claimcheck

Receipts for AI agent runs. After every turn, a receipt: what the agent touched (commands, files, machines, tokens), what it said it did, and whether the two match. Every claim in the agent's final message is marked verified · unverified · pre-existing · contradicted · unchecked — by code, against the tool log, never by another model's opinion. Receipts are signed on your machine and stay on your machine.

Works with Claude Code, Codex CLI, Gemini CLI, Cursor and Hermes Agent. One hook command speaks all of their payloads; one init wires whichever of them you have.

Set it up (one line)

curl -fsSL https://claimcheck.cc/install.sh | sh

Or, if you would rather have your agent do it: install the plugin and say "set up claimcheck".

claimcheck init finds every supported agent in your home folder, backs up its settings file once, and adds two hook lines: after each tool call, and at the end of each turn. claimcheck init --remove takes them out again and leaves everything else as it was.

Then

claimcheck open        # the latest receipt, as a page
claimcheck flagged     # receipts whose headline is not "verified", newest first, with the evidence
claimcheck doctor      # what is wired, how many receipts, any hook errors
claimcheck verify <receipt.json>   # prove a receipt file was never edited (schema · content id · signature)

Receipts live in ~/.claimcheck/receipts/<session>/<turn>.json (+ .html). The live tool log per session is ~/.claimcheck/runs/<session>.jsonl, hash-chained; claimcheck chain <session> recomputes it.

What a receipt says

receipt rcpt_41569cb5b5eca5688eb6dac7 · claude-code · session 5bbefce1…
asked: run the smoke test and report
verdicts: 3 verified · 0 unverified · 1 pre-existing · 0 contradicted · 2 unchecked → pre-existing
- [pre-existing] Added window_width_override=1280 to settings
    evidence: `window_width_override=1280` appears only in content the agent read; never in anything it wrote or ran

• verified — the log contains what the sentence says (the command ran and exited 0; the file was written; the value appears in output).
• unverified — nothing in the log shows it.
• pre-existing — the agent took credit for something it only read.
• contradicted — the log shows the opposite (the command it names failed).
• unchecked — nothing literal to check; counts neither way.

Not a lie detector: verification demand tracks stakes, not error rate. The receipt is the unit of accountability for agent work, the same way a till receipt is for a cashier.

How it stays honest

• The check runs where the log lives. The verifier is a small pure-Python library that runs inside the hook, in milliseconds, offline. No transcript is ever uploaded anywhere.
• Captured live, not read back later. Hooks hand over each tool result before the platform truncates or compacts it; each event carries the hash of the previous one.
• Redacted before hashing. Passwords, tokens, keys and PEM blocks are stripped from every event and from the report before anything is hashed or written, so a receipt can be re-verified without the secret ever existing again.
• Deterministic. Same report + same log = same receipt. The report is split into units by code; a model is never asked for a verdict. 12 real sessions are frozen as golden tests with exactly two real catches and zero false flags. The extractor and verdict rules ship as the claimcheck-core package (binary wheels); without it, receipts are ledger-only and say so in verifier.name.
• Signed. ed25519, one key per machine (~/.claimcheck/key), over the canonical JSON; the receipt's id is its content hash. cryptography or PyNaCl when present, a vendored pure-Python signer otherwise, so a hook works with nothing but python3.
• Open format. spec/receipt-v0.1.schema.json + spec/README.md. Anyone can write a verifier or a viewer.

For the Hermes owner: review cards

hermes_plugin/review_bridge.py is a no-agent cron: every flagged receipt becomes one kanban card for the agent's own profile (skill claimcheck-review) and a line in the Office room. The agent reviews its own receipt against the raw log and answers verifier-right or false-flag with the proof — false flags become verifier rules; the rest becomes the precision record.

Repo map

claimcheck/          the library + CLI (hook, init, receipt, page, verify, sign, capture, store, engine → claimcheck-core)
hermes_plugin/       Hermes Agent plugin (native hooks) + review bridge
hooks/, bin/, .claude-plugin/   Claude Code plugin (hooks.json → bin/claimcheck-hook shim)
plugin.json, skills/            Agent Plugins 1.0 manifest + Agent Skills (claimcheck-setup, claimcheck-review)
spec/                receipt spec v0.1
tests/               document/signature, plugin, hook (6 agents' payloads), installer, bridge, ledger-only fallback

License: Apache-2.0.