Your AI wrote a lot of code. How much of it survived?
AgentTrace maps every Claude Code session to the files it changed, the commits it produced, and whether they stuck — then rolls it into a single survival rate. Plus cost, security flags, and what to fix next. All on your machine, nothing uploaded.
Claude Code edits files, runs commands, opens PRs. Session after session. The transcripts pile up on disk.
Which sessions ran? What did each one change? Did that hour of edits get committed — or silently thrown away? How much did it cost?
It reads the transcripts already on your disk, matches them to your commits, and shows you what landed, what it cost, and how to make the next run better.
It reads what’s already on your disk.
No agent, no instrumentation, no config. AgentTrace discovers your Claude Code transcripts, parses them, and correlates each one with git — so the loop runs on data you already have.
Finds every Claude Code session transcript under ~/.claude/projects. Nothing to install into your agent.
Matches each session to the files it touched and the commits that followed — then checks what’s still on disk.
Rolls outcomes into a survival rate, tallies token spend, and flags risky shell actions or leaked secrets.
Surfaces concrete recommendations — add a skill, trim context, fix a risk — and tracks the ROI you bank.
Free. Two ways in. Pick yours.
The Mac app is signed with a Developer ID certificate and notarized by Apple, so it opens without Gatekeeper warnings. The CLI is a zero-dependency install on Node 20+.
Version 0.1.0 · SHA-256 checksums · All releases
A native window over the full dashboard. Signed & notarized.
Node ≥ 20. Lists sessions, drills into one, or starts the local dashboard.
Most tools count what your agent did. AgentTrace tells you what actually lasted.
Every session is correlated with your git history and labeled Landed, Reverted, Not committed, Not in git, or No edits — then rolled into one number. Token counters can’t tell you this.
Zero backend, zero telemetry. It reads the Claude Code transcripts already on disk and serves a dashboard on 127.0.0.1. The server even rejects any non-loopback request.
Per-session token usage and USD estimates, plus flags for risky shell actions and plaintext secrets an agent ran — before they cost you money or a breach.
Prefer a window? Download the signed, notarized .dmg. Live in the terminal? `npx agenttrace-cli`. Same engine underneath, so your team can use whichever fits.
Everything, running locally.
- Per-session survival rate rolled up across your whole history.
- Git-correlated outcome labels: Landed / Reverted / Not committed / Not in git / No edits.
- Files-still-on-disk check so a “landed” session really landed.
- Per-session input / output / cache token breakdown.
- USD spend estimate from published model rates.
- Total spend across every session, at a glance.
- Flags critical shell actions and plaintext secrets an agent ran.
- Local heuristics surface repeated commands (→ add a skill) and context bloat.
- Proven-ROI tracking: mark a recommendation done and bank the value.
- Group sessions your way (Personal / Business / anything) and rename them.
- Searchable, sortable session list with per-session drill-down.
- An interactive chart — switch the metric, break down by session, group, or project.
- `agenttrace analyze <id>` prints a Markdown or JSON brief to pipe back to your agent.
- A local dashboard on 127.0.0.1 with a per-session API.
- Turn any session outcome into a tamper-evident Ed25519 receipt with ActionProof.
- Verified locally — no server, no account.
One free hub. Focused lenses on the same local data.
The free hub: every Claude Code session, mapped to the files it changed, the commits it produced, and whether the work survived.
The cost lens: per-session token usage and USD spend from published model rates, so you see where the money goes.
The security lens: flags risky shell actions and plaintext secrets an agent ran, before they bite you.
The proof lens: turn any session outcome into a tamper-evident Ed25519 receipt, verified locally.
Straight answers.
Does AgentTrace send my code or sessions anywhere?+
No. It is local-first with zero backend and zero telemetry. It reads the Claude Code transcripts already on your machine (~/.claude/projects), correlates them with your local git history, and renders everything in your terminal or a dashboard on 127.0.0.1. Nothing is uploaded.
What is a “survival rate”?+
Every session is correlated with git and labeled Landed, Reverted, Not committed, Not in git, or No edits — then rolled into one number: the share of your agent’s work that actually survived. Plenty of tools count tokens; AgentTrace tells you whether the work stuck.
Is the Mac app safe to install?+
Yes. The .dmg is signed with a Developer ID certificate, notarized by Apple, and stapled — so macOS Gatekeeper opens it without warnings. SHA-256 checksums are published alongside each release.
Do I have to use the desktop app?+
No. AgentTrace is also a zero-dependency CLI: run `npx agenttrace-cli` (Node ≥ 20) to list sessions, drill into one, or start the local dashboard. The Mac app is just a native window over the same engine.
Which agents does it work with?+
Claude Code today — it reads the session transcripts Claude Code writes to disk. Support for other coding agents is on the roadmap.
How much does it cost?+
AgentTrace is free and open. It’s the free hub of a family that includes focused lenses — TokenKeeper (cost), SessionSentry (security), and ActionProof (cryptographic proof).
It’s free, it’s local, and it reads sessions you already have.
Download the Mac app or run one command. In under a minute you’ll see what your agents actually accomplished — and how much of it survived.