5-agent Claude Code team with closed-loop self-improvement, Obsidian-synced external memory, multi-repo project constellations.

claude plugins add github:braininahat/brains-in-a-hat

# 1. Install the plugin
claude plugins add github:braininahat/brains-in-a-hat

# 2. Bootstrap your vault (one-time, per device)
bash brains-in-a-hat/scripts/vault-init.sh auto

# 3. Set the env var (add to your shell rc to persist)
export BRAINS_VAULT_DIR=~/.hatbrains

# 4. (Optional) Group a multi-repo project — drop this in each repo:
cat > .brains_in_a_hat/project.json <<EOF
{
  "projectGroup": "pulse",
  "repoKey": "pulse-model"
}
EOF

# 5. Open Claude Code in the repo. The new-roster-bootstrap hook auto-spawns
#    Librarian on your first prompt. That's it.

• Neal lives in the main thread and orchestrates. The user talks to Neal; Neal routes by intent (no CODEOWNERS, no file-pattern routing).
• Librarian is an always-on teammate that curates a flat-vault knowledge graph in your private vault repo. It pulls/commits/pushes on every TeammateIdle tick — multi-device sync via git, no Obsidian sync required.
• Critic / Builder / Scout spawn on demand. Critic switches review lenses by prompt parameter (lens=architecture|security|performance|ux|correctness); Builder reads project CLAUDE.md for domain context; Scout routes between context7, web, and own knowledge.
• A SubagentStart hook injects the last N main-thread turns + active learnings into every teammate spawn, so you never have to manually push context.
• A TeammateIdle hook re-engages Librarian to classify new transcript turns, update the vault with curation logic (update/supersede/delete — not just append), and push.

• Not 21 role-specialists with file-path routing — 5 generalist teammates with intent-based dispatch.
• Not a write-only retro archive — a curated vault that updates, supersedes, and deletes as you work.
• Not single-repo state — project constellations that share context across all your repos.
• Not Obsidian-sync-dependent — the vault is a git repo (e.g., a private GitHub repo); git is the sync primitive.
• Not a graph database server — a flat-vault knowledge graph: markdown + YAML frontmatter + wikilinks + Dataview queries at read time. Zero infrastructure.

The vault ($BRAINS_VAULT_DIR, default ~/.hatbrains) is a private git repo. On a new device:

gh repo clone <you>/hatbrains-vault ~/.hatbrains
export BRAINS_VAULT_DIR=~/.hatbrains

vault-init.sh detect (or auto) figures out which case you're in (fresh init vs clone existing).

Worktrees of the same repo share the same vault entries (logical project keying via .brains_in_a_hat/project.json → git remote → basename, in that order).

For multi-repo projects (e.g., PULSE = model + UI + paper), declare a shared projectGroup in each repo's project.json. Cross-repo learnings land in <projectGroup>--shared--learnings.md.

Open Obsidian to the vault root and you see one index.md with Dataview-driven sections: Active now, This week, Open threads, Recent decisions, Wiki entries, Group rollup. No subdirectories — the whole knowledge graph is a flat directory of markdown files. Cross-references are [[wikilinks]]; structure is frontmatter (type, projectGroup, repoKey, tags, status).

• brains-in-a-hat/CLAUDE.md — architecture and hook lifecycle (source of truth for behavior)
• brains-in-a-hat/agents/ — the five teammate persona files (librarian, critic, scout, architect, builder)
• brains-in-a-hat/hooks/neal-persona.md — Neal's full prompt (orchestrator + meta-analyst)
• brains-in-a-hat/scripts/vault-init.sh — vault bootstrap helper
• brains-in-a-hat/scripts/file-plugin-issue.sh — file issues on the plugin repo from any project (used by Builders via UPSTREAM-ISSUE-REQUEST: sentinel)

If you fork this, set these topics for discoverability:

claude-code, claude-code-plugin, claude-code-agents, agent-team, multi-agent, obsidian, obsidian-sync, ai-agents, agentic-ai, anthropic, self-improvement, external-memory, prompt-engineering, developer-tools, closed-loop-agents.

MIT.