Agent Commerce Protocol (ACP)

Open specification and reference implementation for agent-native data commerce.

ACP enables AI agents to autonomously discover, evaluate, and transact with data providers — without human mediation at each access decision. It defines the commercial and governance layer that sits between agent task execution and proprietary data sources.

---

The Problem

Traditional SaaS data products were built for human buyers: pricing pages, manual authentication, signed contracts. AI agents can't navigate any of that. When an agent tasked with M&A research hits a paywall, it has no way to evaluate the terms, decide if they comply with its operating policy, or execute a transaction — so it abandons the access attempt.

ACP fills that gap.

---

Three Schemas, One Flow

Provider publishes          Agent carries             Agent produces
ATS document       ──────▶  AAP policy      ──────▶   ATR record
(what's available,          (what it's                (what was agreed,
 what it costs,              authorized to             audit trail)
 terms of access)            agree to)

Schema	Role	Published by
ATS — Agent Terms Schema	Machine-readable product terms at /.well-known/agent-terms.json	Data providers
AAP — Agent Authorization Policy	Operating authority the agent carries at runtime	Agent creators
ATR — Agent Transaction Record	Tamper-evident audit record of every access decision	Agents (to creator's log endpoint)

---

Real-World Workflow Examples

Three GTM use cases showing buyer and seller perspectives, ATS/AAP/ATR flows, and outcome distributions. Each answers the same three questions:

"Is a human required in order for us to perform this transaction?" "Where can we realistically provide agentic transactions as a customer segment?" "How can we empower our existing customers with this workflow?"

Use Case 1 — M&A Due Diligence Research (Private Equity)

The scenario: A PE firm deploys a research agent to assemble first-pass due diligence packages across three proprietary data sources — financial data, litigation records, and media intelligence — within 90 minutes of a new target entering the pipeline.

Seller opportunity: Per-query pricing aligned to deal-cycle consumption patterns. New agent buyer segment invisible to current CRM pipelines.

Human required? ~2% — first-time provider access above spend threshold only. 97% fully autonomous.

ACP path: DISC → PARSE → EVAL → APPROVED → PAY (prepaid credit / x402) → ATR

→ Full flow diagram and schema examples

---

Use Case 2 — Supplier Contract Screening (Legal / Compliance)

The scenario: A Fortune 500 legal ops team deploys a compliance screening agent to run annual and quarterly supplier screening across OFAC sanctions, ESG violations, adverse media, and litigation records — 200 suppliers, three data sources.

Seller opportunity: Continuous quarterly consumption replacing annual project-based access. ATR audit record satisfies Chief Compliance Officer regulatory defensibility requirements directly.

Human required? ~6% — partial sanctions matches, preliminary ESG investigations, ambiguous jurisdiction only. 89% approved autonomous, 5.5% auto-rejected.

ACP path: DISC → PARSE → EVAL → APPROVED / REJECTED / ESCALATED → PAY (prepaid credit wallet) → ATR

→ Full flow diagram and schema examples

---

Use Case 3 — Commercial Underwriting Enrichment (Financial Services / Insurance)

The scenario: A commercial lines carrier deploys an underwriting enrichment agent to pull property valuation, CAT exposure, industry loss history, and financial stability data for 150–200 weekly submissions — reducing manual data assembly from 25–40 minutes to 5–10 minutes per submission.

Seller opportunity: Existing API customer expansion. No new contract, no new sales cycle. Outcome-based pricing ($0.45/enriched submission) replaces capped annual license. Monthly prepaid wallet ($2,500 budget) maps to enterprise procurement patterns.

Human required? ~3% escalations (CAT score >85, below-investment-grade financial stability, valuation anomalies). Human remains in loop for binding decision — not for data assembly.

ACP path: DISC → PARSE → EVAL → APPROVED / ESCALATED → PAY (prepaid credit wallet, auto-topup) → ATR

→ Full flow diagram and schema examples

---

Repository Structure

acp/
├── schemas/
│   ├── ats-schema.json          # Agent Terms Schema (JSON Schema draft-07)
│   ├── aap-schema.json          # Agent Authorization Policy
│   └── atr-schema.json          # Agent Transaction Record
│
├── reference/
│   └── acp/
│       ├── evaluator.py         # Core policy evaluation engine
│       ├── atr_builder.py       # ATR production with integrity hashing
│       ├── agent_client.py      # Full ACP flow client
│       └── mock_provider.py     # Mock ATS-enabled data provider servers
│
├── examples/
│   ├── ats/                     # Example ATS documents
│   └── aap/
│       └── enterprise-manda.json  # M&A research agent policy
│
├── tests/
│   └── test_acp.py              # Full test suite (25 tests, stdlib only)
│
└── docs/
    ├── acp-flow.md              # Mermaid flow diagram + decision reference
    └── workflow-examples/
        ├── uc1-pe-research.md
        ├── uc2-legal-compliance.md
        └── uc3-financial-services.md

---

Quick Start

No external dependencies. Python 3.10+ stdlib only.

git clone https://github.com/fxops-ai/agent-commerce
cd agent-commerce

# Run the full test suite
PYTHONPATH=reference python3 tests/test_acp.py

# Run a live demo (starts mock servers, runs full flow)
PYTHONPATH=reference python3 -c "
import json
from acp import ACPAgentClient
from acp.mock_provider import MockProviderServer

server = MockProviderServer()
server.start()

with open('examples/aap/enterprise-manda.json') as f:
    aap = json.load(f)

# Disable new-provider escalation for demo
aap['escalation_rules']['escalate_on'] = []

client = ACPAgentClient(
    aap=aap,
    agent_id='did:web:acme.com:agents:demo',
    known_providers={'com.capitalintel.dataproducts'}
)

result = client.access(
    'http://localhost:8001',
    task_description='M&A due diligence: company financial analysis'
)

print(json.dumps(result['atr'], indent=2))
server.stop()
"

---

For Data Providers

Publish an ATS document at /.well-known/agent-terms.json on your domain. Agents will discover it automatically during task execution.

