Credence · Free and open source
Check what your AI agent says it did.
Credence reads your Claude Code, Codex and Cursor sessions, finds every claim the agent made (“all tests pass”, “it’s live”, “I fixed the bug”) and checks each one against what actually happened. It runs on your machine.
curl -fsSL https://tokenmanagementgroup.com/credence/install.sh | sh
Paste it into a terminal and answer two questions. Your reports open in the browser, and Claude Code shows a one-line check under each reply that makes claims.
← All sessions
Add CSV export to the billing page
The agent claimed 5 things. 2 are backed by the session record, 1 is partly backed, 1 is yours to confirm and 1 is contradicted by it.
What the agent claimed
- ✗41:30
“All tests pass.”
Contradicted. The last test run,
npm testat 41:12, failed: 41 passed, 1 failed. - ◐22:05
“I cleaned up the date formatting.”
Partly backed. 3 files were changed here (dates.ts, format.ts and Export.tsx). A quick look will match them to this claim.
- ?35:47
“I updated the webhook settings in your Stripe dashboard.”
Yours to confirm. The work went through the browser tool.
- ✓18:20
“I added the export button to
billing/Export.tsx.”Backed by evidence. Export.tsx was changed in the session.
- ✓45:02
“The preview is live at https://billing-preview.example.dev.”
Backed by evidence. The link came back from
npx wrangler deployat 44:51.
The problem
Agents report on their own work. Credence checks it.
A long session ends with a confident summary. Some of it happened exactly as described, some of it is hopeful, and some of it is invented. Reading the whole transcript to sort it out takes longer than the work did.
- Did the tests really pass, or did the agent just say so?
- Was that link actually deployed, or made up?
- Which of these twelve claims should I double-check?
What you get
A plain answer for every claim.
Every claim, in the agent’s words
Credence pulls the concrete claims out of the agent’s replies to you, so you see exactly what you were told: changes, test runs, builds, deploys and commits.
A verdict and a reason
Each claim is checked against what the coding tool recorded. You get one verdict and one sentence explaining it.
Problems first
Contradicted claims come first, so you can review a two-hour session in a minute and know which parts need a second look.
Four verdicts
Clear about where every verdict comes from.
Every verdict comes straight from the session record. When a claim reaches beyond it, Credence hands that one to you.
Backed
The record shows it happened: the test run passed, the file changed, the link came back from the deploy.
Partly backed
Related work is in the record. A quick look ties it to this exact claim.
Yours to confirm
The work happened in a browser or another outside tool, so Credence hands it to you to confirm.
Contradicted
The record shows something different: the tests failed, the file stayed the same, or the link came from the agent alone.
How it works
From session to verdict, in four steps.
Capture
Credence follows your sessions through the coding tool’s own hooks and transcripts. Secrets are removed before anything is stored.
Find the claims
It reads the agent’s replies and picks out each concrete claim about work that was done.
Check
Each claim is compared with the recorded commands, file changes, test results and links. In a repo, tests can be re-run in a clean copy.
Review
Confirm or correct each claim. Every review goes into a signed ledger you can export and verify.
Why we built it
Putting a value on AI’s work starts with knowing what it did.
Token Management Group helps companies tie AI spend to the work it produces. Credence is the first step: an honest record of what an agent actually delivered, backed by evidence.
It’s free and open source under Apache 2.0, and it starts in minutes. Anyone can run it, read the code or build on it.
- Local-first. Everything runs and stays on your machine. You choose what to share and when.
- Deterministic at the core. The core checks run on evidence alone. Add an optional AI judge with your own key or a local model.
- Resists gaming. Test commands come from your repo, tests written in the same session must survive mutation checks, and contradicted claims stay contradicted until you rewrite them to match the evidence.
- Extensible. Write verifier plugins, import from other tools, or post reviewed claims on pull requests with the GitHub Action.
See what your agent really did.
Install Credence and open the demo report in a couple of minutes.