params", func(t *testing.T) {
+		viper.Reset()
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				SystemPrompt: "You are a coding assistant.",
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.SystemPrompt != "You are a coding assistant." {
+			t.Errorf("expected system prompt to be set, got %q", config.SystemPrompt)
+		}
+	})
+
+	t.Run("explicit system prompt takes precedence", func(t *testing.T) {
+		viper.Reset()
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				SystemPrompt: "Model-specific prompt",
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString:  "custom/test-model",
+			SystemPrompt: "Global prompt",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		// Global system prompt should NOT be overridden because config
+		// already has a non-empty SystemPrompt.
+		if config.SystemPrompt != "Global prompt" {
+			t.Errorf("expected global prompt preserved, got %q", config.SystemPrompt)
+		}
+	})
+
+	t.Run("system prompt from file path", func(t *testing.T) {
+		viper.Reset()
+
+		// Create a temp file with a system prompt
+		tmpFile, err := os.CreateTemp("", "kit-test-prompt-*.txt")
+		if err != nil {
+			t.Fatal(err)
+		}
+		defer func() { _ = os.Remove(tmpFile.Name()) }()
+		if _, err := tmpFile.WriteString("  Prompt from file  "); err != nil {
+			t.Fatal(err)
+		}
+		_ = tmpFile.Close()
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				SystemPrompt: tmpFile.Name(),
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.SystemPrompt != "Prompt from file" {
+			t.Errorf("expected trimmed file content, got %q", config.SystemPrompt)
+		}
+	})
+
+	t.Run("modelSettings system prompt overrides custom model params", func(t *testing.T) {
+		viper.Reset()
+
+		viper.Set("modelSettings", map[string]any{
+			"custom/test-model": map[string]any{
+				"systemPrompt": "From modelSettings",
+			},
+		})
+
+		modelInfo := &ModelInfo{
+			ID: "test-model",
+			Params: &GenerationParams{
+				SystemPrompt: "From custom model",
+			},
+		}
+
+		config := &ProviderConfig{
+			ModelString: "custom/test-model",
+		}
+
+		ApplyModelSettings(config, modelInfo)
+
+		if config.SystemPrompt != "From modelSettings" {
+			t.Errorf("expected modelSettings prompt, got %q", config.SystemPrompt)
+		}
+	})
+}
@@ -241,6 +241,11 @@ func CreateProvider(ctx context.Context, config *ProviderConfig) (*ProviderResul
 		validateModelConfig(config, modelInfo)
 	}

+	// Apply per-model generation parameter defaults. Model-level params are
+	// only applied for fields where the user hasn't explicitly set a value
+	// via CLI flag or global config.
+	ApplyModelSettings(config, modelInfo)
+
 	// Create the base provider
 	var result *ProviderResult
 	var createErr error
@@ -26,6 +26,11 @@ type ModelInfo struct {
 	ProviderNPM string // Model-specific provider npm override (e.g. "@ai-sdk/anthropic")
 	BaseURL     string // Per-model base URL override (custom models only)
 	APIKey      string // Per-model API key override (custom models only)
+
+	// Params holds per-model generation parameter defaults. These are applied
+	// when the user hasn't explicitly set the corresponding CLI flag or global
+	// config value. Nil pointer fields mean "no model-level default".
+	Params *GenerationParams
 }

 // SupportsCaching returns true if this model family supports prompt caching.
@@ -236,6 +241,18 @@ func (r *ModelsRegistry) LookupModel(provider, modelID string) *ModelInfo {
 	return &modelInfo
 }

+// LookupModelForSettings is a convenience function that parses a
+// "provider/model" string and looks up the ModelInfo in the global registry.
+// Returns nil when the model string is invalid or the model is unknown.
+// Used by Kit.SetModel to pre-apply per-model settings before CreateProvider.
+func LookupModelForSettings(modelString string) *ModelInfo {
+	provider, modelName, err := ParseModelString(modelString)
+	if err != nil {
+		return nil
+	}
+	return GetGlobalRegistry().LookupModel(provider, modelName)
+}
+
 // getRequiredEnvVars returns the required environment variables for a provider.
 func (r *ModelsRegistry) getRequiredEnvVars(provider string) ([]string, error) {
 	providerInfo, exists := r.providers[provider]
@@ -60,15 +60,16 @@ type MCPConnection struct {
 // creation, health monitoring, and cleanup. The pool runs background health checks
 // to proactively identify and remove unhealthy connections.
 type MCPConnectionPool struct {
-	connections map[string]*MCPConnection
-	config      *ConnectionPoolConfig
-	mu          sync.RWMutex
-	model       fantasy.LanguageModel
-	ctx         context.Context
-	cancel      context.CancelFunc
-	debug       bool
-	debugLogger DebugLogger
-	oauthFlow   *OAuthFlowRunner
+	connections       map[string]*MCPConnection
+	config            *ConnectionPoolConfig
+	mu                sync.RWMutex
+	model             fantasy.LanguageModel
+	ctx               context.Context
+	cancel            context.CancelFunc
+	debug             bool
+	debugLogger       DebugLogger
+	oauthFlow         *OAuthFlowRunner
+	tokenStoreFactory TokenStoreFactory // custom factory for per-server token stores (nil = default FileTokenStore)
 }

 // NewMCPConnectionPool creates a new MCP connection pool with the specified configuration.
@@ -76,19 +77,20 @@ type MCPConnectionPool struct {
 // goroutine for periodic health checks that runs until Close is called.
 // The model parameter is used for MCP servers that require sampling support.
 // Thread-safe for concurrent use immediately after creation.
-func NewMCPConnectionPool(config *ConnectionPoolConfig, model fantasy.LanguageModel, debug bool, authHandler MCPAuthHandler) *MCPConnectionPool {
+func NewMCPConnectionPool(config *ConnectionPoolConfig, model fantasy.LanguageModel, debug bool, authHandler MCPAuthHandler, tokenStoreFactory TokenStoreFactory) *MCPConnectionPool {
 	if config == nil {
 		config = DefaultConnectionPoolConfig()
 	}

 	ctx, cancel := context.WithCancel(context.Background())
 	pool := &MCPConnectionPool{
-		connections: make(map[string]*MCPConnection),
-		config:      config,
-		model:       model,
-		ctx:         ctx,
-		cancel:      cancel,
-		debug:       debug,
+		connections:       make(map[string]*MCPConnection),
+		config:            config,
+		model:             model,
+		ctx:               ctx,
+		cancel:            cancel,
+		debug:             debug,
+		tokenStoreFactory: tokenStoreFactory,
 	}

 	if authHandler != nil {
@@ -367,7 +369,7 @@ func (p *MCPConnectionPool) createSSEClient(ctx context.Context, serverConfig co
 	// scopes are discovered automatically via dynamic client registration and
 	// server metadata (RFC 9728).
 	if p.oauthFlow != nil {
-		tokenStore, tsErr := NewFileTokenStore(serverConfig.URL)
+		tokenStore, tsErr := p.createTokenStore(serverConfig.URL)
 		if tsErr != nil {
 			return nil, fmt.Errorf("failed to create token store: %w", tsErr)
 		}
@@ -414,7 +416,7 @@ func (p *MCPConnectionPool) createStreamableClient(ctx context.Context, serverCo
 	// scopes are discovered automatically via dynamic client registration and
 	// server metadata (RFC 9728).
 	if p.oauthFlow != nil {
-		tokenStore, tsErr := NewFileTokenStore(serverConfig.URL)
+		tokenStore, tsErr := p.createTokenStore(serverConfig.URL)
 		if tsErr != nil {
 			return nil, fmt.Errorf("failed to create token store: %w", tsErr)
 		}
@@ -437,6 +439,16 @@ func (p *MCPConnectionPool) createStreamableClient(ctx context.Context, serverCo
 	return streamableClient, nil
 }

