Get started

Your first report in a few minutes.

One line installs Credence and sets it up. Answer two questions, and your reports open in the browser.

Install and set up

Paste this into a terminal on macOS or Linux. It downloads the ready-to-run program for your machine and verifies its checksum.

curl -fsSL https://tokenmanagementgroup.com/credence/install.sh | sh

Then it asks two questions. Press Enter for yes:

Connect Credence to Claude Code? This adds capture hooks and a /credence command, and starts a small background service. Your settings get a backup first.
Check your sessions from the last 7 days now? Credence reads your recent sessions and checks every claim in them.

Your reports open in the browser. Everything stays on your machine. You can also read the script or download a build from the releases page.

Keep working

Use Claude Code as usual. When a reply makes claims, a one-line Credence note appears under it:

Credence: 2 claims in this reply: 1 backed by evidence, 1 contradicted. ✗ “All 6 tests pass.” The last test run, npm test at 7:41, failed: 5 passed, 1 failed. Type /credence for every claim.

Type /credence in Claude Code to see every claim in the session, each with its verdict and reason. You can also simply ask Claude “is that really done?” and it runs the check itself. Codex and Cursor sessions get full reports in the browser.

Open your reports anytime

Type this in any terminal. The first time, it runs the setup; after that, it opens your reports.

credence

Or read the latest report right in the terminal:

credence show

See an example

The demo tells a short story. Sam runs a bakery and asks Claude Code to update the bakery’s website: holiday hours, a phone-friendly layout, a cake order form and the hours on Google. Credence checks the seven things the agent says it did and opens the report, with a verdict and a one-line reason for each. Everything stays inside the demo’s own folder.

credence demo

Other ways to set up

Your setup, your way.

Preview every change first

This lists each change to your coding tools’ settings. Run credence init afterwards to apply them.

credence init --dry-run

Install now, set up later

This installs the program only. Type credence whenever you’re ready to answer the two questions.

curl -fsSL https://tokenmanagementgroup.com/credence/install.sh | CREDENCE_NO_SETUP=1 sh

Build from source

With Go 1.27 or later. The second line adds Go’s program folder to your PATH, and the third runs the setup.

go install github.com/BTONGK/credence/cmd/credence@latest
export PATH="$HOME/go/bin:$PATH"
credence

Check older sessions

Reach further back than the first week whenever you like.

credence backfill --since 30d

Keep notes to the browser

This turns off the note under Claude Code replies; reports and /credence stay as they are.

credence config set notify.replies off

Remove Credence

This removes the hooks, the /credence command and the background service. Add --purge to delete your stored sessions too.

credence uninstall

Optional

Add an AI judge.

The core checks run on evidence alone. For claims that call for judgment, you can add a judge that reviews the evidence blind to the agent’s own confidence. It uses your own key or a local model, and stays within a monthly budget you set.

With Claude, using the key in ANTHROPIC_API_KEY:

credence config set judge.provider anthropic
credence config set judge.model claude-opus-5-5
credence config set budget.monthly_usd 20

With a local model, through any OpenAI-compatible server:

credence config set auditor.provider openai-compatible
credence config set auditor.base_url http://127.0.0.1:11434/v1

Commands

The ones you’ll use most.

Everything in the browser view has a terminal equivalent, and most commands take --json. Run credence help for the full list.

Command	What it does
`credence`	Set up the first time; after that, open your reports in the browser.
`/credence`	In Claude Code: every claim in the current session, with its verdict and reason.
`credence check`	The same check in a terminal, for the newest session in the current folder.
`credence show [session]`	Print the latest report in the terminal, or the session you name.
`credence review [session]`	Go through claims with the keyboard and confirm or correct each one.
`credence verify <claim> --rerun`	Check a claim again, re-running its tests.
`credence ledger`	See how often each agent setup’s claims hold up, and verify the signed history.
`credence export --format json`	Export your reviews to share or archive.
`credence import <format> <path>`	Bring in sessions from transcripts, OpenTelemetry, Agent Trace, git-ai or Entire.
`credence doctor`	Check that capture, the service and your settings are working.
`credence setup`	Run the two-question setup again, for example after installing another coding tool.

Contribute

Credence is built in the open.

Bug reports, support for more coding tools and verifier plugins are all welcome. The claim format is an open spec, so other tools can produce and read the same reports.