Open source · MIT

Prove what your AI agents did.

A receipt for every move your AI makes, which no one can quietly edit later. Not even you.

Every tool call, decision and payment your agent makes gets signed with Ed25519 and chained to the entry before it. Anyone can check the record later, with only your public key, without trusting the agent's own summary or trusting you.

pip install git+https://github.com/cipherxsniper/agentproof.git

Add it to your agent Try to tamper with a log

Ed25519 signaturesSHA-256 hash chainverifies offlinecrash-safe appends

A real signed log: a refund agent

public key

1. An AI refund agent did 6 things. Each one below is signed and linked to the one before it.
2. Your browser checks every signature right now. Nothing is sent anywhere.
3. Press Change the refund to $640, the kind of edit someone might try after the fact, and watch the proof break.

Checking…

How it works

Three ideas, no blockchain required

Sign every action

Each event is serialized deterministically and signed with your Ed25519 key at the moment your code records it.

Chain the entries

Each entry stores the SHA-256 hash of the one before it. Editing, deleting or inserting anything breaks every link after it.

Verify from outside

Anyone with your public key can re-check every signature and link, offline, in Python or right here in the browser.

Hook it into your stack

Five minutes to a verifiable agent

Verify a log

Check anyone's agent log, right here

Log file (one JSON entry per line) Signer's public key (base64)

Paste a log and a public key.

Nothing leaves your browser

The check runs locally with the Web Crypto API: the same rules as agentproof verify. Each entry's link to the previous hash, its own hash, and its Ed25519 signature are all re-computed.

Teams use this to hand an auditor, customer or lender a log they can check without database access and without taking anyone's word for it.

For business data, too

proofchain turns any Postgres table (payouts, invoices, trades) into the same kind of signed chain: pip install proofchain-cli.

Running in production

Built for, and used by, a real autonomous system

5 agents, one signed channel

A two-server trading and research operation where five AI agents coordinate over a signed, hash-chained message bus: 3,900+ messages, each verifiable.

Bitcoin-anchored track record

Daily signal lists are published and timestamped in Bitcoin via OpenTimestamps before their outcome. See the live record ↗

Policy before action

Spending rules are checked and the decision is signed before the agent acts, so a denied payment is on the record too.

Straight answers

What a signature can and can't prove

It proves that the holder of a specific key recorded a specific sequence of events, and that nothing in that sequence was edited, removed or inserted since.
It does not prove that what was recorded is true. If an agent misreports a tool's result and signs it, the signature is still valid. The guarantee is as strong as what you log and where you log it.
Arguments and responses of wrapped calls are stored as hashes, not raw values, so a log can be shared for verification without leaking customer data.
Keep the signing key away from the agent you're auditing when you can (a sidecar process or separate service), so the agent can't rewrite its own history.

Work with me

Want verifiable logs in your agents this week?

I'm Thomas Lee Harvey, the developer behind agentproof. I'll wire signed audit trails, policy checks and a public verify page into your agent stack, and show your team or your auditor how to check them. Email me what your agents do; I reply within a day with a plan and a fixed price.

tommyleeharvey17@gmail.com Star on GitHub ↗