Codewiki MCP

What izzzzzi/codewiki-mcp MCP server does

The izzzzzi/codewiki-mcp MCP server gives an AI assistant access to repository information from codewiki.google. It is intended for open-source repositories that the service has indexed and documented. The server supports three main workflows:

• Find repositories with a text search and receive matching names, URLs, descriptions, and result metadata.
• Retrieve generated wiki documentation either as one aggregate markdown response or as structured pages.
• Ask a natural-language question about a repository and receive an answer from the service.

The fetch and ask operations accept an owner/repo identifier, a GitHub URL, or a natural-language repository description. This makes the server useful when a user knows what a project does but does not know its exact repository path.

How it works

An MCP client sends tool calls to the local server. The server then communicates with codewiki.google over HTTPS and returns the resulting data through MCP. The upstream service uses Google’s batchexecute RPC protocol rather than a conventional REST or GraphQL API; the server constructs the request, removes the response prefix, parses the returned frames, and extracts the payload for each operation.

Natural-language repository resolution uses wink-nlp to extract keywords. When needed, those keywords are sent to the GitHub repository search endpoint, with the top result normalized into a repository URL for codewiki.google. This resolution path is useful for inputs such as a description of a framework rather than a precise repository name.

Requests can retry automatically after server-side 5xx responses, using increasing delays. Client-side 4xx responses are not retried. Successful responses include metadata such as the transferred byte count and elapsed time, while failures use codes including validation, RPC failure, timeout, and NLP resolution errors.

Setup and configuration

The package can run directly through npm without a separate installation:

npx -y codewiki-mcp@latest

By default, it uses stdio transport. The command can instead start Streamable HTTP with --http, or SSE with --sse; both modes accept --port, and the HTTP endpoint can be changed with --endpoint. The documented default port is 3000 and the default endpoint is /mcp.

The README provides configurations for Cursor, Claude Desktop, Windsurf, and VS Code using the same npx command. It also documents a source workflow that clones the repository, installs npm dependencies, builds the project, and runs node dist/cli.js.

Optional environment settings include CODEWIKI_REQUEST_TIMEOUT, CODEWIKI_MAX_RETRIES, and GITHUB_TOKEN. The provided material does not identify any environment variable as mandatory for starting the server.

Tools and capabilities

codewiki_search_repos requires a search string and accepts a result limit from 1 to 50, defaulting to 10.

codewiki_fetch_repo requires a repository identifier. Its mode defaults to aggregate, returning full markdown; pages returns structured JSON page data instead.

codewiki_ask_repo requires a repository identifier and a question. An optional history array supports up to 20 earlier role/content messages, allowing follow-up questions to retain conversation context.

The izzzzzi/codewiki-mcp MCP server supports stdio, Streamable HTTP, and SSE transports. Docker usage is also documented, including HTTP startup and passing configuration through environment variables.

Limitations and notes

Results depend on codewiki.google’s indexed repository coverage and generated documentation. The server is a client for that service; it does not itself generate a repository wiki. Natural-language lookup also depends on keyword extraction and the GitHub Search API, so an imprecise description may fail to resolve or select an unexpected top result.

The upstream RPC interface is internal to codewiki.google rather than a documented REST or GraphQL API. Calls may therefore return service-specific RPC errors, including 404 failures, timeouts, or NLP resolution failures. The izzzzzi/codewiki-mcp MCP server reports response size and timing metadata, but the supplied material does not describe caching, offline operation, or support for private repositories.