What is MCP Server?
The Model Context Protocol (MCP) is an open standard that allows AI tools like Cursor, Windsurf, and Claude Desktop to connect to external data sources and call them as structured tools during a conversation. Kelu exposes a per-knowledge-base MCP server that makes your documentation queryable from any MCP-aware environment.
How Kelu works with MCP Server
Everything you can do once the integration is connected.
Per-knowledge-base MCP endpoint
Kelu serves MCP at POST https://app.kelu.dev/mcp, authenticated with a client key (X-Client-Key header) that names the knowledge base.
Search, ask_question, list_sources and feedback tools
The MCP server's tools include a search tool named after your knowledge base (search_docs also works) for hybrid semantic search, ask_question for a full RAG-grounded answer, list_sources for metadata about indexed content, and feedback for reporting wrong answers.
Works with Cursor, Windsurf, and Claude Desktop
Add your Kelu MCP server to any MCP-aware IDE or AI assistant. Developers can query your docs inline without leaving their editor.
Client key authentication
Secure the MCP endpoint with a client key that has Allow MCP access, or create a Public server whose readers sign in with Google or GitHub. Key-authenticated and Public servers never return restricted documents.
How to set it up
- 1
In Kelu, open the knowledge base's Integrations tab > MCP Server. Under Create a server, type the end of its address (for example acme), pick a sign-in type, and click Create.
- 2
Under Install it, pick your tool (Claude Code, Cursor, VS Code, or Codex) and copy the command or config. For an API key server it already contains your client key; a Public server signs each reader in on first use.
- 3
Test by asking a question in your IDE — the AI will call the Kelu search tool or ask_question and return a grounded answer from your docs.
Related integrations
Web Widget
The Kelu web widget embeds in any HTML page. One script tag, full style customization, mobile-ready, WCAG AA accessible.
React Component
Install @keluhq/react from npm. Drop in the <KeluChat /> panel, or build your own UI with the useKeluChat hook.
JavaScript SDK
Install @keluhq/sdk from npm for chat, search, and feedback in browsers, React Native, and Node.js 18 or later.
Got questions?
MCP Server — Frequently Asked Questions
Common questions about connecting Kelu to MCP Server.
Kelu's MCP server works with any tool that implements the Model Context Protocol — including Cursor, Windsurf, and Claude Desktop. If your IDE or AI assistant supports MCP server configuration, you can connect it to Kelu by adding the endpoint URL and X-Client-Key header.
In Cursor's settings, navigate to the MCP server configuration section and add a new server entry with the Kelu endpoint URL (https://app.kelu.dev/mcp) and set the X-Client-Key header to your knowledge base's client key — the key is what selects the knowledge base. Cursor will then be able to call the search tool (named after your knowledge base) and ask_question during conversations.
The MCP server's tools include a search tool named after your knowledge base, such as search_acme_api_knowledge_sources (search_docs also works), for hybrid vector + full-text semantic search returning ranked chunks, ask_question (full RAG-grounded answer with citations from your knowledge base), list_sources (metadata about all indexed content in the knowledge base), and feedback (records a wrong answer or missing documentation). A get_documents tool is off by default.
With a client key, sent in the X-Client-Key header or as Authorization: Bearer. Create one in the Kelu dashboard under the knowledge base's Integrations tab > Client Keys, with Allow MCP access ticked (the default). The key belongs to one knowledge base. A Public server created under Integrations > MCP Server needs no key: each reader signs in with Google or GitHub on first use. Key-authenticated and Public servers never return restricted documents.
Yes. The MCP endpoint follows the Model Context Protocol specification, which uses JSON-RPC 2.0 over HTTP POST. Any MCP client that sends properly formatted JSON-RPC requests with the correct authentication header will receive structured tool results from your Kelu knowledge base.