🤖 feat: add Google provider with Gemini 3 support (#677)

Add Google as a model provider with support for Gemini 3.

## Changes

- Add `@ai-sdk/google` dependency
- Add Google provider creation in AIService with lazy loading pattern
matching Anthropic/OpenAI
- Add `GoogleProviderOptions` to `MuxProviderOptions` type system
- Add `GOOGLE_API_KEY` and `GOOGLE_BASE_URL` environment variable
support
- Add `GEMINI_3_PRO` to `KNOWN_MODELS` with `gemini-2.5-pro` tokenizer
- Add Google fallback to tokenizer (`google/gemini-2.5-pro`)
- Add test coverage for Gemini 3 tokenization and thinking policy

## Usage

Configure via `~/.mux/providers.jsonc`:
```json
{
  "google": {
    "apiKey": "YOUR_GOOGLE_API_KEY"
  }
}
```

Or via environment variable:
```bash
export GOOGLE_API_KEY="your-api-key"
```

Reference the model as:
- `google:gemini-3-pro-preview`
- `gemini-3` (alias)
- `gemini-3-pro` (alias)

## Infrastructure

The following infrastructure already supported Google:
- ✅ Tool definitions (google_search)
- ✅ Slash command registry
- ✅ Model display formatting

## Testing

- ✅ Tokenizer tests pass
- ✅ Thinking policy tests pass
- ✅ Type checking passes

_Generated with `mux`_

Signed-off-by: Thomas Kosiewski <tk@coder.com>

Author: ThomasK33

bun.lock

@@ -21,6 +21,28 @@ GPT-5 family of models:
 
 - `openai:gpt-5`
 - `openai:gpt-5-pro`
+
+#### Google (Cloud)
+
+Access Gemini models directly via Google's generative AI API:
+
+- `google:gemini-3-pro-preview`
+- `google:gemini-2.5-pro`
+- `google:gemini-2.5-flash`
+
+**Setup:**
+
+1. Get your API key from [Google AI Studio](https://aistudio.google.com/)
+2. Add to `~/.mux/providers.jsonc`:
+
+```jsonc
+{
+  "google": {
+    "apiKey": "AIza...",
+  },
+}
+```
+
 - `openai:gpt-5-codex`
 
 **Note:** Anthropic models are better supported than GPT-5 class models due to an outstanding issue in the Vercel AI SDK.
@@ -141,6 +163,10 @@ All providers are configured in `~/.mux/providers.jsonc`. Example configurations
   "openai": {
     "apiKey": "sk-...",
   },
+  // Required for Google models
+  "google": {
+    "apiKey": "AIza...",
+  },
   // Required for OpenRouter models
   "openrouter": {
     "apiKey": "sk-or-v1-...",

docs/models.md

@@ -46,6 +46,7 @@
   },
   "dependencies": {
     "@ai-sdk/anthropic": "^2.0.44",
+    "@ai-sdk/google": "^2.0.28",
     "@ai-sdk/openai": "^2.0.66",
     "@openrouter/ai-sdk-provider": "^1.2.2",
     "ghostty-web": "^0.1.1",

package.json

@@ -3,14 +3,14 @@ import { usePersistedState, readPersistedState, updatePersistedState } from "./u
 import { MODEL_ABBREVIATIONS } from "@/browser/utils/slashCommands/registry";
 import { defaultModel } from "@/common/utils/ai/models";
 
-const MAX_LRU_SIZE = 8;
+const MAX_LRU_SIZE = 12;
 const LRU_KEY = "model-lru";
 
 // Default models from abbreviations (for initial LRU population)
-// Ensure defaultModel is first, then fill with other abbreviations
+// Ensure defaultModel is first, then fill with other abbreviations (deduplicated)
 const DEFAULT_MODELS = [
   defaultModel,
-  ...Object.values(MODEL_ABBREVIATIONS).filter((m) => m !== defaultModel),
+  ...Array.from(new Set(Object.values(MODEL_ABBREVIATIONS))).filter((m) => m !== defaultModel),
 ].slice(0, MAX_LRU_SIZE);
 function persistModels(models: string[]): void {
   updatePersistedState(LRU_KEY, models.slice(0, MAX_LRU_SIZE));

src/browser/hooks/useModelLRU.ts

@@ -46,6 +46,12 @@ describe("getThinkingPolicyForModel", () => {
       "medium",
       "high",
     ]);
+    expect(getThinkingPolicyForModel("google:gemini-3-pro-preview-11-2025")).toEqual([
+      "off",
+      "low",
+      "medium",
+      "high",
+    ]);
   });
 });
 

src/browser/utils/thinking/policy.test.ts

@@ -2,7 +2,7 @@
  * Centralized model metadata. Update model versions here and everywhere else will follow.
  */
 
