Open Agent Protocol

The vendor neutral standard for how autonomous agents discover, invoke, pay for, govern, and audit the tools and services they use.

Specification · RFCs · Papers · Reference · Conformance · Website · Discussions

What OAP is

OAP is a single conformance surface for the agent economy. It unifies the operational, commercial, legal, and ethical interfaces that today are implemented in fragmented and incompatible ways by the Model Context Protocol, the Agent2Agent Protocol, proprietary function calling, and bespoke enterprise integrations.

OAP covers identity, capability description, structured invocation, monetization, multi agent coordination, confidentiality enforcement, and tamper evident auditing. It is designed to be the default interoperability layer for autonomous agents, comparable in scope to what HTTP, OAuth, and TLS represent for the human web.

Why OAP exists

OAP is additive. It does not replace existing protocols. It composes with them through reference adapters that ship in this repository.

Architecture

                            +-------------------------------+
                            |        AGENT PRINCIPALS       |
                            |  (humans, organizations, AI)  |
                            +---------------+---------------+
                                            |
                                  signed Intents (AQL)
                                            |
       +------------------------------------v------------------------------------+
       |                           AGENT QUERY LANGUAGE                          |
       |  closed operator set, projection, budget, quality floor, signature      |
       +------------------------------------+------------------------------------+
                                            |
                          +-----------------+-----------------+
                          |                                   |
                     RESOLVER                            MATCH BROKER
                   (single provider)                  (verifiable index)
                          |                                   |
                          +-----------------+-----------------+
                                            |
                            +---------------v----------------+
                            |          MANIFESTS             |
                            |  capability, pricing, policy,  |
                            |  identity, conformance level   |
                            +---------------+----------------+
                                            |
                          +-----------------+-----------------+
                          |                 |                 |
                  REQUEST ENVELOPE   POLICY STACK      AUDIT LAYER
                  signed invocation   4 layers         tamper evident
                                                       receipts and
                                                       decision records
                                            |
                            +---------------v----------------+
                            |     COMMERCE AND SETTLEMENT    |
                            |  offers, wallets, statements,  |
                            |  refunds, dispute resolution   |
                            +--------------------------------+

The Policy Stack has four layers, evaluated in order: Universal Prohibitions, Organizational Policy, Persona Policy, Personal Policy. Every Action carries a Decision Record explaining which layer admitted or denied it. Every billable Action produces a Receipt that is hash chained into a tamper evident Audit Log.

Quick start

Validate a Manifest in 60 seconds

git clone https://github.com/openagentprotocol-OAP/oap-spec
cd oap-spec/reference/validator
npm install
node ajv-validate.js ../../examples/team-crm-collaborative/manifest.json

Resolve an Intent against a candidate set

cd oap-spec/reference/aql
npm install
node bin/oap-aql.js resolve \
  ../../examples/aql/discovery-intent.json \
  ../../examples/aql/candidates.json

Returns a signed Intent Response with ranked candidates and a per candidate Decision Record. The same evaluator runs in the browser at openagentprotocol.eu/aql.

Run the full conformance suite

cd oap-spec/test-suite
npm install
node runner.js

Runs schema fixtures, the AQL behavior probe, and the charter assertions. Exits non zero on any failure.

Map an existing MCP server to an OAP Manifest

cd oap-spec/adapters/mcp
node oap-mcp-adapter.js --mcp-tools ./your-mcp-tools.json --out manifest.json

The adapter emits a Level 0 Compatible Manifest. Add pricing, policy, and audit blocks to ascend conformance levels.

Repository layout

oap-spec/
  spec/v1.0/                 Normative specification: OAP-CORE-1.0.md
  schemas/v1.0/              23 normative JSON Schemas (2020-12)
  rfcs/                      26 RFCs covering sessions, commerce, policy, AQL, indexes, NC profile, registry
  papers/                    10 whitepapers on theory, economics, governance
  reference/
    agent/                   Reference Agent runtime
    server/                  Reference Resolver server
    aql/                     Agent Query Language reference parser, evaluator, CLI
    validator/               AJV based schema validator
    mappings/                Cross protocol mappings
  adapters/
    mcp/                     Model Context Protocol adapter
    a2a/                     Agent2Agent adapter
    openai-functions/        OpenAI Function Calling adapter
    langgraph/               LangGraph adapter
  examples/                  Worked Manifests, Receipts, AQL Intents, scenarios
  test-suite/                Conformance runner, schema fixtures, behavior probes
  governance/                Charter, RFC process, Working Groups
  .github/                   Issue templates, PR template, release notes, CI

Specification, RFCs, and Papers

The 12 whitepapers cover accountability, the agent web, agent context mobility, confidentiality, databases for the agent economy, economics, governance, interoperability, the policy stack, the presentation layer (storefront to canvas), verifiable conformance, and the storefront to manifest transition.

Conformance levels

Conformance is machine verifiable. The reference test suite runs schema, behavior, and charter assertions against any implementation.

Reference implementations

All reference code is published under the Apache License 2.0. Implementers may vendor or fork without restriction.

Governance

The Open Agent Protocol is community driven. There is no foundation, association, corporation, or other legal entity that owns, controls, or speaks for OAP. There are no membership dues and no licensing fees. The specification is maintained by an open community of contributors through the public RFC process and through CI enforced quality gates in the open source repositories under the openagentprotocol-OAP GitHub organization.

Protocol changes follow an open RFC process with publicly archived discussion. Decisions are taken by rough consensus on RFCs and confirmed by a Peer Review Quorum of at least three Maintainers from at least three distinct organizations, as defined in governance/WORKING-GROUPS.md. Working Group Coordinators have no veto power; their role is to triage, facilitate consensus, and shepherd RFCs through the process.

Services that the protocol depends on, including the OAP Registry (RFC 0026), the Trust Score reference implementation, the Sampling and Outcome Verification probes, and the Revocation Service, are open source software that any community member may deploy. Multiple competing instances are presumed and encouraged.

Working Groups are documented in governance/WORKING-GROUPS.md. The RFC process is documented in governance/RFC-PROCESS.md.

Community

Substantive protocol changes follow the RFC process. Editorial corrections may be proposed directly through pull requests against the latest published draft.

Contributing

OAP is an open standard. Contributions from individuals, companies, civil society, and government participants are welcome. Begin with CONTRIBUTING.md and the Code of Conduct.

Licensing

The specification text is published under the Creative Commons Attribution 4.0 International License. See LICENSE-SPEC.

Reference implementations, schemas, adapters, and tooling in this repository are published under the Apache License 2.0. See LICENSE-CODE.

Citation

A machine readable citation file is provided at CITATION.cff. The recommended human readable form is:

Open Agent Protocol Working Groups. Open Agent Protocol, Version 1.0 Public Working Draft. OAP-CORE-1.0. 2026. https://openagentprotocol.eu

Status

This is a Public Working Draft. Implementations are encouraged. Non breaking refinements are expected before the Final designation is granted.