MCP server

raglite mcp serves your documents to any MCP client over stdio. The agent can then search them, list them, and (when you give it an LLM) ask questions about them.

Claude Code

TypeScript

npx fetches everything it needs, including the MCP SDK and local embeddings:

claude mcp add docs -- npx -y -p raglite-toolkit -p @modelcontextprotocol/sdk -p @huggingface/transformers \
  raglite mcp /absolute/path/to/docs --store-dir /absolute/path/to/raglite-index

Python

pip install 'raglite-toolkit[mcp,local]'
claude mcp add docs -- raglite mcp /absolute/path/to/docs --store-dir /absolute/path/to/raglite-index

Claude Desktop, Cursor and VS Code

They use the same command in their MCP settings file:

TypeScript

{
  "mcpServers": {
    "docs": {
      "command": "npx",
      "args": [
        "-y", "-p", "raglite-toolkit", "-p", "@modelcontextprotocol/sdk", "-p", "@huggingface/transformers",
        "raglite", "mcp", "/absolute/path/to/docs", "--store-dir", "/absolute/path/to/raglite-index"
      ]
    }
  }
}

In a project that already has raglite-toolkit and @modelcontextprotocol/sdk installed, npx raglite mcp ./docs --store-dir ./.raglite is enough.

Python

{
  "mcpServers": {
    "docs": {
      "command": "raglite",
      "args": ["mcp", "/absolute/path/to/docs", "--store-dir", "/absolute/path/to/raglite-index"]
    }
  }
}

If raglite is not on the client’s PATH, use its full path, for example /path/to/.venv/bin/raglite.

• Use absolute paths, and pass --store-dir. MCP clients may start the server in any directory, and the index is kept in ./.raglite by default.
• Run the command once in a terminal first. The first start indexes the documents (and, with npx, installs packages), which can take longer than a client waits. Stop it with Ctrl+C once it is ready; later starts reuse the index.
• Add --embed-provider, --embed-model and --embed-key to use a hosted embedding model, as with raglite index.

Tools

Tool	Inputs	Returns
search	query; optional topK (1–50, default 5) and mode (vector, keyword, hybrid)	JSON passages: source, chunk, page, section, score, text
list_sources	none	JSON list: source (file name or URL), path, chunks
ask	question; optional topK and mode	The answer, citing passages as [1], [2]

ask is offered only when the server has an LLM: pass --llm-provider (and --llm-model, --llm-key as needed). Without one, the client’s own model reads the search results, which is usually what you want.

--mode sets the default retrieval mode for both tools.

From code

TypeScript

import { DocumentCollection, createMcpServer, serveMcp } from "raglite-toolkit";

const docs = new DocumentCollection("./docs");
await docs.build();

await serveMcp(docs); // stdio, until the client disconnects

// Or get the McpServer and connect your own transport:
const server = await createMcpServer(docs, { llm: { provider: "anthropic" } });

Needs npm install @modelcontextprotocol/sdk.

Python

from raglite import DocumentCollection, create_mcp_server, serve_mcp

docs = DocumentCollection("./docs")
docs.build()

serve_mcp(docs)  # stdio, until the client disconnects

# Or get the MCPServer and run another transport:
server = create_mcp_server(docs, llm={"provider": "anthropic"})

Needs the [mcp] extra.

stdout carries the MCP protocol, so RAGLite’s logs go to stderr.