hjess/fil

Fork 0

Files

Henrik Jess Nielsen b4c07d3693

Deploy fil (kreuzberg) / deploy (push) Successful in 49s

Details

Nomad changes

2026-06-01 23:40:55 +02:00

5.1 KiB

Raw Permalink Blame History

MCP Integration v5.0.0

Kreuzberg speaks Model Context Protocol. That means any AI agent — Claude, Cursor, a custom LangChain pipeline — can extract documents, generate embeddings, and manage caches through a standard tool interface without writing extraction code.

Two commands to get started:

pip install "kreuzberg[all]"
kreuzberg mcp

That's it. You now have an MCP server running over stdio, ready for any compatible client.

How It Works

The MCP server wraps Kreuzberg's extraction engine behind standard tools, running as a child process over stdin/stdout with JSON-RPC messages — no HTTP ports or configuration needed.

flowchart LR
    A["AI Agent\n(Claude, Cursor, etc.)"] -->|"JSON-RPC\nover stdio"| B["kreuzberg mcp"]
    B --> C["Extraction Engine"]
    B --> D["Embedding Engine"]
    B --> E["Cache Layer"]

Server Modes

Stdio (Default)

The standard mode for local AI tools. The agent spawns kreuzberg mcp as a subprocess and communicates over pipes.

kreuzberg mcp
kreuzberg mcp --config kreuzberg.toml

This is what Claude Desktop, Cursor, and most MCP clients expect.

HTTP Transport

!!! Info "Feature flag: mcp-http" HTTP transport requires the mcp-http feature flag at build time.

For remote deployments or multi-client setups where stdio doesn't work — shared servers, team environments, cloud-hosted agents — HTTP transport exposes the same tool interface over the network.

Tools

Kreuzberg exposes 13 tools via MCP. All extraction tools accept an optional config object to override defaults:

Extraction: extract_file, extract_bytes, batch_extract_files, detect_mime_type, extract_structured Embeddings: embed_text Chunking: chunk_text Cache: cache_stats, cache_clear, cache_manifest, cache_warm Metadata: list_formats, get_version

extract_structured requires the server to be built with the liter-llm feature. Full parameter schemas are discoverable at runtime via the MCP client's list_tools call.

Connecting AI Tools

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "kreuzberg": {
      "command": "kreuzberg",
      "args": ["mcp"]
    }
  }
}

Restart Claude. Kreuzberg's tools appear automatically — ask Claude to "extract text from invoice.pdf" and it will call extract_file behind the scenes.

Cursor

Add to .cursor/mcp.json in your project root:

{
  "mcpServers": {
    "kreuzberg": {
      "command": "kreuzberg",
      "args": ["mcp"]
    }
  }
}

Python MCP Client

For building custom agent pipelines, use the official mcp Python SDK:

import asyncio
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

async def main() -> None:
    server_params = StdioServerParameters(
        command="kreuzberg", args=["mcp"]
    )

    async with stdio_client(server_params) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()

            tools = await session.list_tools()
            print(f"Available: {[t.name for t in tools.tools]}")

            result = await session.call_tool(
                "extract_file",
                arguments={"path": "document.pdf"},
            )
            print(result)

asyncio.run(main())

Spawning from Python

If your application manages the server lifecycle directly:

import subprocess

process = subprocess.Popen(
    ["python", "-m", "kreuzberg", "mcp"],
    stdout=subprocess.PIPE,
    stderr=subprocess.PIPE,
)
print(f"MCP server running (PID {process.pid})")

Configuration

Pass a TOML config file to set extraction defaults for all tools:

kreuzberg mcp --config kreuzberg.toml

Individual tool calls override file defaults via a config parameter. See ExtractionConfig Reference for all available fields.

Running in Docker

docker run ghcr.io/kreuzberg-dev/kreuzberg:latest mcp

docker run \
  -v $(pwd)/kreuzberg.toml:/config/kreuzberg.toml \
  ghcr.io/kreuzberg-dev/kreuzberg:latest \
  mcp --config /config/kreuzberg.toml

For production, use Compose with a persistent cache volume so embedding models don't re-download on restart:

services:
  kreuzberg-mcp:
    image: ghcr.io/kreuzberg-dev/kreuzberg:latest
    command: mcp --config /config/kreuzberg.toml
    volumes:
      - ./kreuzberg.toml:/config/kreuzberg.toml:ro
      - cache-data:/app/.kreuzberg
    restart: unless-stopped

volumes:
  cache-data:

5.1 KiB Raw Permalink Blame History