See what your agent actually did.
Drop in a Claude Code session. See every action it took and what went wrong.
Sessions live in ~/.claude/projects/. On a Mac, press Cmd+Shift+. in the picker to see hidden folders.
curl -fsSLO https://stacktrace.ai/trace/share-session.py && python3 share-session.py- Finds your newest Claude Code session and opens it on this page.
- The file goes from your computer straight to this tab, never to our servers. The script stops after one read, or two minutes. Inspect
share-session.py.
Try a sample.
No session handy? Each sample is a short session with one known problem.
-
Did my agent leak a secret?
Asked to fix a deploy, it read .env.staging and sent the token to an outside API.
HighYou'll see: a token sent to an outside server
-
Did it delete outside the project?
Asked to fix a flaky build, it deleted a cache in a folder next door and, with sudo, a global install.
HighYou'll see: a delete outside the project
-
Did it switch credentials mid-session?
The Linear connector call failed first, then the agent reconnected with a different token.
MediumYou'll see: a connector that recovered on new credentials
-
Is it stuck in a loop?
It reran the failing tests four times before it checked whether the database was even up.
LowYou'll see: the same call retried, unchanged
-
Why does it keep refusing?
A safety check blocked a password rotation three times in a row.
HighYou'll see: blocked by a safety check
-
Why did it stop halfway?
It tagged the release, then hit the team's spend limit.
HighYou'll see: stopped by the spend limit
How it works.
- Engine loads.
stacktrace-cli0.9.0 on WebAssembly. Served from stacktrace.ai, sha256-pinned. - Network removed. The worker deletes
fetch, XHR and WebSocket. The results frame runs underconnect-src 'none'. - File stays in memory. Read via the File API into the worker. Never written or sent.
- Same as the CLI. The detector and page of
stacktrace monitor.
- Verify it. Open developer tools, then Network. No requests after the engine loads.
- What we count. That a check ran, the file's size band, and which rules matched. Never the file.
- Not in the browser. Model-reviewed rules, vulnerability lookups and the Inventory tab. Use the CLI.
Get this on every session, automatically.
Stacktrace runs these checks while your agents work and tells you in Slack or Claude Code as it happens.
Try Stacktrace free- Starting Python in your browser
- Loading the Stacktrace engine (about 20 MB, once)
- Switching off the engine's network access
- Reading your session, on your computer
- Checking what your agent did