How it works
The home, the brain, the rails, and skills and packs. What each one is and where to read more.
What makes an agent useful is not which model answers. It is what the agent knows before it starts, what it is allowed to do, and what gets recorded when it acts. Those live in files around the model. The model is a setting.
Your host is the agent CLI you run: Claude Code or Codex, and one home works from either. The plugin adds four things to it.
The home
The home is the agent's folder, the one you launch your host from. It is plain markdown, so nothing needs a server and it survives a change of model or tool.
| Piece | Where | Read more |
|---|---|---|
| Who the agent is, who you are, and the operating manual | SOUL.md, IDENTITY.md, USER.md, AGENTS.md | Agent home |
| The compiled wiki: notes about you, concept articles, active memory | knowledge/ | How the brain works |
| Raw intake: saved sessions, the inbox, the feedback log | knowledge/raw/ | Capture |
| Projects and tasks, one file each | projects/ | Tasks |
| Short-lived outputs: briefings, plans, audits | reports/ | Daily rhythm |
The brain
The brain is what the agent knows: a wiki compiled from real work and loaded before every session. It is not a search box over your documents. Search finds fragments each time it is asked; the brain keeps a compiled understanding, so the answer exists before the question.
- Capture. Every session is saved raw when you exit. Notes, files and links you hand over land in the same intake.
- Compile. A compile pass distils the raw intake into the wiki and prunes stale entries, so the index stays small enough to load. Sync runs it for you.
- Load. The wiki is in context before your first prompt, on either host. What loads when.
- Feedback. Corrections become learnings that load every session, and a review skill turns them into changes to the agent itself. See Self-evolution.
The rails
The rails are what make the agent safe to use: what it may do, and what gets written down when it acts.
| Rail | What it does | Read more |
|---|---|---|
| The manual | AGENTS.md says how the agent works and how you like to work. Every host reads it at session start. | Agent home |
| Permissions | Allow and deny lists, a sandboxed shell, and keys in a store the agent's own tools cannot read. | Configuration |
| Hooks | Small programs the host runs at set moments: one loads the context, one saves the session, one checks every shell command. Skills advise; hooks enforce. | Hooks |
| The record | Every session saved to a dated file, every task change threaded, the folder under git. The dashboard shows it on one page. | History |
In Claude Code two more ride along: mods (instant commands and progress above the prompt) and notices (reminders for upkeep only you can start).
Skills and packs
A skill is a workflow the agent runs when you ask: a briefing, a plan, a review. A tool is something it can reach: a pull request, a database, a browser. A pack is a group of tools that stays off until you add its key. See Skills, MCP tools and Connect it.
How a session runs
You launch from the home. A hook loads the brain and the task board. You type. Skills run when you ask; tools fetch and write, and hand prompts back for your own session to think with, so nothing bills a metered key. On exit the session is saved for the next compile.
Work is tasks, not chats: one markdown file per task, grouped into projects, with parallel work in worktrees. The architecture page shows how the host, the plugin and the home fit.
One plugin, your own agent
The plugin is the same for every home. What goes into a brain is yours. A seed bundle hands an agent's character and starting knowledge to a teammate, and several homes keep separate brains apart on one machine.