[Feature Request] Native MCP Client Support via SSE/HTTP Tunnels in Google AI Studio

Google AI Studio
1

Native MCP (Model Context Protocol) Client Integration via SSE / HTTP Tunnels

Component: Google AI Studio > Tools & Function Calling > External Connectors
Type: Feature Request (FR)
Priority: P2 / Severity: S2
Reporter: Community Developer & Gemini Agent Collaborative Request

  1. PROBLEM STATEMENT & EXECUTIVE SUMMARY

Currently, Google AI Studio supports Function Calling, Google Search, and Code Execution. However, testing custom external tools requires developers to manually author and maintain JSON schemas (OpenAPI/declarative tool definitions) and manually route execution back and forth.

Meanwhile, the open-source Model Context Protocol (MCP) has emerged as the de facto industry standard for LLM tooling (adopted by Anthropic, Cursor, Zed, Sourcegraph, OpenAI).

Developers running local environments (CLI agents, local databases, dev-servers) lack a frictionless way to connect their live local tools directly to Gemini 2.0 Flash / Pro in AI Studio during prototyping.

  1. PROPOSED SOLUTION (High-Level PRD)

Introduce a native “Add MCP Server” capability within the AI Studio “Tools” sidebar:

  1. Allow developers to specify an MCP Server Endpoint (HTTP/HTTPS URL supporting Server-Sent Events / SSE, compatible with Cloudflare Tunnel, ngrok, Loophole, or public relays).
  2. AI Studio acts as an MCP Client: performs discovery handshake (initialize → tools/list), automatically populates Gemini’s system instructions with declared tool signatures.
  3. During inference, when Gemini emits a tool call, AI Studio dispatches tools/call over the tunnel, captures the SSE stream/result, and injects the output back into the context loop.
  1. USER JOURNEY & WORKFLOW
  1. Developer runs a local MCP server exposing local tools (e.g., File Relay, SQLite, Git, Terminal) and opens an HTTPS tunnel:
    npx @modelcontextprotocol/server-filesystem /path/to/dir → https://xyz.ngrok-free.app/sse
  2. Developer navigates to Google AI Studio > “Tools” dropdown > Clicks “+ Add Custom MCP Server”.
  3. Developer enters:
    • Endpoint URL: https://xyz.ngrok-free.app/sse
    • Optional Auth Header: Bearer <token>
  4. AI Studio tests connection, fetches tools schema, and displays active tools with toggle switches.
  5. In chat, Gemini autonomously invokes local MCP tools with zero manual schema authoring.
  1. STRATEGIC & COMPETITIVE VALUE (BRD Context)
  • Ecosystem Moat & Parity: Aligns Google AI Studio with modern multi-agent development stacks.
  • Showcasing Gemini 2.0 Strengths: Gemini’s massive 1M–2M token context window and high speed make it uniquely suited to ingest large MCP tool payloads (full file trees, deep logs, database outputs) without hitting token limits that degrade competing models.
  • Developer Velocity: Cuts prototyping time for multi-agent workflows from hours (manual bridging) to under 60 seconds.
  1. FUNCTIONAL REQUIREMENTS & ARCHITECTURE (FRD Context)

[FR-1] Protocol Compliance:

  • Support MCP Specification (version 2024-11-05 and newer) over HTTP/SSE transport.
  • Support core primitives: tools/list, tools/call. (Optional Phase 2: resources/read, prompts/get).

[FR-2] Security & Networking:

  • Validate outgoing HTTPS certificates for tunneled endpoints.
  • Configurable per-call timeout (default: 30s) to prevent hanging inference threads.
  • Explicit user consent banner: “External tools may send prompts and execute actions on external servers.”
  • Support custom HTTP Authorization headers (Bearer / API-Key).

[FR-3] UI/UX Integration:

  • Visual badge indicating connected MCP servers and active tool counts.
  • Collapsible inspect panel showing the raw JSON-RPC request/response payload for debugging.
  1. ACCEPTANCE CRITERIA
  • User can paste a standard MCP SSE endpoint into AI Studio settings.
  • AI Studio successfully queries tools/list and registers functions into Gemini’s tool registry.
  • Gemini can trigger a tool call, and the resulting payload is correctly returned to the model.
  • Clear error toasts displayed if tunnel drops, returns non-200, or times out.