ActionProof MCP server

ActionProof

A tamper-proof audit trail for AI agents. Verifiable observability: every action your agent takes gets a cryptographically signed receipt you can verify offline, anywhere — zero backend.

Observability tools (LangSmith, Langfuse, Arize) show you what your agent reportedly did — traces recorded inside their platform, on their word. But those logs are self-asserted: an agent, a bug, or an attacker can write anything into them, and you can't prove after the fact that the record wasn't edited.

ActionProof adds the missing layer: verifiable observability. Each action — email sent, form filed, payment made — gets a tamper-evident, Ed25519-signed receipt capturing what was done, by which agent, when, and on whose authority. Edit any field and verification fails. It's an audit trail you (or an auditor, a user, or a counterparty) can trust without trusting the agent, the vendor, or us.

Built for the compliance floor that's coming — the EU AI Act (Article 12) and ISO 42001 require traceable, tamper-evident logs for automated decisions. ActionProof produces exactly that, as a portable primitive rather than a walled-garden platform.

Install

npm install actionproof      # TypeScript / JavaScript
pip install actionproof      # Python

Receipts are cross-compatible: one signed in TypeScript verifies in Python, and vice-versa.

Quick start (TypeScript)

import { attest, verify, generateKeypair } from "actionproof";

const agent = generateKeypair();               // agent's identity = its key (did:key)

const receipt = attest(agent, {
  type: "email.send",
  summary: "Sent renewal quote to jane@acme.com",
  params: { to: "jane@acme.com", amount: 4200 }, // hashed, not stored in clear
  result: { smtp: 250 },
  outcome: "ok",
});

verify(receipt);            // -> { valid: true, agent: "did:key:z6Mk..." }

Quick start (Python)

from actionproof import attest, verify, generate_keypair

agent = generate_keypair()

receipt = attest(
    agent,
    type="email.send",
    summary="Sent renewal quote to jane@acme.com",
    params={"to": "jane@acme.com", "amount": 4200},  # hashed, not stored in clear
    result={"smtp": 250},
    outcome="ok",
)

verify(receipt)             # -> VerifyResult(valid=True, agent="did:key:z6Mk...")

Edit any field of that receipt and verify returns invalid. That's the whole idea.

Where it fits: the verifiable layer of agent observability

ActionProof complements your observability stack rather than replacing it. Keep using LangSmith / Langfuse / Arize for rich traces, latency, and cost — then attach an ActionProof receipt to the actions that matter (the ones that move money, change state, or touch a user's data) so that part of your trail is tamper-evident and independently verifiable.

It's a proof, not just a log entry — the difference between "our dashboard says the agent did this" and "here's a signed receipt anyone can verify."

Design principles

Shortened. The full README is on GitHub.