agents.txt is a discovery manifest a store publishes at /agents.txt. It is the first file an AI agent reads on arrival, and it points the agent to everything it needs to transact: the store's Terminal (its agent API), the identity issuers the store trusts, its checkout surface, and its machine contract. Think of robots.txt for crawlers, and agents.txt for buyers.
robots.txt governs crawling: which bots may read your content, and which paths are off limits. It has been the web's convention for crawlers for decades.
agents.txt answers a different question. When an agent arrives not to read but to transact, it needs to know where your agent surface is: your API, your checkout, the credentials you accept. agents.txt is that pointer. It does not replace robots.txt; it sits beside it for a new kind of visitor.
An agent that wants to buy from you, rather than index you, starts at agents.txt.
It is a short, plain-text file. Each line is a pointer the agent follows. The core fields:
| Field | What it points to |
|---|---|
| Terminal | The store's agent API, where identity, rate limits, and capabilities are enforced |
| KYA-Issuers | The know-your-agent credential issuers the store trusts |
| OpenAPI | The machine contract, so an agent can build a typed client with no out-of-band lookup |
| Checkout | The checkout surface an agent calls to complete a purchase |
| Directory | A discovery endpoint for finding businesses and their Terminals |
| MCP-Endpoint | The hosted MCP endpoint, for agents that speak the tool-call transport |
| Pricing-Hint | Non-binding pricing signal; the canonical terms live at the Terminal |
The file itself is a thin pointer, not a permissions standard. It does not grant or deny access. Identity verification, rate limits, and capability rules all happen at the Terminal the manifest points to. The manifest just gets the agent to the right door.
Once the agent has the manifest, it follows the pointers:
Every step is machine-readable and consented, which is what makes an agents.txt store legible to a buyer's agent instead of something it has to scrape. See how the checkout protocols fit together.
Serve a plain-text file at /agents.txt (and, by convention, at /.well-known/agents.txt) with the pointers above. That is the whole mechanism: no build step, no framework.
Facet stewards the open agents.txt manifest and publishes a correct one for a store the moment it is mounted, pointed at that store's Terminal, the trusted issuers, and its checkout. This is part of the layer Facet installs above an existing store: the store stays exactly as it is for human shoppers, and the agent-facing surface, discovery manifest included, is generated and kept current for it.
Mount your store with Facet and it publishes a correct agents.txt for you, pointed at your Terminal, checkout, and trusted issuers. About five minutes, no theme changes.
Make your store agent-ready →
