Vercel AI Docs MCP

What IvanAmador/vercel-ai-docs-mcp MCP server does

IvanAmador/vercel-ai-docs-mcp MCP server gives MCP-compatible assistants access to the Vercel AI SDK documentation. It supports two query styles: an agent query that searches the indexed content and synthesizes an answer, and a direct query that returns similarity-search results from the documentation index.

The server also exposes session memory controls. Agent requests can include a session identifier so related questions retain conversational context. The clear-memory tool removes stored context for one session, or for every session when no identifier is provided.

This project is intended for developers working with the Vercel AI SDK who want documentation answers inside an MCP client rather than manually searching the documentation site.

How it works

The project has separate components for retrieving documentation, building the search index, answering questions, and exposing MCP tools. Its indexing workflow fetches and processes Vercel AI SDK documentation, then stores embeddings in a FAISS vector index. Direct queries search that index using semantic similarity.

Agent queries use the indexed documentation as context for responses generated with a Google Gemini model. The server stores session data in the project’s session files so subsequent requests can continue a conversation. The index must exist before normal querying; the documented troubleshooting step for a missing index is to run the index build command.

Setup and configuration

The IvanAmador/vercel-ai-docs-mcp MCP server requires Node.js 18 or newer, npm, and a Google Gemini API key. Set GOOGLE_GENERATIVE_AI_API_KEY in a .env file at the project root. The key is obtained through Google AI Studio and is used for Gemini model access.

The documented local setup is:

1. Clone the repository and enter the project directory.
2. Run npm install.
3. Compile the TypeScript project with npm run build.
4. Create the documentation index with npm run build:index.
5. Start the server with npm run start.

Claude Desktop and Cursor can run the built server with Node and the path to dist/main.js. Their MCP configurations need the absolute project path and the Google API key in the server environment. The README also identifies npm run dev:index and npm run dev as development scripts.

Tools and capabilities

The MCP server provides three primary tools:

• agent-query accepts a natural-language question and an optional sessionId. It searches the documentation and produces a synthesized answer.
• direct-query accepts a query and an optional result limit, returning direct similarity-search results.
• clear-memory clears one session when given a sessionId, or all sessions when the parameter is omitted.

The indexing components can fetch and process documentation, create the FAISS index, and manage the resulting local files. This makes the server suitable for documentation lookup, API usage questions, and follow-up questions that depend on earlier context.

Limitations and notes

A Google API key is required for the Gemini-backed agent functionality, and Google API rate limits can cause errors. A valid key must have access to the configured Gemini model. The documentation index is a local project artifact and needs to be built before the server can use it for queries.

The README explicitly documents Claude Desktop and Cursor configurations. Other MCP clients may work if they implement the Model Context Protocol, but their setup details are not provided here. Configuration paths must be absolute in the shown client examples, and clients may need to be restarted after configuration changes.