+// createTokenStore creates a token store for the given server URL.
+// If a custom TokenStoreFactory is configured, it is used; otherwise the
+// default file-backed token store is created.
+func (p *MCPConnectionPool) createTokenStore(serverURL string) (transport.TokenStore, error) {
+	if p.tokenStoreFactory != nil {
+		return p.tokenStoreFactory(serverURL)
+	}
+	return NewFileTokenStore(serverURL)
+}
+
 // initializeClient initializes the client
 func (p *MCPConnectionPool) initializeClient(ctx context.Context, client client.MCPClient) error {
 	initCtx, cancel := context.WithTimeout(ctx, 5*time.Minute)
@@ -583,6 +595,27 @@ func (p *MCPConnectionPool) GetClients() map[string]client.MCPClient {
 	return clients
 }

+// RemoveConnection closes and removes a single connection from the pool.
+// Returns an error if the connection does not exist or if closing fails.
+// Thread-safe for concurrent use.
+func (p *MCPConnectionPool) RemoveConnection(serverName string) error {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	conn, exists := p.connections[serverName]
+	if !exists {
+		return fmt.Errorf("connection %q not found in pool", serverName)
+	}
+
+	err := conn.client.Close()
+	delete(p.connections, serverName)
+
+	if p.debugLogger != nil && p.debugLogger.IsDebugEnabled() {
+		p.debugLogger.LogDebug(fmt.Sprintf("[POOL] Removed connection %s", serverName))
+	}
+	return err
+}
+
 // Close gracefully shuts down the connection pool, closing all client connections
 // and stopping the background health check goroutine. It attempts to close all
 // connections even if some fail, logging any errors encountered.
@@ -20,19 +20,25 @@ import (
 // pooling, health checks, tool name prefixing to avoid conflicts, and sampling support for LLM interactions.
 // Thread-safe for concurrent tool invocations.
 type MCPToolManager struct {
-	connectionPool *MCPConnectionPool
-	tools          []fantasy.AgentTool
-	toolMap        map[string]*toolMapping // maps prefixed tool names to their server and original name
-	mu             sync.Mutex              // protects tools and toolMap during parallel loading
-	model          fantasy.LanguageModel   // LLM model for sampling
-	authHandler    MCPAuthHandler          // OAuth handler for remote servers (nil = no OAuth)
-	config         *config.Config
-	debug          bool
-	debugLogger    DebugLogger
+	connectionPool    *MCPConnectionPool
+	tools             []fantasy.AgentTool
+	toolMap           map[string]*toolMapping // maps prefixed tool names to their server and original name
+	mu                sync.Mutex              // protects tools and toolMap during parallel loading
+	model             fantasy.LanguageModel   // LLM model for sampling
+	authHandler       MCPAuthHandler          // OAuth handler for remote servers (nil = no OAuth)
+	tokenStoreFactory TokenStoreFactory       // factory for creating per-server token stores (nil = default FileTokenStore)
+	config            *config.Config
+	debug             bool
+	debugLogger       DebugLogger

 	// onServerLoaded, if non-nil, is called when each server finishes loading.
 	// Called with server name, tool count, and error (nil on success).
 	onServerLoaded func(serverName string, toolCount int, err error)
+
+	// onToolsChanged, if non-nil, is called after AddServer or RemoveServer
+	// mutates the tool list. The agent layer uses this to trigger a
+	// rebuildFantasyAgent so the LLM sees the updated tools.
+	onToolsChanged func()
 }

 // toolMapping stores the mapping between prefixed tool names and their original details
@@ -69,6 +75,14 @@ func (m *MCPToolManager) SetAuthHandler(handler MCPAuthHandler) {
 	m.authHandler = handler
 }

+// SetTokenStoreFactory sets a custom factory for creating per-server OAuth token
+// stores. When set, the factory is called for each remote MCP server instead of
+// using the default file-based token store. This method should be called before
+// LoadTools.
+func (m *MCPToolManager) SetTokenStoreFactory(factory TokenStoreFactory) {
+	m.tokenStoreFactory = factory
+}
+
 // SetDebugLogger sets the debug logger for the tool manager.
 // The logger will be used to output detailed debugging information about MCP connections,
 // tool loading, and execution. If a connection pool exists, it will also be configured
@@ -87,6 +101,126 @@ func (m *MCPToolManager) SetOnServerLoaded(cb func(serverName string, toolCount
 	m.onServerLoaded = cb
 }

+// SetOnToolsChanged sets the callback that's invoked after AddServer or
+// RemoveServer mutates the tool list. The agent layer uses this to trigger
+// a rebuild of the fantasy agent so the LLM sees the updated tool set.
+func (m *MCPToolManager) SetOnToolsChanged(cb func()) {
+	m.onToolsChanged = cb
+}
+
+// AddServer connects to a new MCP server at runtime and loads its tools.
+// The server's tools are immediately available to the agent after this call.
+// Returns the number of tools loaded from the server.
+//
+// If the connection pool has not been initialised yet (i.e. LoadTools was never
+// called), AddServer creates one automatically using the manager's current
+// configuration.
+//
+// Returns an error if a server with the same name is already loaded, or if
+// the connection or tool loading fails.
+func (m *MCPToolManager) AddServer(ctx context.Context, name string, cfg config.MCPServerConfig) (int, error) {
+	m.mu.Lock()
+	// Check for duplicate.
+	if _, exists := m.toolMap[name+"__"]; exists {
+		m.mu.Unlock()
+		return 0, fmt.Errorf("MCP server %q is already loaded", name)
+	}
+	// More thorough duplicate check: scan toolMap for any key with the server prefix.
+	prefix := name + "__"
+	for k := range m.toolMap {
+		if len(k) >= len(prefix) && k[:len(prefix)] == prefix {
+			m.mu.Unlock()
+			return 0, fmt.Errorf("MCP server %q is already loaded", name)
+		}
+	}
+	m.mu.Unlock()
+
+	// Lazily create the connection pool if LoadTools was never called.
+	m.ensureConnectionPool()
+
+	count, err := m.loadServerTools(ctx, name, cfg)
+	if err != nil {
+		return 0, fmt.Errorf("failed to add MCP server %q: %w", name, err)
+	}
+
+	// Notify listeners.
+	if m.onServerLoaded != nil {
+		m.onServerLoaded(name, count, nil)
+	}
+	if m.onToolsChanged != nil {
+		m.onToolsChanged()
+	}
+
+	return count, nil
+}
+
+// RemoveServer disconnects an MCP server and removes all its tools.
+// After this call the agent will no longer see or be able to call tools from
+// the named server. Returns an error if the server is not loaded.
+func (m *MCPToolManager) RemoveServer(name string) error {
+	prefix := name + "__"
+
+	m.mu.Lock()
+
+	// Check the server actually has tools loaded.
+	found := false
+	for k := range m.toolMap {
+		if len(k) >= len(prefix) && k[:len(prefix)] == prefix {
+			found = true
+			break
+		}
+	}
+	if !found {
+		m.mu.Unlock()
+		return fmt.Errorf("MCP server %q is not loaded", name)
+	}
+
+	// Remove tools belonging to this server.
+	newTools := make([]fantasy.AgentTool, 0, len(m.tools))
+	for _, t := range m.tools {
+		if len(t.Info().Name) < len(prefix) || t.Info().Name[:len(prefix)] != prefix {
+			newTools = append(newTools, t)
+		}
+	}
+	m.tools = newTools
+
+	// Remove tool mappings.
+	for k := range m.toolMap {
+		if len(k) >= len(prefix) && k[:len(prefix)] == prefix {
+			delete(m.toolMap, k)
+		}
+	}
+	m.mu.Unlock()
+
+	// Close the connection in the pool (best-effort).
+	if m.connectionPool != nil {
+		_ = m.connectionPool.RemoveConnection(name)
+	}
+
+	if m.onToolsChanged != nil {
+		m.onToolsChanged()
+	}
+
+	return nil
+}
+
+// ensureConnectionPool lazily creates a connection pool if one does not exist.
+// This allows AddServer to work even if LoadTools was never called.
+func (m *MCPToolManager) ensureConnectionPool() {
+	if m.connectionPool != nil {
+		return
+	}
+	debug := false
+	if m.config != nil {
+		debug = m.config.Debug
+	}
+	if m.debugLogger == nil {
+		m.debugLogger = NewSimpleDebugLogger(debug)
+	}
+	m.connectionPool = NewMCPConnectionPool(DefaultConnectionPoolConfig(), m.model, debug, m.authHandler, m.tokenStoreFactory)
+	m.connectionPool.SetDebugLogger(m.debugLogger)
+}
+
 // LoadTools loads tools from all configured MCP servers based on the provided configuration.
 // It initializes the connection pool, connects to each configured server, and loads their tools.
 // Tools from different servers are prefixed with the server name to avoid naming conflicts.
@@ -99,7 +233,7 @@ func (m *MCPToolManager) LoadTools(ctx context.Context, cfg *config.Config) erro
 	if m.debugLogger == nil {
 		m.debugLogger = NewSimpleDebugLogger(cfg.Debug)
 	}
-	m.connectionPool = NewMCPConnectionPool(DefaultConnectionPoolConfig(), m.model, cfg.Debug, m.authHandler)
+	m.connectionPool = NewMCPConnectionPool(DefaultConnectionPoolConfig(), m.model, cfg.Debug, m.authHandler, m.tokenStoreFactory)
 	m.connectionPool.SetDebugLogger(m.debugLogger)

