An open specification proposal for the machine web
Make your content verifiably readable by AI agents.
index-ai is an open specification that lets websites expose a structured layer for agents — who you are, what content exists, and — where you declare it — exactly how big each piece is, measured and verified rather than guessed.
License: MIT (code) · CC-BY-4.0 (spec)Validator: @hardmachinelabs/index-ai-validator
index-ai intro — the Agent View in action, playing in loop with audio (unmute via the video controls if autoplay starts muted).
One site, four rungs of conformance
Start at Level 1, climb as far as you need. Partial conformance is valid to ship.
1 AI Manifest ~15 minutes
Identity, publisher, freshness, policy. One HTTP call answers: “What is this site?”
/.well-known/index-ai.json
2a Agent Index ~1–2 hours, static site
A flat list of content nodes — each with a clean-text endpoint and an exact measured size. An agent can navigate directly, no scraping.
/agent-index.json
2b Agent Graph +2–4 hours
Typed relationships between nodes forming a navigable graph. From “hotels” to “hotels-casablanca” to “hotels-casablanca-luxury” — without parsing HTML.
/agent-index.json + relations
3 Query Interface MCP
A typed API over the Agent View. The agent sends a query with filters and receives exactly the data requested — minimum possible tokens.
