Claude Code

Setup is harness-agnostic — see the quickstart for creating a hall and opening your first session. This page covers what is specific to Claude Code once you're there.

What lands in .claude/

A Claude Code session materialises a .claude/ directory at the view directory's root, alongside the promoted repos:

.claude/commands/ holds only ivar-*.md — the glob is deliberately narrow so a user's own commands in that directory stay untouched and committable. .claude/skills/ carries every skill from both roots — the hall's committed skills and your personal ones — materialised under its bare id; the directory itself is gitignored because each child is a derived symlink back into .ivar/. CLAUDE.md is the provider-native instructions file Claude Code reads at startup.

The hall's MCP servers land beside it in .mcp.json, rewritten in full from ivar.json on every sync — see MCP servers.

The read-only guard for repos you haven't promoted is enforced by .claude/hooks/ivar-execution-guard.sh, registered as a PreToolUse hook in .claude/settings.json. It's a real hook plus a real cleared write bit — not a prompt asking the agent to behave.

Add it to an existing hall

The quickstart covers ivar init. If the hall already exists with a different provider, add Claude Code alongside it:

ivar provider add claude-code

Where /ivar-discovery comes from

It is a slash command materialised from .claude/commands/ivar-discovery.md, so you type it in the Claude Code chat rather than in a shell.