 	// Load all servers in parallel. Each server connection (subprocess
@@ -290,6 +424,9 @@ func (m *MCPToolManager) GetLoadedServerNames() []string {
 // proper cleanup of stdio processes, network connections, and other resources.
 // It is safe to call Close multiple times.
 func (m *MCPToolManager) Close() error {
+	if m.connectionPool == nil {
+		return nil
+	}
 	return m.connectionPool.Close()
 }

@@ -0,0 +1,323 @@
+package tools
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+	"runtime"
+	"slices"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/mark3labs/kit/internal/config"
+)
+
+// testdataDir returns the absolute path to the testdata directory.
+func testdataDir(t *testing.T) string {
+	t.Helper()
+	_, file, _, ok := runtime.Caller(0)
+	if !ok {
+		t.Fatal("cannot determine test file path")
+	}
+	return filepath.Join(filepath.Dir(file), "testdata")
+}
+
+// echoServerConfig returns an MCPServerConfig for the test echo MCP server.
+func echoServerConfig(t *testing.T) config.MCPServerConfig {
+	t.Helper()
+	script := filepath.Join(testdataDir(t), "echo_server.py")
+	if _, err := os.Stat(script); err != nil {
+		t.Skipf("echo_server.py not found: %v", err)
+	}
+	return config.MCPServerConfig{
+		Command: []string{"python3", script},
+	}
+}
+
+// TestMCPToolManager_AddServer_Integration tests adding a real MCP server
+// at runtime and verifying tools are loaded.
+func TestMCPToolManager_AddServer_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	manager := NewMCPToolManager()
+	defer func() { _ = manager.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Track callbacks.
+	var mu sync.Mutex
+	var loadedServer string
+	var loadedCount int
+	toolsChangedCount := 0
+
+	manager.SetOnServerLoaded(func(name string, count int, err error) {
+		mu.Lock()
+		loadedServer = name
+		loadedCount = count
+		mu.Unlock()
+	})
+	manager.SetOnToolsChanged(func() {
+		mu.Lock()
+		toolsChangedCount++
+		mu.Unlock()
+	})
+
+	// Add the server.
+	count, err := manager.AddServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("AddServer failed: %v", err)
+	}
+
+	if count != 2 {
+		t.Errorf("Expected 2 tools from echo server, got %d", count)
+	}
+
+	// Verify callbacks fired.
+	mu.Lock()
+	if loadedServer != "echo" {
+		t.Errorf("Expected onServerLoaded for 'echo', got %q", loadedServer)
+	}
+	if loadedCount != 2 {
+		t.Errorf("Expected onServerLoaded count=2, got %d", loadedCount)
+	}
+	if toolsChangedCount != 1 {
+		t.Errorf("Expected onToolsChanged called once, got %d", toolsChangedCount)
+	}
+	mu.Unlock()
+
+	// Verify tools are accessible.
+	tools := manager.GetTools()
+	if len(tools) != 2 {
+		t.Fatalf("Expected 2 tools, got %d", len(tools))
+	}
+
+	// Verify tool names are prefixed.
+	toolNames := make(map[string]bool)
+	for _, tool := range tools {
+		toolNames[tool.Info().Name] = true
+	}
+	if !toolNames["echo__echo"] {
+		t.Error("Expected tool 'echo__echo'")
+	}
+	if !toolNames["echo__greet"] {
+		t.Error("Expected tool 'echo__greet'")
+	}
+
+	// Verify server appears in loaded names.
+	names := manager.GetLoadedServerNames()
+	if !slices.Contains(names, "echo") {
+		t.Errorf("Expected 'echo' in loaded server names, got: %v", names)
+	}
+}
+
+// TestMCPToolManager_RemoveServer_Integration tests removing a real MCP server
+// and verifying tools are cleaned up.
+func TestMCPToolManager_RemoveServer_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	manager := NewMCPToolManager()
+	defer func() { _ = manager.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add the server first.
+	count, err := manager.AddServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("AddServer failed: %v", err)
+	}
+	if count != 2 {
+		t.Fatalf("Expected 2 tools, got %d", count)
+	}
+
+	var mu sync.Mutex
+	toolsChangedCount := 0
+	manager.SetOnToolsChanged(func() {
+		mu.Lock()
+		toolsChangedCount++
+		mu.Unlock()
+	})
+
+	// Remove the server.
+	err = manager.RemoveServer("echo")
+	if err != nil {
+		t.Fatalf("RemoveServer failed: %v", err)
+	}
+
+	// Verify tools are gone.
+	tools := manager.GetTools()
+	if len(tools) != 0 {
+		t.Errorf("Expected 0 tools after removal, got %d", len(tools))
+	}
+
+	// Verify callback fired.
+	mu.Lock()
+	if toolsChangedCount != 1 {
+		t.Errorf("Expected onToolsChanged called once, got %d", toolsChangedCount)
+	}
+	mu.Unlock()
+
+	// Verify server is gone from loaded names.
+	names := manager.GetLoadedServerNames()
+	for _, n := range names {
+		if n == "echo" {
+			t.Error("Server 'echo' should not appear in loaded names after removal")
+		}
+	}
+
+	// Removing again should error.
+	err = manager.RemoveServer("echo")
+	if err == nil {
+		t.Fatal("Expected error removing already-removed server")
+	}
+	if !strings.Contains(err.Error(), "not loaded") {
+		t.Errorf("Expected 'not loaded' error, got: %v", err)
+	}
+}
+
+// TestMCPToolManager_AddRemoveMultiple_Integration tests adding and removing
+// multiple servers, verifying tool isolation.
+func TestMCPToolManager_AddRemoveMultiple_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	manager := NewMCPToolManager()
+	defer func() { _ = manager.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add two servers with the same binary but different names.
+	count1, err := manager.AddServer(ctx, "server-a", cfg)
+	if err != nil {
+		t.Fatalf("AddServer server-a failed: %v", err)
+	}
+	count2, err := manager.AddServer(ctx, "server-b", cfg)
+	if err != nil {
+		t.Fatalf("AddServer server-b failed: %v", err)
+	}
+
+	totalTools := count1 + count2
+	if totalTools != 4 {
+		t.Fatalf("Expected 4 total tools (2+2), got %d", totalTools)
+	}
+
+	tools := manager.GetTools()
+	if len(tools) != 4 {
+		t.Fatalf("Expected 4 tools, got %d", len(tools))
+	}
+
+	// Remove server-a, verify server-b tools remain.
+	err = manager.RemoveServer("server-a")
+	if err != nil {
+		t.Fatalf("RemoveServer server-a failed: %v", err)
+	}
+
+	tools = manager.GetTools()
+	if len(tools) != 2 {
+		t.Fatalf("Expected 2 tools after removing server-a, got %d", len(tools))
+	}
+
+	// Remaining tools should all be from server-b.
+	for _, tool := range tools {
+		if !strings.HasPrefix(tool.Info().Name, "server-b__") {
+			t.Errorf("Expected tool from server-b, got: %s", tool.Info().Name)
+		}
+	}
+
+	// Remove server-b.
+	err = manager.RemoveServer("server-b")
+	if err != nil {
+		t.Fatalf("RemoveServer server-b failed: %v", err)
+	}
+
+	tools = manager.GetTools()
+	if len(tools) != 0 {
+		t.Errorf("Expected 0 tools after removing all servers, got %d", len(tools))
+	}
+}
+
+// TestMCPToolManager_AddServer_DuplicateDetection_Integration tests that
+// adding a server with the same name as an already loaded server errors.
