AI shopping agents decide what to recommend and where to buy. What they lack is a trust and facts layer: a machine-readable place where a store states what it sells, what its policies are — and where those facts can be checked. AgentFront is that layer. One JSON file, any platform, readable by any agent.
Like llms.txt, this is a proposed standard: published openly by Frictionless, versioned, free to adopt, no lock-in. The manifest works whether you host it yourself or let Frictionless host and verify it.
Self-hosted: publish /agentfront.json at your domain root, following the JSON schema. Hosted & verified: Shopify merchants opt in via the Frictionless app — the manifest is then served at frictionlessai.net/agent-profile/<your-domain> with facts computed by deterministic checks and marked verified_by.
| field | type | meaning |
|---|---|---|
profile | const | "agentfront" |
version | integer | spec version, currently 1 |
store | string | the store's domain |
ai_readiness | object | score 0–100, tier, per-check evidence |
policy_urls | object | refund, shipping — agents check terms first |
products | array ≤100 | product entry-point URLs |
verified_by | string | third party that computed the facts (optional) |
source | enum | self or hosted |
curl "https://frictionlessai.net/api/v1/agentfront-validate?url=yourstore.com"
Returns valid, errors and warnings against the reference validator. A spec without a validator is a PDF; this one ships with infrastructure.
Agents can call our checks directly as MCP tools — “can I buy from store X, and should I recommend it?” answered from live deterministic checks, quotes included, nothing invented.
POST https://frictionlessai.net/mcp (JSON-RPC 2.0, Streamable HTTP) tools: store_ai_readiness · store_answer_audit · store_prompt_coverage