Your agent is guessing what it's allowed to do. Eventually it guesses wrong.

OpenTerms is an open standard for machine-readable permission rules. Your agent checks a domain before it acts and gets back: allowed, denied, or escalate.

pip install openterms-py

import openterms
result = openterms.check("github.com", "scrape_data")
# → allowed / denied / not_specified

Get started Read the spec →

511 services indexed. Open standard, free SDK, no lock-in.

Robots.txt told crawlers where they could go. OpenTerms tells agents what they can do.

FAILURE 01

Acts where it shouldn't — and gets blocked

The agent attempts an action the service forbids, gets rate-limited or banned, and the workflow fails. The engineer debugs a problem a permission check would have prevented.
FAILURE 02

Acts where it shouldn't — and it goes through

The action succeeds when it should not have. Now there is data to roll back, a dispute to manage, or a relationship to repair. Cleanup costs more than prevention.
FAILURE 03

Doesn't act where it should — and stalls

An over-cautious agent defers an action it was permitted to take. The workflow halts, a human is paged, and throughput drops for no reason.

How it works

A service publishes its permissions

Any service drops an openterms.json file at /.well-known/ declaring what agents may do — or what they may not.

An agent checks it — one call — before acting

Before any action, the agent calls openterms.check(domain, action) and gets back allow / deny / escalate. No model in the loop.

Deterministic: The check is a lookup, not a model judgment — the same question always returns the same answer, no AI guessing in the loop.

/.well-known/openterms.json

{ "service": "github.com", "openterms_version": "0.3.0", "permissions": { "read_content": true, "scrape_data": false, "api_access": { "allowed": true, "requires_auth": true } } }

Check is the first of three. As your agents do more, the questions get bigger.

What were agents permitted to do?

A lookup before every action. Deterministic, fast, no model required.

Engineering Live · Public alpha

What did they actually do?

Signed receipts of every action — independently verifiable, reviewable history.

Security & risk Private beta

By what authority did they act?

Delegated-authority commitments, bound to the receipts that record what they did.

Legal & counsel Private alpha / not generally available

Each product stands alone. Together they form a fuller evidence chain. The standards beneath them are open.

Built on published cryptographic standards — not proprietary formats. Anyone can verify, and nothing locks you in.

RFC 8785 RFC 7517 Ed25519 JCS Vendor-neutral

View on GitHub →