MCP & AI Agents
MCP & AI Agents
Indexed integrates with AI assistants through the Model Context Protocol (MCP) — an open standard that lets AI tools use external data sources and services.
What Is MCP?
MCP (Model Context Protocol) is a standard created by Anthropic that lets AI assistants like Claude Desktop, Cursor, and Cline call external tools. Instead of the AI only knowing what's in its training data and your current conversation, MCP lets it reach out to specialized services for additional context.
Think of it like this: without MCP, asking Claude about your team's deployment process gets you a generic answer. With Indexed as an MCP server, Claude searches your actual documentation and Jira tickets and answers with your team's real procedures.
How Indexed Uses MCP
Indexed runs as an MCP server that exposes two tools:
search(query)
Searches across all your collections. The AI assistant sends a natural language query, and Indexed returns the most relevant chunks from your indexed documents.
search_collection(collection, query)
Searches a specific collection. Useful when the AI assistant wants to target a particular source — for example, searching only Jira tickets or only Confluence pages.
Transports
MCP supports different transport methods for communication between the AI client and the Indexed server:
| Transport | Use case | Command |
|---|---|---|
| stdio (default) | Claude Desktop, Cursor, Cline — local single-user | indexed mcp run |
| HTTP | Docker, remote access, shared servers | indexed mcp run --transport http --port 8000 |
| SSE | Server-Sent Events for streaming | indexed mcp run --transport sse |
| Streamable HTTP | Newer HTTP-based streaming | indexed mcp run --transport streamable-http |
For most users, stdio is the right choice. It's what Claude Desktop and most editors expect. Use HTTP when running Indexed in Docker or on a remote server.
The MCP Inspector
Indexed includes a development mode that launches the MCP Inspector — a web UI for testing your MCP tools:
indexed mcp devThis is useful for debugging your setup before connecting an AI client.
How AI Assistants Decide to Search
When you connect Indexed to Claude Desktop (or another MCP client), the AI assistant can see the available tools and their descriptions. It decides whether to use them based on your conversation:
- If you ask "What's our deployment process?", Claude recognizes this might be in your indexed docs and calls
search("deployment process") - If you ask "What Jira tickets are about rate limiting?", Claude might call
search_collection("eng-tickets", "rate limiting") - If you ask "What's 2 + 2?", Claude won't search — it doesn't need your docs for that
The AI makes these decisions automatically. You don't need to explicitly tell it to search.