cendor-mcp

The cendor-mcp server provides a read-only implementation of the Model Context Protocol (MCP) for connecting agent-mode AI coding assistants to Cendor's live documentation and API call-shapes. It is designed to support assistants such as Claude Code, Cursor, GitHub Copilot (in agent mode), and Windsurf Cascade by supplying them with accurate, up-to-date information about Cendor libraries, ensuring that generated code aligns with the current published package versions on PyPI and npm. The tool operates in a pull-based manner: the assistant sends queries to the server, which responds with information, but no user code or data is sent to the server, and the server never initiates communication with the client. cendor-mcp offers several specific functions to AI assistants. These include full-text search over Cendor documentation, retrieval of complete documentation pages in markdown, access to the correct API call-shapes for specific symbols (with both correct and common incorrect forms), provision of runnable, CI-typechecked code snippets for tasks, and listing of cookbook recipes that can be copied and run offline. Each response is built from the authoritative documentation source and is stamped with the current published package versions, preventing discrepancies between documentation and available packages. The server can be accessed remotely via a hosted endpoint (https://mcp.cendor.ai) for zero-install integration or run locally for offline use, with all documentation bundled and no network communication required. Local operation is supported via Node (npx) or Python (uvx), allowing flexibility for different development environments. The tool is open source and distributed under the Apache-2.0 license. It is intended as optional developer tooling and is not required by Cendor libraries at runtime. cendor-mcp is particularly aimed at developers integrating or building AI coding assistants that need access to live, accurate documentation and call-shape information for Cendor libraries. Its design emphasizes privacy, as no codebase data leaves the user's machine in local mode, and it operates strictly in agent mode, not supporting inline autocomplete scenarios. The project is maintained by Raghav Mishra (PowerAI Labs) and is available "as is" under the stated license.