+func TestMCPToolManager_AddServer_DuplicateDetection_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	manager := NewMCPToolManager()
+	defer func() { _ = manager.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add the server.
+	_, err := manager.AddServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("First AddServer failed: %v", err)
+	}
+
+	// Try to add again with the same name.
+	_, err = manager.AddServer(ctx, "echo", cfg)
+	if err == nil {
+		t.Fatal("Expected error adding duplicate server")
+	}
+	if !strings.Contains(err.Error(), "already loaded") {
+		t.Errorf("Expected 'already loaded' error, got: %v", err)
+	}
+}
+
+// TestMCPToolManager_AddAfterRemove_Integration tests that a server can be
+// re-added after being removed.
+func TestMCPToolManager_AddAfterRemove_Integration(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping integration test in short mode")
+	}
+
+	manager := NewMCPToolManager()
+	defer func() { _ = manager.Close() }()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	cfg := echoServerConfig(t)
+
+	// Add, remove, re-add.
+	_, err := manager.AddServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("First AddServer failed: %v", err)
+	}
+
+	err = manager.RemoveServer("echo")
+	if err != nil {
+		t.Fatalf("RemoveServer failed: %v", err)
+	}
+
+	count, err := manager.AddServer(ctx, "echo", cfg)
+	if err != nil {
+		t.Fatalf("Re-AddServer failed: %v", err)
+	}
+	if count != 2 {
+		t.Errorf("Expected 2 tools on re-add, got %d", count)
+	}
+
+	tools := manager.GetTools()
+	if len(tools) != 2 {
+		t.Errorf("Expected 2 tools after re-add, got %d", len(tools))
+	}
+}
@@ -0,0 +1,155 @@
+package tools
+
+import (
+	"context"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/mark3labs/kit/internal/config"
+)
+
+// TestMCPToolManager_AddServer_DuplicateName verifies that adding a server
+// with a name that already exists returns an error.
+func TestMCPToolManager_AddServer_DuplicateName(t *testing.T) {
+	manager := NewMCPToolManager()
+
+	cfg := config.MCPServerConfig{
+		Command: []string{"non-existent-command"},
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	// First add will fail (bad command), but let's test the duplicate detection
+	// by simulating a loaded server via LoadTools first.
+	loadCfg := &config.Config{
+		MCPServers: map[string]config.MCPServerConfig{
+			"test-server": cfg,
+		},
+	}
+	// This will fail to load but creates the connection pool.
+	_ = manager.LoadTools(ctx, loadCfg)
+
+	// Now try to add the same server name — the tools didn't load (bad command),
+	// so AddServer should not find a duplicate and should fail with connection error.
+	_, err := manager.AddServer(ctx, "test-server", cfg)
+	if err == nil {
+		t.Fatal("Expected error when adding server with bad command, got nil")
+	}
+	// It should be a connection error, not a duplicate error.
+	if strings.Contains(err.Error(), "already loaded") {
+		t.Fatalf("Should not report duplicate since server failed to load initially: %v", err)
+	}
+}
+
+// TestMCPToolManager_RemoveServer_NotLoaded verifies that removing a server
+// that doesn't exist returns an appropriate error.
+func TestMCPToolManager_RemoveServer_NotLoaded(t *testing.T) {
+	manager := NewMCPToolManager()
+
+	err := manager.RemoveServer("nonexistent")
+	if err == nil {
+		t.Fatal("Expected error when removing non-existent server, got nil")
+	}
+	if !strings.Contains(err.Error(), "not loaded") {
+		t.Errorf("Expected 'not loaded' error, got: %v", err)
+	}
+}
+
+// TestMCPToolManager_AddServer_CreatesConnectionPool verifies that AddServer
+// lazily creates a connection pool when LoadTools was never called.
+func TestMCPToolManager_AddServer_CreatesConnectionPool(t *testing.T) {
+	manager := NewMCPToolManager()
+
+	// Connection pool should be nil initially.
+	if manager.connectionPool != nil {
+		t.Fatal("Expected nil connection pool before any operation")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	// AddServer with a bad command — should fail, but the pool should be created.
+	_, err := manager.AddServer(ctx, "lazy-server", config.MCPServerConfig{
+		Command: []string{"non-existent-command"},
+	})
+	if err == nil {
+		t.Fatal("Expected error for bad command")
+	}
+
+	// Connection pool should have been created.
+	if manager.connectionPool == nil {
+		t.Fatal("Expected connection pool to be created lazily by AddServer")
+	}
+}
+
+// TestMCPToolManager_OnToolsChanged_Callback verifies that the onToolsChanged
+// callback fires on RemoveServer (we can't easily test AddServer with a real
+// MCP server, but we can test the callback wiring).
+func TestMCPToolManager_OnToolsChanged_Callback(t *testing.T) {
+	manager := NewMCPToolManager()
+
+	var mu sync.Mutex
+	callCount := 0
+	manager.SetOnToolsChanged(func() {
+		mu.Lock()
+		callCount++
+		mu.Unlock()
+	})
+
+	// RemoveServer on non-existent should NOT fire callback.
+	_ = manager.RemoveServer("nonexistent")
+
+	mu.Lock()
+	if callCount != 0 {
+		t.Errorf("Expected 0 callback calls for failed remove, got %d", callCount)
+	}
+	mu.Unlock()
+}
+
+// TestMCPToolManager_Close_NilPool verifies Close is safe when the connection
+// pool was never initialized.
+func TestMCPToolManager_Close_NilPool(t *testing.T) {
+	manager := NewMCPToolManager()
+	err := manager.Close()
+	if err != nil {
+		t.Fatalf("Expected nil error from Close with nil pool, got: %v", err)
+	}
+}
+
+// TestMCPConnectionPool_RemoveConnection_NotFound verifies that removing a
+// non-existent connection returns an error.
+func TestMCPConnectionPool_RemoveConnection_NotFound(t *testing.T) {
+	pool := NewMCPConnectionPool(DefaultConnectionPoolConfig(), nil, false, nil, nil)
+	defer func() { _ = pool.Close() }()
+
+	err := pool.RemoveConnection("nonexistent")
+	if err == nil {
+		t.Fatal("Expected error for non-existent connection")
+	}
+	if !strings.Contains(err.Error(), "not found") {
+		t.Errorf("Expected 'not found' error, got: %v", err)
+	}
+}
+
+// TestMCPToolManager_EnsureConnectionPool_Idempotent verifies that
+// ensureConnectionPool doesn't recreate an existing pool.
+func TestMCPToolManager_EnsureConnectionPool_Idempotent(t *testing.T) {
+	manager := NewMCPToolManager()
+
+	// First call creates the pool.
+	manager.ensureConnectionPool()
+	pool1 := manager.connectionPool
+	if pool1 == nil {
+		t.Fatal("Expected pool to be created")
+	}
+
+	// Second call should be a no-op.
+	manager.ensureConnectionPool()
+	pool2 := manager.connectionPool
+	if pool1 != pool2 {
+		t.Fatal("Expected ensureConnectionPool to be idempotent")
+	}
+}
@@ -6,6 +6,7 @@ import (
 	"net/url"

