Discover
Find capabilities by outcome, not by vendor, UI or implementation language.
OPEN DRAFT · v0.1
The capability layer
for an agent-native web.
ACX (Agent Capability Exchange) is an open draft for machine-readable contracts that let AI agents discover, understand, and invoke software and web-service capabilities. Contracts describe inputs, outputs, permissions, costs, side effects, evidence, and recovery.
Read the specification →See a manifest View on GitHub →{
"acx": "0.1",
"provider": "urn:acx:example:printworks",
"capabilities": [{
"id": "print.document",
"risk": "consequential",
"bindings": ["mcp", "https"],
"economics": { "model": "quote" },
"evidence": ["signed-receipt"]
}]
}
THE MISSING LAYER
Agents can talk. Tools can execute.
The world still needs a common contract for action.
Understand
Read deterministic contracts for permissions, price, consequences and guarantees before execution.
Act & prove
Execute through existing protocols, then receive machine-verifiable evidence and recovery information.
COMPOSE, DON'T REPLACE
A semantic layer over the standards already winning.
ACX does not replace MCP, A2A, HTTP, OAuth, AP2 or Verifiable Credentials. It gives agents one portable description of a capability and points to the appropriate protocols for execution, authorization, identity and settlement.
ACX capability contract
MCP · A2A · HTTPS · gRPC execution
OAuth · OIDC · VC · AP2 trust & settlement
Services · software · machines · robots the world
CORE MODEL
One contract. Eight questions.
WHAT
What outcome can you produce?
HOW
How can an agent invoke it?
WHO
Who may invoke it?
HOW MUCH
What will it cost?
EFFECTS
What changes in the world?
RISK
What approval is required?
PROOF
How is success verified?
UNDO
How can it be recovered?
THE CAPABILITY MANIFEST
Readable by agents.
Auditable by humans.
Providers publish a signed manifest at a well-known URL. Stable capability IDs and JSON Schema-compatible contracts describe inputs and outputs. Consequential operations declare effects before commitment.
Explore the schema →"capability": {
"id": "lab.run_assay",
"inputSchema": { ... },
"effects": ["consumes-material"],
"authority": {"scopes":["lab:execute"],"approval":"human"},
"economics": {"max":"125.00","currency":"USD"},
"evidence": {"receipt":"signed"},
"recovery": {"reversible":false}
}
DESIGN PRINCIPLES
Model-agnostic
No dependency on a specific LLM, framework or cloud.
Protocol-composable
Reuse established transports and trust standards.
Least authority
Agents receive the minimum permission necessary.
Consequences first
Cost, side effects and reversibility are declared before commitment.
Evidence by default
Consequential actions produce receipts software can verify later.
Open governance
Royalty-free specification, public changes and independent conformance tests.
FROM PROJECT TO STANDARD
Built to outlive its founders.
Phase 0 — Draft. Manifest, threat model, schemas, reference server and validator.
Phase 1 — Interop. Independent implementations and public plugfests.
Phase 2 — Neutral home. Trademark/specification stewardship under neutral governance and royalty-free patent commitments.
Phase 3 — Standardization. Submit mature components to appropriate standards organizations while experimental extensions remain fast-moving.
BUILD THE AGENT-NATIVE WEB
Don't build another walled garden.
Publish a capability.
Start with one service, one manifest and one interoperable action.
Implement ACX v0.1 →Read the launch plan