Bring your Claude Code setup

Put the skills, subagents and hooks you keep in your local ~/.claude into the workspace’s Claude Code configuration folder, and sessions use them as they are. The skills you placed can be checked on a Console settings screen.

How it works

What you do	What the agent does
Put skill, subagent and hook files in the configuration folder	Loads and uses them from the next session it starts
Check the skills you placed in ⚙Settings → Skills	—
Ask an agent to edit your own skills and the like	Edits your personal skills, subagents and hooks, only when asked

What goes where

Part of it is covered by settings screens.

Where it lives in your local Claude Code	In the workspace
The global `CLAUDE.md`	⚙Settings → Agent instructions (Write down how you work, once)
MCP server configuration	⚙Settings → MCP servers (used by sessions) (Add an MCP server to your sessions)
Skills, subagents, hooks	Files in the configuration folder (the table below)

The configuration folder is /var/lib/af/claude. It can differ per deployment, so check by running echo $CLAUDE_CONFIG_DIR in a shell. It is the counterpart of ~/.claude on your own machine.

What	Where	Shared with the team
Hooks, skills and subagents for the team	The repository’s `.claude/` (committed)	Yes
Your skills	`<config folder>/skills/<name>/SKILL.md`	No
Your subagents	`<config folder>/agents/<name>.md`	No
Your hooks	The `hooks` key of `<config folder>/settings.json`	No

Claude reads these when a session starts. A change applies to sessions started afterwards
Agent Fleet does not inspect these files. Before relying on them, try yours in a real session
Skills in skills/ also appear in the session view’s skill picker
The folder lives on dedicated storage. Stop / start, recreating the workspace and cleaning home leave it alone

Edit them

Edit from a shell or a Terminal (CLI) session. The Console’s file browser does not show this folder, because it also holds sign-in data.

You can also ask an agent to edit them. An agent touches these files only when asked, and only your personal skills, subagents and hooks. Credentials and Agent Fleet’s own hook entries are off-limits to agents too.

Check your skills

⚙Settings → Skills lists each agent’s personal skills, read-only. Besides Claude, Codex, OpenCode and Muse each have their own tab.

The Claude tab of Skills: the folders read (/var/lib/af/claude/skills and commands), your skills, a command shown as shadowed, and an Agent Fleet skill

What it shows	Contents
Folders read	The folders the agent reads skills from. One that does not exist yet is marked “not created”, so you know where to create it
Your skills	Skills you placed
Agent Fleet skills	Skills Agent Fleet writes at every start (`af-*`). Edits to them are overwritten
Bundled with the CLI	Skills built into the CLI. They have no files

A skill that loses to another of the same name is struck through as “shadowed”, and the winner is named
Open a skill for its path, its SKILL.md and the other files in its folder
Cursor, Antigravity, Kiro and Copilot have no personal skills folder

Skills kept in a repository are not listed here. Right-click the repository’s row in the project tree and open “Skills…” to see, per agent, the folders it reads, the copy that wins and the files’ contents. You can also check how the list changes when a session starts in a subdirectory.

Writing hooks

Do not edit Agent Fleet’s hooks. settings.json holds entries that report session state to the Console: the ones whose command runs session-status or session-push-notification. Without them that session’s state no longer shows in the Console (the Agent adds them back at start)
The PreModelSwitch entry that keeps /model from stopping at claude’s “Switch model?” is Agent Fleet’s too. After a switch, the next message rereads the whole history at the uncached price
Turn the RTK entry (rtk hook claude) on and off with RTK on the Claude Code card, not by editing (Check prices and usage)
Your own hooks and other keys stay. A hook of yours on the same matcher (Bash, say) survives RTK being switched on and off
Do not put the word session-status in a command. It would be mistaken for an Agent Fleet entry
Write them for the workspace. You can rely only on what the image has (bash, jq, node, python) and what you installed into ~/.local yourself. /usr, /opt, /tmp and the like revert to the image, so keep scripts in your home
Start claude -p from a hook as env -u AF_SESSION_NAME claude -p …. Otherwise the child claude’s activity is reported as the parent session’s, and its state can go wrong. Run your hook once and check that the session’s state stays right