AI assistants

The extension ships an MCP server that lets AI assistants record captures and query trace data: timelines, events, user traces, priority inversions, SQL over a recording. An assistant that speaks MCP can then answer "why is this task late" from the recording itself.

Connecting

• GitHub Copilot agent mode picks the server up automatically: the extension registers it as an MCP server definition provider.
• Claude Code, Cursor, Claude Desktop and anything else that speaks MCP: run BKPT Debug: Connect AI Assistant and pick your client. The command writes or shows the configuration the client needs (the server is a stdio process bundled with the extension), and copies it to the clipboard where the client has no config file to write.

What the assistant can do

The server exposes the va_* tools: list the connected probes and the known target names, list recordings, record a capture with a given connection (hardware trace included: PC sampling, exception trace, DWT counters and data watches, ITM ports), poll firmware symbols, snapshot the RAM ring, and query a recording for its summary, timeline, events, user traces, priority inversions, CPU statistics, timers, communication paths, verdicts, derived series, ITM console text, DWT watches, exceptions and counters, SWO load, the PC profile, and raw SQL. It can also check a hardware-trace setup without a target (the register image a capture would program), distil a recording into a fingerprint baseline and gate later runs against it. Everything goes through the same engine and the same JSON the command line produces, so what the assistant sees is what the views show.

Tips

• Ask for the summary first; a recording's event count and losses tell the assistant whether the capture is worth analysing.
• The assistant records with the connection settings you give it; keep a .vacf per board in the project so it never has to guess.