Skip to content

AgentIsOK

AgentIsOK — gold AI above OK with a green checkmark

Open clearance for authorized agents. Evidence travels. The origin decides.

Conformance License: Apache-2.0

AgentIsOK proposes a machine-verifiable alternative to a human CAPTCHA when a website is willing to admit authorized automation. The website asks for bounded evidence about the agent's authority and exact request, then applies its own policy. Ordinary account, abuse, and application controls still apply.

The name compresses to AI / OK. “OK” means OK for this action, on this origin, under this policy, at this time. The logo identifies the project; it is not a safety certification or global trust score.

Status: Pre-alpha research prototype. The candidate protocol, two implemented evidence profiles, reference verifier, and shadow adapter are available. Production enforcement, independent interoperability, and origin outcome evidence are not. This code has not received an independent security review and must not authorize production actions. See implementation limits and evidence gates.

Start here

Why this exists

Human challenges are an awkward interface for agents acting for people and organizations. When existing API access or signed-agent rules are insufficient, those agents need a legitimate path through the automation boundary without impersonating humans or inheriting unrestricted credentials.

AgentIsOK is a challenge and clearance layer:

protected request
  -> 401 + origin challenge
  -> presentation POST (HTTP Message Signature + mandate)
  -> local verification and policy
  -> allow / allow with obligations / step up / deny / indeterminate
  -> optional sender-constrained retry

The central boundary is simple:

AgentIsOK verifies and composes evidence. The receiving service decides what is OK.

Web Bot Auth addresses automated-client authentication; its IETF charter excludes end-user authentication and an intent vocabulary. AgentIsOK explores a complementary challenge, mandate, and obligation contract for origin policy. Its Web Bot Auth adapter remains a sketch, and this repo does not claim interoperability with a deployed Web Bot Auth service.

Why it is open

Agent access to the web is too consequential to depend on one company's permission, directory, hosted verifier, or private policy. A conforming implementation must be independently implementable. Compatibility cannot require an AgentIsOK account, a commercial API key, a proprietary allowlist, or an AgentIsOK-operated control plane.

The working technical name is Agent Clearance Protocol. Do not abbreviate it ACP; that short name is already used by unrelated agent protocols. Identifiers use the token agent-clearance.

Design principles

  • Origin authority: Evidence informs; the receiving service decides.
  • Request-bound trust: “OK” applies to one action and context, not an actor forever.
  • Separation of powers: Identity, delegation, request integrity, risk, rate limits, and payment authority remain distinct.
  • No mandatory gatekeeper: Local verification, self-hosting, multiple issuers, and competing implementations are first-class requirements.
  • Minimum disclosure: Participants learn only what the action requires.
  • Protocol reuse: Existing Internet and identity standards are composed rather than replaced.
  • Explicit failure: Denial, insufficient evidence, and operational failure are not conflated.
  • Enforceable obligations: Limits returned by policy must be applied by the enforcement point. Unknown critical obligations fail closed.

See PRINCIPLES.md for the complete set.

What exists now

The first discriminating loop is in this repository:

  1. A test origin challenges a bounded availability.read.
  2. An agent responder signs the presentation HTTP request (HTTP Message Signatures profile).
  3. It presents an origin-scoped mandate (mandate profile).
  4. Local policy returns allow_with_obligations for the read and requires step-up before reservation.commit.
  5. Twenty-six conformance vectors cover replay, validity windows, malformed mandates, exact request binding, omitted or excessive scope, profile confusion, missing evidence, unenforceable obligations, untrusted issuers, and indeterminate revocation.
Surface Implemented today Still required
Python reference In-process presentation verification, bounded decisions, test artifact issuance HTTP serving, artifact redemption, actual rate/result enforcement, production keys and replay storage
Evidence profiles RFC 9421 subset and origin-scoped test mandates Approval/issuance integration, live revocation policy, independent security review
TypeScript edge adapter Sampled comparison, bounded waits, minimized metrics Origin evaluator with access to the actual signed bytes and challenge state
Conformance Deterministic fixtures and negative cases A materially independent verifier and responder
Pilot Qualification, rollout, and measurement contracts An origin-owned experiment with independent outcome labels

Run the reference loop

Requires Python 3.11+ and a repository checkout. The fixtures use a fixed evaluation time and public test keys; they do not contact an origin or issuer.

git clone https://github.com/Univeracity/agentisok.git
cd agentisok
python3 -m venv .venv
. .venv/bin/activate
pip install -e "./reference[test]"
pytest -q reference/tests
agent-clearance-conformance

Expected conformance summary: 26 passed, 0 failed. The TypeScript adapter requires Node.js 24+; run npm ci && npm test from integrations/typescript-edge.

Passing fixtures establishes prototype behavior. The product hypothesis is whether an origin can recover desirable agent traffic within its abuse, privacy, and operating-cost budgets. Shadow mode can estimate decision disagreements; it cannot demonstrate recovered completion while the incumbent still blocks the request. That origin evidence does not exist yet.

What to do next

The next work is evidence-driven:

  • origin-side discovery: real workflows where legitimate agents are blocked or challenged;
  • one shadow-mode design partner, with existing controls still deciding and the origin able to reconstruct the proof;
  • a second independent verifier before anyone calls this mature.

See CONTRIBUTING.md.

Repository map

Document Purpose
PRINCIPLES.md Non-negotiable design and ecosystem invariants
SCOPE.md What AgentIsOK does and deliberately does not do
ARCHITECTURE.md Roles, trust boundaries, data flow, and deployment modes
THREAT-MODEL.md Technical, privacy, governance, and ecosystem threats
WHITEPAPER.md Motivation, thesis, proposed design, and adoption path
ROADMAP.md Evidence required to advance, reshape, or stop
spec/agent-clearance-protocol.md Protocol skeleton and candidate HTTP binding
spec/profiles/ HTTP Message Signatures, origin-scoped mandate, Web Bot Auth adapter
reference/ Prototype origin, responder, and conformance runner
conformance/ Positive and negative vectors
pilot/ Design-partner, shadow-mode, outreach, and metrics package
integrations/typescript-edge/ Deployable Fetch-standard shadow adapter
CONTRIBUTING.md How to participate
IPR-POLICY.md Contribution, patent, and standards commitments
GOVERNANCE.md Open decision-making and anti-capture safeguards
SECURITY.md Vulnerability reporting and security status
CODE_OF_CONDUCT.md Conduct expectations
MAINTAINERS.md Current stewardship
assets/brand/ Logo SVGs, PNGs, social preview, and usage guidance

Standards posture

The protocol profiles HTTP Message Signatures and composes adjacent work such as Digest Fields, OAuth proof of possession and rich authorization, Privacy Pass architecture, and Web Bot Auth as an optional request-integrity adapter. Unstable drafts are not treated as settled foundations.

The project will seek interoperability experience before pursuing formal standardization in an appropriate existing venue.

License

Copyright 2026 the AgentIsOK contributors.

Licensed under the Apache License 2.0. The license applies to code, specifications, documentation, examples, and test material in this repository unless a file states otherwise.

About

Open, origin-authoritative clearance protocol for legitimate AI agents

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages