Detrix is an MCP server that gives AI coding agents (Claude Code, Cursor, Windsurf) a debugger they can control directly. The agent sets observation points on variables in a running process, captures values at runtime, and queries results when ready — no restarts, no code changes at debug time, no print statements.

Here's what that looks like in Claude Code:

  me:    "The order-service is reporting inflated revenue."

  claude: sets observation points on txn.Amount, txn.Unit, total
          → detrix wake order-service:8091
          → detrix observe txn.Amount, txn.Unit, total
          → detrix query_metrics

  claude: "The pricing API changed. It now returns cents for some
           amounts. The code fetches txn.Unit but discards it."

I didn't tell it what to observe. It read the source, picked the variables, watched them live, and found the root cause.

Under the hood it's DAP logpoints — most debuggers can already evaluate expressions without stopping the target. Detrix exposes that over MCP. Pull-based: nothing streams into context, the agent queries when ready.

Overhead is near-zero — logpoints are non-breaking, no pauses, configurable throttling on hot paths. Expression safety layer with allow/block lists prevents accidental exposure of secrets or API keys.

Works locally and remotely — Detrix server runs alongside your service in Docker or on a remote host, so the agent on your laptop observes a process running in the cloud.

Python (debugpy), Go (Delve), Rust (lldb-dap). Ships with a Claude Code skill that makes Claude reach for Detrix before suggesting print statements.

Built this after debugging a trading bot migration where the broker silently changed their price format — the add-prints-restart-wait loop on timing-sensitive code is what Detrix eliminates.

13 Rust crates, 2,300+ tests. MIT licensed. Docker demo with video in `examples/docker-demo/`.

Would love feedback on the DAP approach — are there edge cases in your stack where this would break down?