 	"github.com/mark3labs/mcp-go/client"
+	"github.com/mark3labs/mcp-go/client/transport"
 )

 // MCPAuthHandler is the internal interface for handling MCP OAuth flows.
@@ -21,6 +22,12 @@ type MCPAuthHandler interface {
 	HandleAuth(ctx context.Context, serverName string, authURL string) (callbackURL string, err error)
 }

+// TokenStoreFactory creates a transport.TokenStore for a given MCP server URL.
+// When provided to the connection pool, it is called once per remote MCP server
+// instead of using the default file-based token store. Implementations can
+// return any transport.TokenStore — in-memory, database-backed, encrypted, etc.
+type TokenStoreFactory func(serverURL string) (transport.TokenStore, error)
+
 // OAuthFlowRunner handles the OAuth authorization flow when an MCP server
 // returns an OAuthAuthorizationRequiredError. It coordinates dynamic client
 // registration, PKCE generation, user authorization (via MCPAuthHandler),
@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Minimal MCP server over stdio for testing. Exposes one tool: echo."""
+import json
+import sys
+
+
+def read_message():
+    """Read a JSON-RPC message from stdin."""
+    line = sys.stdin.readline()
+    if not line:
+        return None
+    return json.loads(line.strip())
+
+
+def write_message(msg):
+    """Write a JSON-RPC message to stdout."""
+    sys.stdout.write(json.dumps(msg) + "\n")
+    sys.stdout.flush()
+
+
+def handle(msg):
+    method = msg.get("method", "")
+    mid = msg.get("id")
+
+    if method == "initialize":
+        write_message({
+            "jsonrpc": "2.0",
+            "id": mid,
+            "result": {
+                "protocolVersion": "2024-11-05",
+                "capabilities": {"tools": {}},
+                "serverInfo": {"name": "test-echo", "version": "1.0.0"},
+            },
+        })
+    elif method == "notifications/initialized":
+        pass  # no response needed
+    elif method == "tools/list":
+        write_message({
+            "jsonrpc": "2.0",
+            "id": mid,
+            "result": {
+                "tools": [
+                    {
+                        "name": "echo",
+                        "description": "Echoes the input text back.",
+                        "inputSchema": {
+                            "type": "object",
+                            "properties": {
+                                "text": {"type": "string", "description": "Text to echo"}
+                            },
+                            "required": ["text"],
+                        },
+                    },
+                    {
+                        "name": "greet",
+                        "description": "Returns a greeting.",
+                        "inputSchema": {
+                            "type": "object",
+                            "properties": {
+                                "name": {"type": "string", "description": "Name to greet"}
+                            },
+                            "required": ["name"],
+                        },
+                    },
+                ]
+            },
+        })
+    elif method == "tools/call":
+        tool_name = msg["params"]["name"]
+        args = msg["params"].get("arguments", {})
+        if tool_name == "echo":
+            text = args.get("text", "")
+            write_message({
+                "jsonrpc": "2.0",
+                "id": mid,
+                "result": {
+                    "content": [{"type": "text", "text": text}]
+                },
+            })
+        elif tool_name == "greet":
+            name = args.get("name", "World")
+            write_message({
+                "jsonrpc": "2.0",
+                "id": mid,
+                "result": {
+                    "content": [{"type": "text", "text": f"Hello, {name}!"}]
+                },
+            })
+        else:
+            write_message({
+                "jsonrpc": "2.0",
+                "id": mid,
+                "error": {"code": -32601, "message": f"Unknown tool: {tool_name}"},
+            })
+    elif method == "ping":
+        write_message({"jsonrpc": "2.0", "id": mid, "result": {}})
+    else:
+        if mid is not None:
+            write_message({
+                "jsonrpc": "2.0",
+                "id": mid,
+                "error": {"code": -32601, "message": f"Unknown method: {method}"},
+            })
+
+
+if __name__ == "__main__":
+    while True:
+        msg = read_message()
+        if msg is None:
+            break
+        handle(msg)
@@ -139,7 +139,9 @@ func (h *CLIEventHandler) Handle(msg tea.Msg) {
 		case "block":
 			h.cli.DisplayExtensionBlock(e.Text, e.BorderColor, e.Subtitle)
 		default:
-			fmt.Println(e.Text)
+			// Route unstyled extension prints through the system message
+			// renderer so they get consistent formatting and timestamps.
+			h.cli.DisplayInfo(e.Text)
 		}

 	case app.StepCompleteEvent:
@@ -109,9 +109,7 @@ func SetupCLI(opts *CLISetupOptions) (*CLI, error) {
 		}
 	}

-	fmt.Println("")
-
-	// Display model info
+	// Display model info (the system message block provides its own spacing).
 	if provider != "unknown" && model != "unknown" {
 		cli.DisplayInfo(fmt.Sprintf("Model loaded: %s (%s)", provider, model))
 	}
@@ -51,6 +51,12 @@ type Kit struct {
 	authHandler    MCPAuthHandler // OAuth handler for remote MCP servers (may need Close)
 	opts           *Options       // stored for reload operations (skills, etc.)

