Attest

Attest — prove what your AI agent actually did

[ci] [pypi] [npm] [docs] [license]

Decide → Gate → Verify → Attest. Your agent sends the email, updates the CRM, moves the money. Attest decides whether it may, gates it behind a human when it matters, reads back from the system of record to check it actually happened, and records tamper-evident evidence either way.

Attest never executes your action. Your tool does. That is why it works with any app and any framework on day one.

Your agent says it worked. Did it?

Two identical CRM writes, both reported success by the tool, both approved by a human:

$ attest ledger#2  2026-09-19 10:16:46  someweirdcrm.create → leads   ask  approved  unverified   a09e7641db12#1  2026-09-19 10:16:46  someweirdcrm.create → leads   ask  approved  verified     77a6d99a60f6

The second one read back clean. The first did not, and the ledger says which field disagreed:

{ "level": "unverified", "matched": false,  "evidence": { "compared": 4, "exists": true, "failed": ["stage"],    "fields": { "stage": { "want": "qualified", "got": "new", "ok": false } } } }

A trace would have shown two green steps. That gap is the entire product.

Install

pip install attestlayer          # Pythonnpm  install attestlayer         # TypeScript — byte-identical ledger format

import attest
@attest.action(system="gmail", verb="send", target="to")def send_email(to, subject, body): ...

import { Attest } from "attestlayer";
const at = new Attest({ agent: "followup-agent@v3", readers: { gmail: process.env.GMAIL_TOKEN! } });const sendEmail = at.wrap({ system: "gmail", verb: "send", target: "to" },                          async ({ to, subject, body }) => gmail.send({ to, subject, body }));

That is the whole change. The first external send now pauses for a human, and the ledger row says acknowledged, or verified once Attest can read back with the agent's own credentials (Attest(readers={"gmail": service})). Nothing else in your code moves.

Verification is layered, and it never overstates

Every row states exactly how much was checked. A check that could not run is never dressed up as a pass.

Only a contradiction earns unverified. Missing credentials, a read-only scope or an unsupported verb degrade to acknowledged, because "we could not check" and "it did not happen" are different facts.

Where it plugs in

Human confirmation lands where the team already works: Slack, the web inbox, a webhook, a LangGraph interrupt, or the console. Approvers can edit the parameters before approving, and the edit is recorded.

The record itself

Hash-chained ledger, SQLite locally and Postgres per organisation in the cloud. Signed checkpoints let you prune old rows and still verify the chain. Optional Ed25519 signing, and external anchoring to a file, a git repo or an HTTP endpoint. Exports: JSON, CSV, the IETF draft-sharif-agent-audit-trail JSONL format, and an EU AI Act event-log pack.

Overhead is roughly a sixth of a millisecond per action, measured in benchmarks/ (published numbers).

Try it locally

python examples/unknown_app.py                         # an app Attest has never seen, L1 → L3ATTEST_AUTO_APPROVE=1 python examples/langgraph_agent.pycd deploy && cp .env.example .env && docker compose up  # cloud API :8400 + dashboard :3400

Attest Cloud

The SDK is MIT and works standalone with a local ledger. Attest Cloud adds the shared ledger, the confirm inbox, versioned org policy, agent keys and compliance exports. Your vendor tokens never reach it: read-back happens in your process with your own credentials, and only hashes and previews are sent. Self-host it from deploy/, or read DEPLOY.md.

Install channels

Documentation

Full docs at dev-prathap.github.io/ATTEST — quickstart · verification levels · policy · read-back recipes · ledger & exports · hardening

Repository

pytest -q && (cd cloud && pytest -q)     # 355 + 37 offline testspytest tests/live                        # 8 live suites; need real credentials, skipped without them

Contributing

Read-back recipes are the easiest place to start: each one teaches Attest how to confirm a write in one more app, and needs nothing but that app's read API. See CONTRIBUTING.md, and SECURITY.md for reporting a vulnerability.

Principles

• We never execute your action. Your tool executes; we decide, gate, verify and record.
• Coverage is universal. Verification depth is layered and stated honestly on every row.
• Open-source SDK under MIT. Paid cloud for the shared ledger, inbox, policy and exports.

Design notes

The thinking behind the product, kept in the open: vision · product · universal adapter · architecture · market · build plan · decisions · phase plan

Lineage

Attest's core mechanisms are extracted from two working codebases: DO (policy engine, read-back verification pairs, evidence ledger) and DeerFlow (tool receipts, verification patterns, MCP and chat-channel adapters), which run against real Gmail, Slack, HubSpot, Notion, Linear and Google Workspace. Attest ships its own live suites for those systems in tests/live/; they need real credentials and are skipped without them.