diff --git a/cmd/root.go b/cmd/root.go
index 3126aa7c..7e4bffb9 100644
--- a/cmd/root.go
+++ b/cmd/root.go
@@ -735,12 +735,27 @@ func runNormalMode(ctx context.Context) error {
 		viper.Set("model", "custom/custom")
 	}
 
-	// When --provider-url is set with an explicit --model that lacks a provider
-	// prefix (no "/"), auto-prefix with "custom/" for OpenAI-compatible endpoints.
+	// When --provider-url is set with an explicit --model, route through the
+	// "custom" provider (OpenAI-compatible wire). This honors the user's
+	// intent: passing a custom URL means "use THIS endpoint", not "speak
+	// the Google/Anthropic/etc. wire protocol against this endpoint".
+	//
+	// Any provider prefix on the model is stripped so a model name that
+	// happens to collide with a known provider (e.g. `google/gemma-4-12b`
+	// served by LM Studio) still resolves correctly. If you genuinely need
+	// to point a non-OpenAI wire (Anthropic, Google, ...) at a proxy URL,
+	// use the explicit `custom/<name>` form to opt out of the rewrite by
+	// configuring the proxy as that provider in your config file instead.
 	if viper.GetString("provider-url") != "" && modelFlagChanged {
 		model := viper.GetString("model")
-		if model != "" && !strings.Contains(model, "/") {
-			viper.Set("model", "custom/"+model)
+		if model != "" {
+			name := model
+			if _, after, ok := strings.Cut(model, "/"); ok {
+				name = after
+			}
+			if !strings.HasPrefix(model, "custom/") {
+				viper.Set("model", "custom/"+name)
+			}
 		}
 	}
 
diff --git a/www/pages/providers.md b/www/pages/providers.md
index 8b028a06..99d208b5 100644
--- a/www/pages/providers.md
+++ b/www/pages/providers.md
@@ -133,6 +133,15 @@ For self-hosted or proxy endpoints:
 kit --provider-url "https://my-proxy.example.com/v1" --model openai/gpt-4o
 ```
 
+When `--provider-url` is set with an explicit `--model`, Kit routes through the
+`custom` (OpenAI-compatible) wire and strips any provider prefix from the model
+name. So `openai/gpt-4o`, `google/gemma-4-12b`, and bare `gpt-4o` all resolve
+to the same endpoint — Kit treats `--provider-url` as authoritative about *where*
+to send the request, and the model string as just the upstream model id.
+
+This avoids name collisions when a local server (LM Studio, Ollama, vLLM, ...)
+happens to expose a model whose name matches a known cloud provider.
+
 When `--provider-url` is provided without `--model`, Kit automatically defaults to `custom/custom`:
 
 ```bash