The 2026 API Discovery Stack: OpenAPI + llms.txt + RFC 9727 Explained

Five well-known files that let any AI agent or partner discover your API without an onboarding call.

In 2026, "how do I get started?" should be a machine question. The answer is a stack of five well-known files that turn your domain into a self-describing API surface. unLocked CRM publishes all of them.

1. /openapi.json — The Contract

The OpenAPI 3.1 spec. Every endpoint, schema, security flow, and webhook in one machine-readable document.

2. /.well-known/api-catalog — The Index (RFC 9727)

A linkset that points at every API surface we expose: the OpenAPI spec, the MCP server, the webhooks contract, and the changelog. Tools that follow RFC 9727 (Bruno, Insomnia, Cursor) discover the entire stack from this single URL.

3. /.well-known/agent-skills/index.json — The Capability Manifest

AgentSkills v0.2 manifest describing what an agent can do with our API in human-readable terms. Claude and Goose use it to populate tool descriptions without parsing the full OpenAPI doc.

4. /llms.txt and /llms-full.txt — The Narrative

llms.txt is a curated, markdown-formatted summary of our docs optimized for LLM context windows. llms-full.txt is the everything-and-the-kitchen-sink version. Both are fetched by ChatGPT, Claude, and Perplexity at answer time.

5. /.well-known/ai-plugin.json — The Legacy Bridge

Maintained for backward compatibility with the original ChatGPT plugin spec. Newer agents use the api-catalog, but ai-plugin.json still gets us into a few directories.

Why All Five

Different agents prefer different surfaces. Cursor reads the api-catalog. ChatGPT reads ai-plugin and the OpenAPI spec. Claude reads llms.txt for context, then calls MCP for execution. Publishing all five means we're discoverable no matter which client lands first.

Total cost: a few static files and a deploy script. Total payoff: every AI agent on earth can integrate without a sales call.

Related Reading