Minimum viable ATS:

{
  "ats_version": "0.1.0",
  "provider": { "id": "com.yourco.data", "name": "Your Co Data", "jurisdiction": "US",
                "certifications": ["SOC2_TYPE2"] },
  "product": { "id": "your-product", "name": "Your Data Product",
               "category": "financial_company_data",
               "description": "What the data is and what it covers.",
               "freshness": { "update_frequency": "daily" } },
  "pricing": { "currency": "USD",
               "settlement_methods": [
                 { "method": "prepaid_credit", "min_purchase_usd": 5.00,
                   "topup_via": ["stripe", "invoice_net30"] }
               ],
               "models": [{ "model_id": "per-query", "type": "per_query",
                            "unit_price": 0.05, "unit": "single record" }] },
  "access": { "protocol": "rest_json", "endpoint_uri": "https://api.yourco.com/data",
              "auth_method": "api_key" },
  "usage_policy": { "permitted_uses": ["internal_analysis"],
                    "prohibited_uses": ["model_training"],
                    "redistribution": "derived_only",
                    "downstream_agents": "same_org_only" },
  "legal": { "terms_uri": "https://yourco.com/agent-terms",
             "governing_law": "State of Delaware, United States",
             "auto_renew": false,
             "agent_binding": "agent_acts_as_agent_of_buyer" }
}

See the full SaaS Provider Implementation Guide and complete ATS schema for all available fields.

---

Payment Rail Integration

ACP is payment-rail agnostic. The authorization decision (ATS + AAP → ATR) is independent of how money moves. Supported settlement methods in the current reference implementation:

Method	Best for	Notes
prepaid_credit	Per-query, per-record micropayments	Recommended for enterprise. Human tops up wallet via Stripe or invoice; agent draws autonomously. Maps to standard procurement patterns.
x402	Autonomous micropayments, crypto-native	HTTP-level protocol. No human required per transaction. Ideal for high-frequency sub-cent queries.
ap2	Google ecosystem agents	Google Agent Payment Protocol.
stripe	Subscriptions, prepaid wallet top-ups	Not recommended for direct per-query agent billing. Use as topup mechanism for prepaid_credit instead.
usdc	Stablecoin settlement	Via Nevermined or direct.

Note on Stripe: Stripe's per-transaction fee structure (2.9% + $0.30) makes it unsuitable for direct per-query micropayments. Use prepaid_credit with topup_via: ["stripe"] for enterprise-friendly billing that maps to standard procurement workflows. See Issue #2 — Prepaid Credit Wallet for full schema proposal.

---

Adopters & Implementations

This section tracks known ATS publications, agent implementations, and community contributions. If you've published an ATS document or built an ACP-compatible agent, open a PR to add your entry below.

Data Providers (ATS Published)

No entries yet. Be the first — open a PR to add your ATS endpoint.

Provider	Domain	Data Category	Pricing Model	Date Added
—	—	—	—	—

Agent Implementations (AAP Deployed)

No entries yet. If you've deployed an agent with an ACP-compatible AAP policy, open a PR to add your entry.

Organization	Use Case	Agent Type	AAP Scope	Date Added
—	—	—	—	—

Community Contributions

Contributor	Contribution	Link
—	—	—

---

To add your implementation:

1. Fork this repository
2. Edit this README — add a row to the relevant table above
3. Open a pull request with the title [Adopter] Your Organization Name

For ATS publication validation, use the ATS schema to confirm your document is spec-compliant before submitting.

---

Open Issues & Roadmap

Issue	Title	Milestone
#1	Add prepaid_credit as first-class settlement pattern	v0.2.0

v0.2.0 priorities:

• Prepaid credit wallet settlement pattern (structured settlement_methods field)
• TypeScript reference implementation
• ACP validator CLI tool
• LLM-backed semantic policy evaluation
• Enterprise AAP templates — healthcare, legal, financial services

v1.0.0 milestone:

• Stable schema version (no breaking field changes)
• IANA registration of agent-terms well-known URI suffix
• Formal interoperability test suite

---

Spec Version

ACP v0.1.0 — Draft for public comment.

Schema stability: ATS, AAP, ATR schemas are not yet stable. Field names and enum values may change before v1.0.0. Breaking changes will be communicated via the versioning block in the schemas.

---

License

• Code (reference implementation, tests, tooling): Apache 2.0
• Specification (schemas, documentation): CC BY 4.0

Both licenses permit free use, modification, and distribution. Attribution required for specification derivatives.

---

Contributing

Issues and PRs welcome at github.com/fxops-ai/agent-commerce.

Priority areas for contribution:

• Real-world ATS documents from production data providers
• TypeScript reference implementation
• LLM-backed semantic policy evaluation (replacing keyword heuristic in _check_task_relevance)
• ACP validator CLI tool
• Enterprise AAP templates for healthcare, legal, and financial services verticals

---

Further Reading

• Full technical breakdown — @SaaSCRO on X
• The agent buyer is coming — LinkedIn
• fxops.ai/agent-commerce