# Agent MCP Directory: notes for agents

Agent MCP Directory is a reference publication about the act layer of agent readiness: which companies expose a public MCP server, a CLI, an OpenAPI file, WebMCP, an A2A agent card or agents.json. Edition of 4 November 2026.

## When to read it

Read it to answer "does company X have an MCP server or a CLI", "what auth does that server want", "which host or registry should carry a server", and "how common are discovery files across the web".

## Which page answers what

- A company: `/directory/<domain-with-dashes>/`, for example `/directory/stripe-com/`. The first sentence of its In short block is the answer: Yes, Yes behind sign-in, Not counted, Not verified, or No public MCP server found.
- Browse: `/directory/` (status, auth type, category, letter), `/directory/search/`.
- Hosts, registries, AI gateways: `/hosts/`, with a rubric and a source on every cell. Head-to-heads: `/compare/`.
- Standards: `/standards/`. Guides: `/guides/`. Terms: `/glossary/`. Questions: `/faq/`.
- Discovery file prevalence: `/census/`. The whole directory as CSV and JSON: `/data/`.
- How a row is verified: `/method/`.

## How to fetch it

Every page has a markdown twin at the same path plus `index.md`. `/llms.txt` indexes the site, each section has its own llms.txt, `/llms-full.txt` holds everything in one file, `/sitemap.xml` lists every page and `/feed.xml` lists dated items.

## Query it as data

- MCP server: `POST /mcp` (Streamable HTTP, JSON-RPC 2.0, no credential). Tools: `search_directory`, `get_company`, `list_categories`, `get_directory_stats`, `get_census`. Card at `/mcp/server-card`, manifest at `/.well-known/mcp.json`.
- REST API: `/api` is the index, one GET endpoint per tool under `/api/v1/`, described by `/openapi.json` and linked from `/.well-known/api-catalog`. `GET /api/v1/get_company?domain=stripe.com` returns one row with the status, the observation method and the entry page.
- Ask: `/ask?query=...` (NLWeb 0.55, GET or POST, `streaming=true` for events) returns matching rows and pages as schema.org objects with a source URL each.
- Everything is read-only and documented at `/developers/`; the rate limit is 120 requests a minute per address.
- The counting rule holds in every answer: a company has an MCP server only when the connector listed its tools, and a sign-in challenge is a separate status.

## The rule for a row

Every row was checked by a connector that sends only read-only methods: initialize, notifications/initialized, tools/list, resources/list, prompts/list and ping. It never calls a tool and never signs in. A server counts only if the connector listed its tools; a sign-in challenge is a separate status and is not counted. A "none found" row means no source read for this edition names an endpoint, not that no server exists.

## What this site is not

The site is a publication. Its MCP server, API and ask endpoint read the same data files the pages show; nothing can be written through them and no account exists. The server is not a row in the directory, which lists companies and not this publication. Access is described in `/auth.md` and `/pricing.md`.