staxtrace

Offline, local-first observability and memory toolkit for AI coding agents.

staxtrace ingests session logs from 20 coding-agent providers into one local SQLite store, then builds four pillars on top of it. Local-first from the first commit: everything runs on your machine, no account, no telemetry, nothing leaves ~/.stackunderflow/.

The four pillars

• Cost analytics — what a run actually cost, broken down by tool, model, project and day, summed from the session logs rather than from a context meter
• Time-travel playback — step-by-step filesystem reconstruction of any past session
• Agent memory — a local memory layer your coding agents query mid-task, so a lesson learned in one session is available in the next
• Chat sidebar — an offline conversation over your own history

Why

The Schmidhuber build shipped with a token count read off the wrong meter. Getting the real number meant walking every JSONL by hand. staxtrace exists so that costs a command instead of an afternoon.

I wrote about the thinking behind this in building-stackunderflow.

Links

• GitHub
• Blog post