Agent Discovery eXchange (AX) is an open, internet-native discovery protocol that enables AI agents to advertise their capabilities and supported interaction protocols in a standardized, machine-readable format using existing HTTPS infrastructure.
AX is intentionally scoped to discovery only. It does not define execution semantics, trust models, governance workflows, or economic mechanisms. Instead, AX provides a shared discovery substrate that registries, exchanges, marketplaces, and coordination layers can consume prior to registration, onboarding, trust evaluation, or execution decisions.
By decoupling discovery from coordination, AX enables agent ecosystems to scale openly while allowing downstream systems to differentiate on trust, governance, policy, and execution.
AX is being developed as an open, vendor-neutral standard and is developed as an Internet-Draft standard to maximize visiblity and global adoption.
AX is guided by a small set of principles:
- Discovery before execution
- Decentralisation over registries
- Declaration over assumption
- Policy at the edge
- Evolution over finality
These principles are reflected in both the specification and its release process.
Agents publish a discovery document at a well-known HTTPS location under a domain they control:
https://<domain>/.well-known/agent-exchange
The AX document:
- describes the agent’s capabilities,
- lists supported interaction protocols,
- includes optional descriptive metadata,
- and may include capability hashes for change detection.
AX documents are:
- retrieved over HTTPS,
- accessible without authentication,
- suitable for crawling, indexing, and caching.
AX discovery is non-authoritative. Publishing an AX document does not imply trust, endorsement, availability, pricing, or authorization to invoke the agent.
- A discovery and metadata exchange mechanism
- Protocol-agnostic
- Execution-neutral
- Built on existing internet standards (DNS, HTTPS, .well-known)
- Designed for incremental adoption
At its core, AX allows an agent to publish a machine-readable document that answers:
Who am I?
What can I do?
How can you interact with me?
Under what constraints and trust assumptions?
AX explicitly does not:
- Define agent execution semantics
- Define task or message formats
- Replace existing protocols (MCP, A2A, GraphQL, REST, etc.)
- Act as an identity provider or authorization framework
- Require central registries or governance bodies
- Impose a specific architecture or product model
AX is intentionally narrow in scope.
AX occupies a distinct layer in the agent ecosystem:
+-----------------------------+
| Agent Reasoning & Tools |
| (LLMs, planners, tools) |
+-----------------------------+
| Execution Protocols |
| (MCP, A2A, GraphQL, REST) |
+-----------------------------+
| AX — Discovery & |
| Capability Exchange |
+-----------------------------+
| Internet Infrastructure |
| (DNS, HTTPS, PKI) |
+-----------------------------+
AX answers “how do I find and understand an agent?”
Execution protocols answer “how do I talk to it once I do?”
AX is specified using Internet-Drafts and a JSON schema.
-
Draft-00: Core discovery mechanism
Defines HTTPS.well-knowndiscovery semantics and document structure. -
Draft-01: Capability hashing
Adds capability hashing to support efficient change detection without altering discovery semantics. -
Draft-02: Trust & Attestation
Adds optional trust tier declarations, attestation statements for provenance and verification, explicit support for multiple independent attestation authorities, and clear guidance that trust signals are advisory inputs rather than enforced controls. -
AX JSON Schema
Defines the structure of the AX document.
To be AX-compliant, an agent:
- MUST expose a valid AX document at the
.well-knownHTTPS location - MUST serve the document over HTTPS using UTF-8 JSON
- MUST NOT imply trust, authorization, or endorsement through discovery
- SHOULD support standard HTTP caching semantics
- MAY include capability hashes as defined in Draft-01
AX deliberately avoids defining any registration, trust, or execution requirements.
The smallest useful AX document an agent can publish:
{
"record_type": "AX",
"version": "1.0",
"agent": {
"name": "Example Tax Filing Agent",
"description": "Provides automated tax preparation and filing assistance.",
"provider": "ExampleCo"
},
"endpoints": [
{
"protocol": "a2a",
"url": "https://api.example.com/agents/tax-filing/a2a",
"auth": ["OIDC"],
"content_type": "application/json"
}
]
}Agents may support multiple interaction protocols simultaneously.
{
"record_type": "AX",
"version": "1.0",
"agent": {
"name": "Research Assistant Agent",
"description": "Analyzes documents and answers research questions.",
"provider": "ExampleCo"
},
"endpoints": [
{
"protocol": "graphql",
"url": "https://api.example.com/graphql",
"auth": ["OIDC", "JWT"],
"content_type": "application/json"
},
{
"protocol": "mcp",
"url": "https://api.example.com/agents/research/mcp",
"auth": ["mTLS"]
},
{
"protocol": "a2a",
"url": "https://api.example.com/agents/research/a2a",
"auth": ["OIDC"]
},
{
"protocol": "rest",
"url": "https://api.example.com/agents/research",
"auth": ["OAuth2"],
"content_type": "application/json"
}
],
"schema": {
"graphql_schema_url": "https://api.example.com/graphql/schema.graphql",
"mcp_manifest_url": "https://api.example.com/agents/research/mcp/manifest.json",
"rest_openapi_url": "https://api.example.com/agents/research/openapi.json",
"introspection": true
},
"capabilities": {
"intents": ["document.analysis", "research.qna"],
"async": true,
"supports_callbacks": true,
"callback_modes": ["webhook", "poll"]
}
}Draft-01 introduces capability hashing. Until the schema formally defines the hashing fields, implementations SHOULD carry hashes in extensions to remain forward-compatible.
