mcp — MCP Server

Expose bagx analysis to LLM agents (Claude Code, Claude Desktop, Cursor, etc.) via the Model Context Protocol.

Unlike bagx ask, no LLM API key is required. bagx returns structured JSON; the host agent does the reasoning.

Install

pip install bagx[mcp]

Start the server

bagx mcp

This speaks MCP over stdio (default for local agent integrations).

Tools

Tool	Description
eval_bag(path)	Quality eval → overall score, findings, recommendations
list_topics(path)	Topic names, types, counts, bag duration
detect_anomalies(path, topic?)	GNSS/IMU/rate anomaly report
compare_bags(path_a, path_b)	Side-by-side metric comparison
query_messages(path, topic, start_ns?, end_ns?, limit?)	Decoded messages for one topic

Claude Code

claude mcp add bagx -- bagx mcp

Or add manually to .mcp.json:

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

Use the full path to bagx if it is not on PATH inside Claude Code.

Claude Desktop

Edit claude_desktop_config.json:

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

macOS config path: ~/Library/Application Support/Claude/claude_desktop_config.json

Example agent workflow

1. list_topics — discover /imu, /scan, /odom
2. eval_bag — get readiness score and domain recommendations
3. detect_anomalies — find IMU spikes or GNSS fix loss
4. query_messages — inspect decoded fields around a suspect timestamp

ROS1 bags

Install ROS1 reader support alongside MCP:

pip install bagx[mcp,ros1]

Then point tools at .bag files the same way as .db3 / MCAP directories.