feat(google-gemini): add Google Gemini MCP with full LLM binding support

[viktormarinho]

Summary

• Adds a new google-gemini MCP that provides direct access to Google's Gemini models via the Google Generative AI API
• Implements the same 7 tools as the OpenRouter and Deco AI Gateway MCPs with identical binding schemas (@decocms/bindings/llm), enabling seamless interchangeability
• Accepts a Google AI API key via the authorization header for simple, no-OAuth authentication

Tools

Tool	Description
COLLECTION_LLM_LIST	List Gemini models with filtering and pagination
COLLECTION_LLM_GET	Get a single model by ID
LLM_METADATA	Model capabilities and supported media types
LLM_DO_STREAM	Streaming generation via @ai-sdk/google
LLM_DO_GENERATE	Non-streaming generation via @ai-sdk/google
COMPARE_MODELS	Side-by-side model comparison
RECOMMEND_MODEL	Task-based model recommendations

Implementation details

• Model listing: Calls the Gemini REST API (generativelanguage.googleapis.com/v1beta/models) directly, with pagination and 5-minute cache
• Pricing: Hardcoded pricing table since the Gemini API doesn't expose cost data
• Generation: Uses @ai-sdk/google (createGoogleGenerativeAI) which provides the same doStream()/doGenerate() interface as the OpenRouter provider
• Model IDs: Stripped of the models/ prefix (e.g., gemini-2.5-pro instead of models/gemini-2.5-pro)

Test plan

• Verify bun run check passes (TypeScript compilation)
• Verify bun run fmt and bun run lint pass
• Test model listing with a valid Gemini API key
• Test streaming and non-streaming generation
• Verify model comparison and recommendation tools work

Made with Cursor

---

Summary by cubic

Adds a new Google Gemini MCP with full LLM binding parity, enabling direct access to Gemini models. Drop-in compatible with existing bindings, with streaming and non-streaming generation.

• New Features

  • Implements LLM tools: list, get, metadata, stream, generate; plus compare and recommend.
  • Uses @ai-sdk/google for doStream/doGenerate with @decocms/bindings/llm schemas.
  • Lists models via Gemini REST with pagination and a 5-minute cache; includes a maintained pricing table.
  • Orders well-known models first and exposes multimodal support via LLM_METADATA.

• Migration

  • Pass a Google AI API key in the Authorization header only; “Bearer” prefix is accepted and stripped. No OAuth or env var fallback.
  • Model IDs omit the "models/" prefix (e.g., gemini-2.5-pro).
  • Can replace existing OpenRouter/Deco Gateway MCPs without code changes.

^{Written for commit c36e575. Summary will update on new commits.}

[github-actions]

🚀 Preview Deployments Ready!

Your changes have been deployed to preview environments:

📦 google-gemini

🔗 View Preview

These previews will be automatically updated with new commits to this PR.

---

_{Deployed from commit: a73c87f}

[cubic-dev-ai]

2 issues found across 20 files

Prompt for AI agents (all issues)

Check if these issues are valid — if so, understand the root cause of each and fix them.


<file name="google-gemini/README.md">

<violation number="1" location="google-gemini/README.md:10">
P2: The README instructs users to add the MCP to `deploy.json`, but deployment is automatic in this repo. This new step conflicts with the documented workflow and will mislead users.</violation>
</file>

<file name="google-gemini/server/tools/llm-binding.ts">

<violation number="1" location="google-gemini/server/tools/llm-binding.ts:421">
P1: The `flush()` implementation doesn't match its documented intent. `usage.promise.catch(() => {})` only adds an empty catch handler—it doesn't reject the promise. If the stream ends without a "finish" chunk, the promise hangs forever. Should call `usage.reject(new Error("Stream ended without finish chunk"))` instead.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: The flush() implementation doesn't match its documented intent. usage.promise.catch(() => {}) only adds an empty catch handler—it doesn't reject the promise. If the stream ends without a "finish" chunk, the promise hangs forever. Should call usage.reject(new Error("Stream ended without finish chunk")) instead.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At google-gemini/server/tools/llm-binding.ts, line 421:

<comment>The `flush()` implementation doesn't match its documented intent. `usage.promise.catch(() => {})` only adds an empty catch handler—it doesn't reject the promise. If the stream ends without a "finish" chunk, the promise hangs forever. Should call `usage.reject(new Error("Stream ended without finish chunk"))` instead.</comment>

<file context>
@@ -0,0 +1,664 @@
+    flush() {
+      // If stream finishes without a "finish" chunk, reject the promise
+      // so callers don't hang forever.
+      usage.promise.catch(() => {});
+    },
+  });
</file context>

[cubic-dev-ai]

P2: The README instructs users to add the MCP to deploy.json, but deployment is automatic in this repo. This new step conflicts with the documented workflow and will mislead users.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At google-gemini/README.md, line 10:

<comment>The README instructs users to add the MCP to `deploy.json`, but deployment is automatic in this repo. This new step conflicts with the documented workflow and will mislead users.</comment>

<file context>
@@ -0,0 +1,13 @@
+1. Configure your MCP in `server/types/env.ts`
+2. Implement tools in `server/tools/`
+3. Rename `app.json.example` to `app.json` and customize
+4. Add to `deploy.json` for deployment
+5. Test with `bun run dev`
+
</file context>