llms.txt

A public, agent-discoverable manifest at two well-known paths per endpoint — the front door for any LLM that wants to understand your data before connecting.

What it is

llms.txt is an emerging standard, similar in spirit to robots.txt or sitemap.xml, that gives AI agents a structured, human-readable summary of a site or service. ShipMCP publishes one for every endpoint, automatically, with no auth required.

Where to find it

Every active endpoint serves llms.txt at two paths:

https://shipmcp.io/<slug>/llms.txt
https://shipmcp.io/<slug>/.well-known/llms.txt

The .well-known path follows RFC 8615; the root path matches what most current llms.txt consumers look for.

What it contains

Each llms.txt describes the endpoint in markdown sections an agent can reason over:

• A one-line purpose statement and the source of the data (file name, URL, last refresh)
• The full schema — every table, every column, primary keys, foreign keys, row counts
• The tool catalog — every tools/list entry with arguments and a one-line behaviour summary
• FK relationships described in prose, so agents can plan multi-step joins
• Connection instructions — the MCP URL, the auth shape, a Claude Desktop config snippet

Example

# ShipMCP — `products`

> MCP endpoint exposing products, categories, and orders from products.csv.
> Last ingested: 2026-04-26T18:14:02Z · 1,247 rows.

## Tools

- list_products — paginated listing
- search_products — full-text search across name, description
- filter_products — by category, price range, in_stock, created_at
- get_orders_for_product — FK join: orders.product_id → products.id

## Schema

### `products` — ~1,247 rows
  - id (int4, PK)
  - name (text)
  - price (numeric)
  - category (text)
  - in_stock (boolean)
  - created_at (timestamptz)

…

When it regenerates

• On every successful ingest — initial provisioning and any append job.
• When a token is rotated (the connection snippet always reflects the current shape, but tokens themselves never appear in llms.txt — the file is public).

How agents use it

When you connect Claude Desktop (or any MCP client), the initialize response includes an instructions string pointing at this file. Compliant clients fetch it and put the contents into the agent's context — so even before the first tools/list call, the agent knows your endpoint exists, what it contains, and how to query it.