MCAP MCP Server

Query your robot's MCAP recordings with SQL — straight from your LLM.

Setup

{
  "mcpServers": {
    "mcap-query": {
      "command": "uvx",
      "args": ["mcap-mcp-server[all]"]
    }
  }
}

No install, no database, no API keys. Requires uv.

Related MCP server: GraphJin

Usage

Just talk to your LLM:

• "Tell me what topics are in session_003.mcap"

• "In session_017.mcap find all moments where voltage dropped below 22V"

• "Correlate IMU acceleration with motor current."

• "Compare average battery voltage across my last 5 runs"

• "What version of mcap-mcp-server am I running? Update it"

Tools

Project documentation — configuration, Docker, development setup, and architecture.

Example SQL (under the hood)

-- Time-windowed stats
SELECT (timestamp_us / 1000000) as second,
       AVG(voltage) as avg_v, MIN(voltage) as min_v
FROM battery GROUP BY second ORDER BY second

-- Cross-sensor correlation via ASOF JOIN
SELECT b.timestamp_us, b.voltage, i.linear_acceleration_x
FROM battery b ASOF JOIN imu i ON b.timestamp_us >= i.timestamp_us

-- Multi-recording comparison
SELECT 'run1' as run, AVG(voltage) FROM r1_battery
UNION ALL
SELECT 'run2', AVG(voltage) FROM r2_battery

Performance

Metadata tools (list_recordings, get_recording_info, get_schema) return in < 1 ms regardless of file size. SQL queries execute in 1–20 ms once data is loaded. The one-time load_recording cost scales with file size:

Measured on Apple M4 with JSON-encoded messages, 5 fields per message. Query times are median across aggregation, filter, and window function queries. Memory is the DuckDB in-memory footprint (default budget: 2 GB).

Tip: use topics and start_time/end_time filters on load_recording to load only what you need.

Update

uvx mcap-mcp-server[all] --upgrade

Or ask your LLM — the get_version tool returns the running version and the upgrade command.

License

GNU General Public License v3.0 — see LICENSE.