{
"record_type": "AX",
"version": "1.1",
"agent": {
"name": "Claims Processing Agent",
"description": "Processes insurance claims and returns eligibility decisions.",
"provider": "ExampleCo"
},
"endpoints": [
{
"protocol": "a2a",
"url": "https://api.example.com/agents/claims/a2a",
"auth": ["OIDC"]
}
],
"extensions": {
"ax": {
"capability_hash": "sha256:7f3c2e4c8b1f..."
}
},
"capabilities": {
"intents": ["claims.processing", "claims.eligibility"],
"async": true,
"supports_callbacks": true,
"callback_modes": ["webhook"]
}
}Capability hashing:
- does not change discovery semantics,
- does not imply trust,
- enables consumers to detect meaningful changes efficiently.
Draft-02 introduces trust and attestation. Until the schema formally defines trust and attestation fields, implementations SHOULD carry trust in extensions to remain forward-compatible.
{
"record_type": "AX",
"version": "1.0",
"agent": {
"name": "Travel Planner Agent",
"description": "Plans itineraries and provides travel recommendations based on user preferences.",
"provider": "ExampleCo"
},
"endpoints": [
{
"protocol": "rest",
"url": "https://api.example.com/agents/travel-planner",
"auth": ["OAuth2"],
"content_type": "application/json"
}
],
"capabilities": {
"intents": [
"travel.itinerary.plan",
"travel.recommendation.generate"
],
"async": true,
"supports_callbacks": true,
"callback_modes": ["webhook"]
},
"schema": {
"rest_openapi_url": "https://api.example.com/agents/travel-planner/openapi.json"
},
"security": {
"issuer": "https://example.com",
"jwks_url": "https://example.com/.well-known/jwks.json"
},
"extensions": {
"ax": {
"trust": {
"trust_tiers": ["self-asserted"],
"attestations": [
{
"type": "self-assertion",
"scope": "agent-metadata",
"issuer": {
"name": "ExampleCo",
"uri": "https://example.com"
},
"issued_at": "2026-01-20T00:00:00Z",
"notes": "Self-asserted trust metadata only. Consumers MUST apply local policy and verification before use."
}
]
}
}
}
}
An agent operator publishes the AX document at:
https://example.com/.well-known/agent-exchange
The resource:
- MUST be accessible via HTTPS
- MUST return UTF-8 JSON
- SHOULD include appropriate cache headers
No authentication is required for discovery.
Registries, exchanges, and indexers typically:
- Discover candidate domains
- Retrieve the AX document
- Validate it against the AX schema
- Index advertised capabilities and protocols
- Decide whether to onboard, register, or ignore the agent
AX provides the input signal; downstream systems make governance decisions.
AX is applicable to many downstream systems, including:
- Enterprise agent registries
- Public or curated agent exchanges
- Marketplaces layered on top of discovery
- Partner-built vertical registries
- Brokerages and federated coordination systems
- Open search indexes for agents
Any system that needs to decide which agents to trust, onboard, or coordinate can use AX as a discovery substrate.
-
Background Reading
Agent Discovery Exchange (AX): Defining the social contract by which agents introduce themselves to the internetDiscovery Is Infrastructure: Why Agents Need Open Signals Before Platforms
From Web Discovery to Agent Tethering: Scaling a Global Network of Agents
-
GitHub Discussions
Community feedback, design notes, and evolution of the specification.
Discussion: Initial public draft of Agent Discovery Exchange (AX)Discussion: Capability hashing
Design Note: AX as a Pre-Registration Discovery Layer for Registries
-
Closed Issues
#4: Align Discovery Mechanism with RFC 8615: Make .well-known HTTPS the Required Path -
Open Issues
#3: Trust Signaling & Attestation
AX is early but intentional.
- It is not an IETF standard
- It is not final
- It is expected to evolve
- Feedback and alternative viewpoints are welcome
AX is published as Internet-Draft–style documents to encourage discussion, experimentation, and independent implementation.
- Publish an AX document at the
.well-knownpath - Validate it against
ax-schema.json - Consume AX documents using a crawler or indexer
- Use capability hashes to detect changes over time
Example templates and utilities are available in the examples/ directory.
AX is implementation-agnostic.
Any system that:
- Publishes an AX document
- Consumes AX documents
- Uses them to inform agent interaction decisions
can claim AX compatibility.
Reference and example implementations may be published separately.
Feedback, issues, and pull requests are welcome. Please use GitHub Discussions for design topics and Issues for concrete proposals.
This project is licensed under MIT-0 license. Specification contributions follow Internet-Draft conventions.
AX defines a missing layer in the agent ecosystem: open, standardized discovery prior to registration.
By decoupling discovery from coordination, AX enables agent ecosystems to scale globally without fragmenting around proprietary registries.