> ## Documentation Index
> Fetch the complete documentation index at: https://humboldt.metaphi.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Project instructions

> Tell Humboldt how this repository works in an `AGENTS.md` at the repository root. Humboldt reads it at the start of every session.

## One file: AGENTS.md

```
repo/
  AGENTS.md
```

Write what a new engineer needs on day one: build and test commands, conventions, where things live, how work is delivered, what to avoid. Plain markdown. Humboldt reads it whole and never changes it on its own.

`CLAUDE.md` and `README.md` at the same level are read too. When more than one is present, `AGENTS.md` has the last word.

## Instructions for one folder

Put an `AGENTS.md` inside the folder. Humboldt reads it the first time it reads or edits a file in that folder or beneath it.

```
repo/
  AGENTS.md                 # read at session start
  batch/
    AGENTS.md               # read when Humboldt first works in batch/
  online/
    AGENTS.md               # read when Humboldt first works in online/
```

Both files apply once read; the folder's file is read last. After a compaction, a folder's file comes back with the next read there.

## Linking your documentation

Keep `AGENTS.md` as the entry point and name your documents in it: a design document, standards, test data, the delivery process. Humboldt reads a named file when the work calls for it.

```markdown theme={"dark"}
# AGENTS.md
- Coding standards: docs/standards.md
- Test data and how to regenerate it: tests/README.md
- Delivery: every accepted change is pushed to `development`; see docs/delivery.md
```

A folder of markdown files on its own, such as `agents/*.md` or `rules/*.md`, is not read. Name the files in `AGENTS.md`, or put the folder-specific ones in that folder's `AGENTS.md`.

## Let Humboldt write it

```bash theme={"dark"}
hum "write an AGENTS.md for this repository"
```

Humboldt reads the repository and drafts the file. Review it, change what you want, commit it. It applies from the next session.

## Where Humboldt starts

Humboldt reads instruction files from the directory you start `hum` in, and every parent up to the repository root. Start `hum` at the root to have the root file apply everywhere. Nothing above the repository root, or in your home directory, is read.

## MCP servers

Tool servers are configured in `~/.hum/config.toml`, separately from instructions. See [MCP servers](/mcp-servers).