AgentTrace

How much of your AI's code actually survived? AgentTrace correlates every Claude Code session with your git history and gives you one number: the share of your agent's work that landed and stuck — versus what got reverted or never committed. Local-first, zero backend, zero telemetry.

You run coding agents all day, but you can't see whether their work lasts. That hour of edits — did it get committed and survive, or silently thrown away? AgentTrace reads the session transcripts already on your disk, matches them to commits, and shows you a survival rate — plus what each session changed, what it cost, and how to make the next one better. All in your terminal or a local dashboard, without sending anything anywhere.

Install

Mac app — download the latest .dmg (signed & notarized; Apple Silicon and Intel).

Command line — requires Node ≥ 20:

npx agenttrace-cli        # list your sessions, newest first

The installed command is agenttrace (the npm package is agenttrace-cli). Or clone and run from source (this path uses Node's TypeScript type-stripping, so it needs Node ≥ 23.6 — no build step):

git clone https://github.com/Burakfenerci5/agenttrace && cd agenttrace
node src/cli.ts           # same as `npx agenttrace-cli`

Quick start

agenttrace                # list every session, with outcome + cost
agenttrace show <id>      # drill into one session (id prefix is enough)
agenttrace serve          # open the local dashboard (http://127.0.0.1:4317)
agenttrace analyze <id>   # print an agent-readable analysis (pipe it to your agent)

Everything runs against ~/.claude/projects on your machine. Nothing is uploaded.

What you get

• Outcome labeling — each session is correlated with git and labeled Landed, Reverted, Not committed, Not in git, or No edits — and rolls them into a single survival rate. Plenty of tools count your tokens; AgentTrace is the one that tells you, per session, whether the work actually survived.
• Cost & tokens — per-session token usage and a USD estimate from published model rates, so you can see where your spend goes.
• Recommendations — local heuristics spot repeated commands (→ add a skill), context bloat (→ trim cache reads), marathon sessions, and uncommitted work. Each comes with a ready-to-paste prompt you can hand straight back to your agent.
• Actions — every meaningful step (file writes, shell commands, deploys) is listed; high-consequence ones (pushes, publishes, rm -rf, payments) are flagged as worth verifying.
• A striking local dashboard — searchable, groupable, an interactive metric chart, multi-session selection to compare cost, and per-session drill-down.

Prove it, don't just observe it

AgentTrace shows you what your agents reportedly did (heuristics over local logs). Its sibling ActionProof lets you prove it: Ed25519-signed, tamper-evident receipts for any action or session, verifiable offline. If ActionProof is installed alongside AgentTrace, the dashboard's Sign button mints a real receipt and shows a ✓ verified badge — the free-to-provable bridge, built in.

Privacy

• Binds to 127.0.0.1 only; rejects any non-loopback Host and cross-origin request.
• Never reads your message content — only event metadata (tools, files, timing, usage).
• No network calls, no analytics, no account. Your transcripts never leave your machine.

Development

npm run typecheck    # tsc --noEmit
npm test             # node:test suite

License

MIT