ai-policy.json

Machine-readable AI agent permissions for websites.

ai-policy.json is a simple JSON file that website owners deploy at /.well-known/ai-policy.json to declare what AI agents can and cannot do with their content. It's the only format designed to grow from content-usage permissions into action-level permissions for agentic AI.

Quick Look

{
  "version": "1.0",
  "domain": "example.com",
  "permissions": {
    "training": "block",
    "search": "allow",
    "ai_input": "allow"
  }
}

Three fields. One file. Any agent can parse it in a single request.

• training: Can AI models train on this content?
• search: Can AI search engines index and cite this content?
• ai_input: Can AI systems use this content at inference time to generate responses?

Why This Exists

The current AI permissions landscape is fragmented across 8+ formats: robots.txt, Cloudflare Content Signals, TDMRep, ai.txt, llms.txt, Terms of Service, meta tags, HTTP headers. Research across 1,000,000+ domains found that approximately 90% have no machine-readable AI policy at all, and thousands have conflicts between different signals.

ai-policy.json consolidates these signals into one structured document and serves as the canonical source of intent. Tools that generate ai-policy.json also generate the corresponding robots.txt directives, Content Signals, TDMRep headers, and ai.txt, so your policy works everywhere, not just with agents that support this format.

How It Works

For website owners: Use a tool like Maango's Policy Builder to set your preferences. You'll get ai-policy.json plus all the other formats (robots.txt additions, Content Signals, TDMRep headers, ai.txt) generated automatically. Deploy them to your site. Done.

For agent developers: Fetch /.well-known/ai-policy.json from any domain to get its AI policy in structured JSON. Or use the Maango API to check any domain. It synthesizes every signal (robots.txt, Content Signals, TDMRep, ToS, and ai-policy.json) into one answer.

# Direct fetch
import httpx
r = httpx.get("https://example.com/.well-known/ai-policy.json")
policy = r.json()
if policy["permissions"]["training"] == "block":
    # don't use this content for training

# Or via Maango API (covers domains without ai-policy.json too)
r = httpx.get(
    "https://api.maango.io/v1/domain/example.com",
    headers={"Authorization": "Bearer YOUR_KEY"}
)

Vocabulary Alignment

The three permission fields align with existing standards:

ai-policy.json	Cloudflare Content Signals	IETF AIPREF
training	ai-train	Foundation Model Production
search	search	Search
ai_input	ai-input	(under discussion)

This is intentional. ai-policy.json is interoperable with the most widely deployed AI permissions infrastructure, not competing with it.

Full Specification

📄 Read the full spec →

The specification covers discovery, field definitions, permission semantics, bot directives, interoperability mappings to other formats, versioning, security considerations, and examples.

Schema Validation

Validate your ai-policy.json against the JSON Schema:

schema/v1.0.json

Examples

File	Scenario
minimal.json	Smallest valid file, just the three required fields
blog.json	Blog that blocks training, allows search and AI input
ecommerce.json	E-commerce site with bot-specific directives
healthcare-block-all.json	Regulated site that blocks everything
open-source-allow-all.json	Open source project that allows everything

What's Next

ai-policy.json v1.0 handles content-usage permissions, i.e. what AI can do with your content. Future versions will extend into action-level permissions: what AI agents can do on your site (browse, submit forms, transact). The capabilities field is reserved for this purpose.

Contributing

We welcome contributions. See CONTRIBUTING.md for how to propose changes.

License

This specification is released under CC0 1.0 Universal (public domain). You can implement, extend, and use it freely.

---

Built by Maango, the permissions layer for the agentic internet.