-type ModelProvider = "anthropic" | "openai";
+type ModelProvider = "anthropic" | "openai" | "google";
 
 interface KnownModelDefinition {
   /** Provider identifier used by SDK factories */
@@ -51,6 +51,7 @@ const MODEL_DEFINITIONS = {
     providerModelId: "gpt-5.1",
     aliases: ["gpt-5.1"],
     warm: true,
+    tokenizerOverride: "openai/gpt-5",
   },
   GPT_PRO: {
     provider: "openai",
@@ -62,12 +63,19 @@ const MODEL_DEFINITIONS = {
     providerModelId: "gpt-5.1-codex",
     aliases: ["codex"],
     warm: true,
+    tokenizerOverride: "openai/gpt-5",
   },
   GPT_MINI: {
     provider: "openai",
     providerModelId: "gpt-5.1-codex-mini",
     aliases: ["codex-mini"],
   },
+  GEMINI_3_PRO: {
+    provider: "google",
+    providerModelId: "gemini-3-pro-preview",
+    aliases: ["gemini-3", "gemini-3-pro"],
+    tokenizerOverride: "google/gemini-2.5-pro",
+  },
 } as const satisfies Record<string, KnownModelDefinition>;
 
 export type KnownModelKey = keyof typeof MODEL_DEFINITIONS;

src/common/constants/knownModels.ts

@@ -27,6 +27,13 @@ export async function importOllama() {
   return await import("ollama-ai-provider-v2");
 }
 
+/**
+ * Dynamically import the Google provider package
+ */
+export async function importGoogle() {
+  return await import("@ai-sdk/google");
+}
+
 /**
  * Dynamically import the OpenRouter provider package
  */
@@ -50,6 +57,7 @@ export async function importOpenRouter() {
 export const PROVIDER_REGISTRY = {
   anthropic: importAnthropic,
   openai: importOpenAI,
+  google: importGoogle,
   ollama: importOllama,
   openrouter: importOpenRouter,
 } as const;

src/common/constants/providers.ts

@@ -29,6 +29,12 @@ export interface OpenAIProviderOptions {
   simulateToolPolicyNoop?: boolean;
 }
 
+/**
+ * Google-specific options
+ */
+// eslint-disable-next-line @typescript-eslint/no-empty-object-type
+export interface GoogleProviderOptions {}
+
 /**
  * Ollama-specific options
  * Currently empty - Ollama is a local service and doesn't require special options.
@@ -52,6 +58,7 @@ export interface MuxProviderOptions {
   /** Provider-specific options */
   anthropic?: AnthropicProviderOptions;
   openai?: OpenAIProviderOptions;
+  google?: GoogleProviderOptions;
   ollama?: OllamaProviderOptions;
   openrouter?: OpenRouterProviderOptions;
 }

src/common/types/providerOptions.ts

@@ -203,6 +203,14 @@ export function buildProviderOptions(
     return options;
   }
 
+  // Build Google-specific options
+  if (provider === "google") {
+    // Google Gemini models don't currently support the same thinking/reasoning
+    // configuration as Anthropic/OpenAI, so return empty options for now
+    log.debug("buildProviderOptions: Google config - no specific options yet");
+    return {};
+  }
+
   // Build OpenRouter-specific options
   if (provider === "openrouter") {
     const reasoningEffort = OPENROUTER_REASONING_EFFORT[effectiveThinking];

src/common/utils/ai/providerOptions.ts

@@ -78,6 +78,18 @@ const buildProvidersFromEnv = (env: NodeJS.ProcessEnv): ProvidersConfig => {
     }
   }
 
+  const googleKey = trim(env.GOOGLE_API_KEY);
+  if (googleKey.length > 0) {
+    const entry: ProviderConfig = { apiKey: googleKey };
+
+    const baseUrl = trim(env.GOOGLE_BASE_URL);
+    if (baseUrl.length > 0) {
+      entry.baseUrl = baseUrl;
+    }
+
+    providers.google = entry;
+  }
+
   return providers;
 };
 
@@ -102,7 +114,7 @@ export const ensureProvidersConfig = (
   const providersFromEnv = buildProvidersFromEnv(env);
   if (!hasAnyConfiguredProvider(providersFromEnv)) {
     throw new Error(
-      "No provider credentials found. Configure providers.jsonc or set ANTHROPIC_API_KEY / OPENAI_API_KEY / OPENROUTER_API_KEY."
+      "No provider credentials found. Configure providers.jsonc or set ANTHROPIC_API_KEY / OPENAI_API_KEY / OPENROUTER_API_KEY / GOOGLE_API_KEY."
     );
   }