llms.txt

/llms.txt is a pointer. It names the endpoints, tools, and collections an agent can use on this host, and links to per-collection index shards. It is not a copy of the site.

An agent that fetches it learns where to connect in roughly a kilobyte, instead of reading the site as text.

Fetch it

curl -s https://<host>/llms.txt

# ArcBlock

> ArcBlock builds decentralized identity, user-owned data, Blocklet
> infrastructure, and AI-native application systems.

## For agents

- MCP (Streamable HTTP): https://<host>/mcp
- Server card: https://<host>/.well-known/mcp.json
- AFS over HTTP (JSON-RPC): POST https://<host>/api/afs/rpc
- Tools: search_content, list_content, get_content
- Collections available to agents: articles, docs, glossary, products
- Anonymous public read; writes require DID-Connect auth

Locales: en, zh, ja, zh-TW (base: en). …

Full-text version (all collections, never truncated): https://<host>/llms-full.txt

## Docs (1268 items)

Index: https://<host>/llms-docs.txt
Full text: https://<host>/llms-docs-full.txt

The document above is 1,142 bytes. The For agents block is the part that matters: it hands over the MCP endpoint, the server card, and the JSON-RPC endpoint, so a client that can do better than reading text does not have to read text.

Shards

Each collection that declared an llms face gets an index shard and a full-text shard.

A collection without the face has no shard:

curl -s -o /dev/null -w '%{http_code} %{size_download}\n' https://<host>/llms-docs.txt      # 200 154869
curl -s -o /dev/null -w '%{http_code} %{size_download}\n' https://<host>/llms-articles.txt  # 404 9

Which collections have the face is a per-blocklet decision; see Declare what agents see.

Sizes are not symmetric

On a site with about 1,300 documents the pointer is around 1 KB, the docs index shard is around 150 KB, and /llms-full.txt is several megabytes.

The shards exist so an agent can take the one collection it needs. llms-full.txt is there for completeness, not as the intended entry point. Fetching it costs more context than connecting over MCP and asking a question.

When to use which

llms.txt names the first two, so an agent that starts at the text surface can move up to a better one.