+	// hasCustomSystemPrompt is true when the user explicitly configured a
+	// system prompt (via --system-prompt flag, config file, or SDK option).
+	// When false, per-model system prompts from modelSettings/customModels
+	// can replace the default prompt on model switch.
+	hasCustomSystemPrompt bool
+
 	// Hook registries — interception layer (see hooks.go).
 	beforeToolCall  *hookRegistry[BeforeToolCallHook, BeforeToolCallResult]
 	afterToolResult *hookRegistry[AfterToolResultHook, AfterToolResultResult]
@@ -140,6 +146,79 @@ func (m *Kit) MCPToolsReady() bool {
 	return m.agent.MCPToolsReady()
 }

+// MCPServerStatus describes the runtime state of a loaded MCP server.
+type MCPServerStatus struct {
+	// Name is the configured server name.
+	Name string
+	// ToolCount is the number of tools loaded from this server.
+	ToolCount int
+}
+
+// AddMCPServer connects to a new MCP server at runtime and makes its tools
+// available to the agent immediately. The server's tools are prefixed with the
+// server name (e.g. "myserver__tool_name") to avoid naming conflicts, matching
+// the behaviour of servers loaded at initialization.
+//
+// Returns the number of tools loaded from the server.
+//
+// AddMCPServer is safe to call while the agent is idle. If a turn is in
+// progress ([Kit.IsGenerating] returns true), the new tools will be visible
+// starting from the next LLM step.
+//
+// Example:
+//
+//	n, err := k.AddMCPServer(ctx, "github", kit.MCPServerConfig{
+//	    Command: []string{"npx", "-y", "@modelcontextprotocol/server-github"},
+//	    Environment: map[string]string{"GITHUB_TOKEN": os.Getenv("GITHUB_TOKEN")},
+//	})
+func (m *Kit) AddMCPServer(ctx context.Context, name string, cfg MCPServerConfig) (int, error) {
+	return m.agent.AddMCPServer(ctx, name, cfg)
+}
+
+// RemoveMCPServer disconnects an MCP server and removes all its tools from
+// the agent. After this call the agent will no longer see or be able to call
+// tools from the named server.
+//
+// RemoveMCPServer is safe to call while the agent is idle. If a turn is in
+// progress, the tools are removed at the next LLM step. Any in-flight tool
+// calls to the removed server will fail gracefully.
+//
+// Returns an error if the named server is not currently loaded.
+func (m *Kit) RemoveMCPServer(name string) error {
+	return m.agent.RemoveMCPServer(name)
+}
+
+// ListMCPServers returns the status of all currently loaded MCP servers.
+// The returned slice is a snapshot; it is safe to read concurrently.
+func (m *Kit) ListMCPServers() []MCPServerStatus {
+	names := m.agent.GetLoadedServerNames()
+	if len(names) == 0 {
+		return nil
+	}
+
+	// Build a tool count per server by scanning tool names for the prefix.
+	toolNames := m.GetToolNames()
+	countByServer := make(map[string]int, len(names))
+	for _, tn := range toolNames {
+		for _, sn := range names {
+			prefix := sn + "__"
+			if len(tn) > len(prefix) && tn[:len(prefix)] == prefix {
+				countByServer[sn]++
+				break
+			}
+		}
+	}
+
+	result := make([]MCPServerStatus, 0, len(names))
+	for _, n := range names {
+		result = append(result, MCPServerStatus{
+			Name:      n,
+			ToolCount: countByServer[n],
+		})
+	}
+	return result
+}
+
 // GetExtensionToolCount returns the number of tools registered by extensions.
 func (m *Kit) GetExtensionToolCount() int {
 	return m.agent.GetExtensionToolCount()
@@ -221,9 +300,12 @@ func iterBranchMessages[T any](tm *session.TreeManager, fn func(*session.Message
 	return results
 }

-// SetModel changes the active model at runtime. The existing tools, system
-// prompt, and session are preserved. The model string should be in
-// "provider/model" format (e.g. "anthropic/claude-sonnet-4-5-20250929").
+// SetModel changes the active model at runtime. The existing tools and
+// session are preserved. When the new model has a per-model system prompt
+// (from modelSettings or customModels params), it is composed with the
+// current AGENTS.md context and skills before being applied.
+// The model string should be in "provider/model" format
+// (e.g. "anthropic/claude-sonnet-4-5-20250929").
 // Returns an error if the model string is invalid or the provider cannot
 // be created.
 func (m *Kit) SetModel(ctx context.Context, modelString string) error {
@@ -239,7 +321,7 @@ func (m *Kit) SetModel(ctx context.Context, modelString string) error {

 	// With message-level caching, thinking and caching can work together.
 	// No need to disable caching when thinking is enabled.
-	config := &models.ProviderConfig{
+	cfg := &models.ProviderConfig{
 		ModelString:    modelString,
 		SystemPrompt:   systemPrompt,
 		ProviderAPIKey: viper.GetString("provider-api-key"),
@@ -249,18 +331,50 @@ func (m *Kit) SetModel(ctx context.Context, modelString string) error {
 		ThinkingLevel:  thinkingLevel,
 		DisableCaching: false, // Caching enabled by default, works with thinking
 	}
-	temperature := float32(viper.GetFloat64("temperature"))
-	config.Temperature = &temperature
-	topP := float32(viper.GetFloat64("top-p"))
-	config.TopP = &topP
-	topK := int32(viper.GetInt("top-k"))
-	config.TopK = &topK
-	frequencyPenalty := float32(viper.GetFloat64("frequency-penalty"))
-	config.FrequencyPenalty = &frequencyPenalty
-	presencePenalty := float32(viper.GetFloat64("presence-penalty"))
-	config.PresencePenalty = &presencePenalty

-	if err := m.agent.SetModel(ctx, config); err != nil {
+	// Only set generation parameter pointers when the user has explicitly
+	// provided a value. This leaves nil pointers for unset params, allowing
+	// per-model defaults (modelSettings / customModels params) to apply.
+	if viper.IsSet("temperature") {
+		v := float32(viper.GetFloat64("temperature"))
+		cfg.Temperature = &v
+	}
+	if viper.IsSet("top-p") {
+		v := float32(viper.GetFloat64("top-p"))
+		cfg.TopP = &v
+	}
+	if viper.IsSet("top-k") {
+		v := int32(viper.GetInt("top-k"))
+		cfg.TopK = &v
+	}
+	if viper.IsSet("frequency-penalty") {
+		v := float32(viper.GetFloat64("frequency-penalty"))
+		cfg.FrequencyPenalty = &v
+	}
+	if viper.IsSet("presence-penalty") {
+		v := float32(viper.GetFloat64("presence-penalty"))
+		cfg.PresencePenalty = &v
+	}
+
+	// When the user hasn't set a custom global system prompt, check for a
+	// per-model system prompt. Pre-apply model settings to discover it,
+	// then compose with AGENTS.md context and skills if found.
+	if !m.hasCustomSystemPrompt {
+		// Temporarily clear the system prompt so ApplyModelSettings can
+		// detect that no explicit prompt is set and apply the per-model one.
+		cfg.SystemPrompt = ""
+		models.ApplyModelSettings(cfg, models.LookupModelForSettings(modelString))
+
+		if cfg.SystemPrompt != "" {
+			// Per-model system prompt found — compose with runtime context.
+			cfg.SystemPrompt = m.composeSystemPrompt(cfg.SystemPrompt)
+		} else {
+			// No per-model prompt — restore the global composed prompt.
+			cfg.SystemPrompt = systemPrompt
+		}
+	}
+
+	if err := m.agent.SetModel(ctx, cfg); err != nil {
 		return err
 	}

