Every nest init scaffolds includes an llms.txt - a one-page, plain-text index of the project written for coding agents. An agent that lands in the directory (or fetches the file from a served nest) learns in one read what this thing is, what data it holds, and how to query it - without crawling the source.

What’s in it

Generated from the decode registry - the same source of truth behind schema.json, /tables, and the MCP schema tool - it contains:

• What this is: a self-hosted blockchain index on <chain>; query it locally, there is no third-party API.
• The contracts: every alias and address the nest indexes.
• The tables: one line per {alias}__{event} table with its columns.
• The live HTTP API: the handful of endpoints that matter (/, /tables, /table/{name}, /entity/{id}, /sql?q=…, /balances, /balance/{address}).
• The MCP pointer: run nuthatch mcp for the tool surface, fully offline.

Lifecycle

• Generated, not authored. init and add both write it; nuthatch schema regenerates it (with schema.json and the semantic.toml footguns) after you hand-edit the config. Don’t maintain it by hand - it always reflects the tables the config actually produces.
• Pinned in the bundle. A nest’s llms.txt is part of its content-addressed inputs, so a published nest carries its own agent documentation, verified by hash like everything else in the registry.

Alongside llms.txt, init also scaffolds a per-nest .claude/skills/nuthatch querying skill - the deeper counterpart for agents that support skills. The distinction from the builder skill: the scaffolded skill teaches an agent to query this nest; the builder skill teaches an agent to drive nuthatch.