@@ -276,6 +390,32 @@ func (m *Kit) SetModel(ctx context.Context, modelString string) error {
 	return nil
 }

+// composeSystemPrompt takes a base system prompt and composes it with the
+// current runtime context: AGENTS.md content, skills metadata, and date/cwd.
+// This mirrors the composition done during Kit.New() initialization.
+func (m *Kit) composeSystemPrompt(basePrompt string) string {
+	cwd, _ := os.Getwd()
+	pb := skills.NewPromptBuilder(basePrompt)
+
+	// Inject AGENTS.md content as project context.
+	for _, cf := range m.contextFiles {
+		pb.WithSection("", fmt.Sprintf("Instructions from: %s\n\n%s", cf.Path, cf.Content))
+	}
+
+	// Inject skills metadata.
+	if len(m.skills) > 0 {
+		pb.WithSkills(m.skills)
+	}
+
+	// Append current date/time and working directory.
+	pb.WithSection("", fmt.Sprintf(
+		"Current date and time: %s\nCurrent working directory: %s",
+		time.Now().Format("Monday, January 2, 2006, 3:04:05 PM MST"), cwd,
+	))
+
+	return pb.Build()
+}
+
 // GetAvailableModels returns a list of known models from the registry. Each
 // entry includes provider, model ID, context limit, and whether the model
 // supports reasoning. This is an advisory list — models not in the registry
@@ -497,6 +637,17 @@ type Options struct {
 	// display a URL in a custom UI, redirect to a web app, etc.).
 	MCPAuthHandler MCPAuthHandler

+	// MCPTokenStoreFactory, if non-nil, is called to create a token store for
+	// each remote MCP server that requires OAuth. The factory receives the
+	// server's URL and returns a [MCPTokenStore] implementation.
+	//
+	// When nil (default), tokens are persisted to a JSON file at
+	// $XDG_CONFIG_HOME/.kit/mcp_tokens.json (or ~/.config/.kit/mcp_tokens.json).
+	//
+	// Use this to store tokens in a database, encrypt them, keep them
+	// in-memory, or write them to a custom file path.
+	MCPTokenStoreFactory MCPTokenStoreFactory
+
 	// OnMCPServerLoaded, if non-nil, is called when each MCP server finishes
 	// loading during Kit initialization. The callback receives the server name,
 	// tool count, and any error. Called from a background goroutine; safe to
@@ -582,16 +733,17 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 	// provider creation, session init) then runs outside the lock, allowing
 	// parallel subagent spawns to proceed concurrently.
 	var (
-		providerConfig *models.ProviderConfig
-		modelString    string
-		cwd            string
-		contextFiles   []*ContextFile
-		loadedSkills   []*Skill
-		mcpConfig      *config.Config
-		debug          bool
-		noExtensions   bool
-		maxSteps       int
-		streaming      bool
+		providerConfig        *models.ProviderConfig
+		modelString           string
+		cwd                   string
+		contextFiles          []*ContextFile
+		loadedSkills          []*Skill
+		mcpConfig             *config.Config
+		debug                 bool
+		noExtensions          bool
+		maxSteps              int
+		streaming             bool
+		hasCustomSystemPrompt bool
 	)

 	if err := func() error {
@@ -647,8 +799,41 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {

 		// Always compose the system prompt with runtime context: base prompt +
 		// AGENTS.md context + skills metadata + date/cwd.
+		//
+		// If the configured model has a per-model system prompt (via
+		// modelSettings or customModels params) and the user hasn't
+		// explicitly set system-prompt, use the per-model prompt as the
+		// base instead of the global default.
 		{
 			basePrompt := viper.GetString("system-prompt")
+
+			// Track whether the user explicitly configured a custom system
+			// prompt. When they haven't (basePrompt is the built-in default
+			// or empty), per-model system prompts can replace it on switch.
+			userSetSystemPrompt := basePrompt != "" && basePrompt != defaultSystemPrompt
+			hasCustomSystemPrompt = userSetSystemPrompt
+
+			// Check for per-model system prompt override when no explicit
+			// global system-prompt was configured by the user.
+			if !userSetSystemPrompt {
+				modelStr := viper.GetString("model")
+				if modelStr != "" {
+					if mi := models.LookupModelForSettings(modelStr); mi != nil {
+						var perModelParams *models.GenerationParams
+						// modelSettings takes priority over custom model params.
+						if ms := models.LoadModelSettingsFromConfig(); ms != nil {
+							perModelParams = ms[modelStr]
+						}
+						if perModelParams == nil && mi.Params != nil {
+							perModelParams = mi.Params
+						}
+						if perModelParams != nil && perModelParams.SystemPrompt != "" {
+							basePrompt = models.LoadSystemPromptValue(perModelParams.SystemPrompt)
+						}
+					}
+				}
+			}
+
 			pb := skills.NewPromptBuilder(basePrompt)

 			// Inject AGENTS.md content as project context.
@@ -745,6 +930,13 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 		}
 	}

+	// Set up custom token store factory for MCP OAuth tokens.
+	// The SDK MCPTokenStoreFactory is structurally identical to
+	// tools.TokenStoreFactory, so it can be assigned directly.
+	if opts.MCPTokenStoreFactory != nil {
+		setupOpts.TokenStoreFactory = tools.TokenStoreFactory(opts.MCPTokenStoreFactory)
+	}
+
 	if opts.CLI != nil {
 		setupOpts.ShowSpinner = opts.CLI.ShowSpinner
 		setupOpts.SpinnerFunc = opts.CLI.SpinnerFunc
@@ -774,24 +966,25 @@ func New(ctx context.Context, opts *Options) (*Kit, error) {
 	}

 	k := &Kit{
-		agent:           agentResult.Agent,
-		session:         sessionManager,
-		modelString:     modelString,
-		events:          newEventBus(),
-		autoCompact:     opts.AutoCompact,
-		compactionOpts:  opts.CompactionOptions,
-		contextFiles:    contextFiles,
-		skills:          loadedSkills,
-		extRunner:       agentResult.ExtRunner,
-		bufferedLogger:  agentResult.BufferedLogger,
-		authHandler:     setupOpts.AuthHandler,
-		opts:            opts,
-		beforeToolCall:  beforeToolCall,
-		afterToolResult: afterToolResult,
-		beforeTurn:      beforeTurn,
-		afterTurn:       afterTurn,
-		contextPrepare:  contextPrepare,
-		beforeCompact:   beforeCompact,
+		agent:                 agentResult.Agent,
+		session:               sessionManager,
+		modelString:           modelString,
+		events:                newEventBus(),
+		autoCompact:           opts.AutoCompact,
+		compactionOpts:        opts.CompactionOptions,
+		contextFiles:          contextFiles,
+		skills:                loadedSkills,
+		extRunner:             agentResult.ExtRunner,
+		bufferedLogger:        agentResult.BufferedLogger,
+		authHandler:           setupOpts.AuthHandler,
+		opts:                  opts,
+		hasCustomSystemPrompt: hasCustomSystemPrompt,
+		beforeToolCall:        beforeToolCall,
+		afterToolResult:       afterToolResult,
+		beforeTurn:            beforeTurn,
+		afterTurn:             afterTurn,
+		contextPrepare:        contextPrepare,
+		beforeCompact:         beforeCompact,
 	}

 	// Bridge extension events to SDK hooks.
@@ -0,0 +1,56 @@
+package kit_test
+
+import (
+	"testing"
+
+	kit "github.com/mark3labs/kit/pkg/kit"
+)
+
+// TestMCPServerStatus_TypeSurface verifies the MCPServerStatus type is
+// accessible and has the expected fields.
+func TestMCPServerStatus_TypeSurface(t *testing.T) {
+	s := kit.MCPServerStatus{
+		Name:      "test-server",
+		ToolCount: 5,
+	}
+	if s.Name != "test-server" {
+		t.Errorf("Expected Name 'test-server', got %q", s.Name)
+	}
+	if s.ToolCount != 5 {
+		t.Errorf("Expected ToolCount 5, got %d", s.ToolCount)
+	}
+}
+
+// TestMCPServerConfig_ForDynamicAdd verifies that MCPServerConfig can be
+// constructed with the expected fields for dynamic server management.
+func TestMCPServerConfig_ForDynamicAdd(t *testing.T) {
+	// Stdio server config.
+	stdio := kit.MCPServerConfig{
+		Command:     []string{"npx", "-y", "@modelcontextprotocol/server-github"},
+		Environment: map[string]string{"GITHUB_TOKEN": "test-token"},
+	}
+	if len(stdio.Command) != 3 {
+		t.Errorf("Expected 3 command parts, got %d", len(stdio.Command))
+	}
+	if stdio.Environment["GITHUB_TOKEN"] != "test-token" {
+		t.Error("Expected GITHUB_TOKEN in environment")
+	}
+
+	// Remote server config.
+	remote := kit.MCPServerConfig{
+		URL:     "https://mcp.example.com/sse",
+		Headers: []string{"Authorization: Bearer test"},
+	}
+	if remote.URL != "https://mcp.example.com/sse" {
+		t.Errorf("Unexpected URL: %s", remote.URL)
+	}
+
+	// Config with tool filtering.
+	filtered := kit.MCPServerConfig{
+		Command:      []string{"some-server"},
+		AllowedTools: []string{"read", "write"},
+	}
+	if len(filtered.AllowedTools) != 2 {
+		t.Errorf("Expected 2 allowed tools, got %d", len(filtered.AllowedTools))
+	}
+}
@@ -11,6 +11,7 @@ import (
 	"github.com/mark3labs/kit/internal/message"
 	"github.com/mark3labs/kit/internal/models"
 	"github.com/mark3labs/kit/internal/session"
+	"github.com/mark3labs/mcp-go/client/transport"
 )

 // ==== Message Types (internal/message/content.go) ====
@@ -204,6 +205,29 @@ type CompactionResult = compaction.CompactionResult
 // CompactionOptions configures compaction behaviour.
 type CompactionOptions = compaction.CompactionOptions

+// ==== MCP OAuth Types ====
+
+// MCPTokenStore persists OAuth tokens for a single MCP server. Implementations
+// must be safe for concurrent use.
+//
+// This is a type alias for the mcp-go transport.TokenStore interface. SDK
+// consumers can implement this interface to provide custom storage backends
+// (database, encrypted file, in-memory, etc.).
+type MCPTokenStore = transport.TokenStore
+
+// MCPToken represents an OAuth token for an MCP server, containing access
+// and refresh tokens along with expiration metadata.
+type MCPToken = transport.Token
+
+// MCPTokenStoreFactory creates an [MCPTokenStore] for a given MCP server URL.
+// It is called once per remote MCP server during connection setup.
+type MCPTokenStoreFactory func(serverURL string) (MCPTokenStore, error)
+
+// ErrMCPNoToken is the sentinel error that [MCPTokenStore] implementations
+// should return from GetToken when no token is stored for the server.
+// Callers can check for this with errors.Is.
+var ErrMCPNoToken = transport.ErrNoToken
+
 // ==== Constructor & Helper Functions ====

 // ParseModelString parses a model string in "provider/model" format.
Author	SHA1	Message	Date
Ed Zynda	e07c94f49d	feat(mcp): add dynamic MCP server loading and unloading - Add AddServer/RemoveServer to MCPToolManager for runtime server management - Add RemoveConnection to MCPConnectionPool for per-server teardown - Add AddMCPServer/RemoveMCPServer/ListMCPServers to Agent and SDK Kit - Lazily create connection pool so AddServer works without prior LoadTools - Wire onToolsChanged callback to trigger agent tool list rebuild - Make MCPToolManager.Close nil-safe when pool was never initialized Tests: - Integration tests with real stdio MCP server (Python echo server) - Agent-level tests using mock LLM model (no API key needed) - Unit tests for error paths, callbacks, idempotency, nil safety - SDK type surface tests	2026-04-09 13:54:11 +03:00
Ed Zynda	b87146a284	feat(sdk): add MCPTokenStoreFactory for custom OAuth token storage - Add MCPTokenStoreFactory option to kit.Options allowing SDK consumers to provide custom token storage backends for remote MCP servers - Thread TokenStoreFactory through the full chain: kit.Options → kitsetup → agent → MCPToolManager → MCPConnectionPool - Add createTokenStore() helper on connection pool that delegates to the factory or falls back to the default FileTokenStore - Export MCPTokenStore, MCPToken, MCPTokenStoreFactory, and ErrMCPNoToken in pkg/kit/types.go following SDK naming conventions - Default behavior (file-based storage) is preserved when factory is nil	2026-04-09 13:27:40 +03:00
Ed Zynda	186d9f7f44	fix(ui): route raw fmt.Print calls through proper renderers - event_handler: route default extension print level through DisplayInfo instead of bare fmt.Println for consistent styling and timestamps - factory: remove orphan fmt.Println("") before system messages; the renderer already manages its own spacing - app: PrintFromExtension non-interactive fallback now respects level, writing errors/info to stderr with prefix to keep stdout clean - app: PrintBlockFromExtension non-interactive fallback writes framed blocks to stderr instead of raw text to stdout	2026-04-09 13:00:23 +03:00
Ed Zynda	3a8ffc2104	feat(models): add per-model system prompt support - Add systemPrompt field to GenerationParams and config structs - On init, replace default system prompt with per-model prompt when user hasn't explicitly set one (via flag, config, or SDK option) - On model switch, detect per-model prompt and compose it with AGENTS.md, skills, and date/cwd context - Fix viper.IsSet bug: BindPFlag causes IsSet to return true for unset flags, so compare against defaultSystemPrompt instead - Agent.SetModel now updates stored system prompt from config - Export LoadModelSettingsFromConfig, LoadSystemPromptValue, and LookupModelForSettings for use by Kit.SetModel - Add tests for prompt apply, precedence, file path, and modelSettings override	2026-04-09 12:35:00 +03:00
Ed Zynda	e54570162e	feat(models): add per-model generation parameter defaults - Add modelSettings config section for attaching generation params (temperature, topP, topK, frequencyPenalty, presencePenalty, maxTokens, stopSequences, thinkingLevel) to any model by provider/model key - Add params field to customModels definitions for inline defaults - Change BuildProviderConfig and SetModel to use viper.IsSet so unset params remain nil, allowing model-level defaults to apply - Wire ApplyModelSettings into CreateProvider with priority order: CLI flags > global config > modelSettings > customModels params - Add GenerationParams to ModelInfo in the registry - Update default config template with modelSettings and customModels params examples	2026-04-09 12:07:42 +03:00