Per-Employee Setup

1. Grant per-person sync access

   Give the employee access to compiled/<person-id>/ only — a per-person deploy key or a per-person remote. Never hand out access to the master or to another person's slice.

   This is a real boundary, not a convention. The master vault is the tenant wall; a person who can read the master can read everyone. Sync access is always to a compiled slice.
2. Clone and open in Obsidian

   On their machine:

   1git clone <their-vault-remote> ~/brain

   Open ~/brain as an Obsidian vault (the pilot uses the Obsidian Git plugin to sync). Everything is plain markdown, so any editor works too. Obsidian's built-in graph view works out of the box — see See your brain as a graph.
3. Point an agent at the vault

   The compiled vault is self-describing — its root AGENTS.md and CLAUDE.md carry the full protocol.

   Claude CodeHermes

   Nothing to install. Open ~/brain in Claude Code; the vault's CLAUDE.md carries the routing and promotion protocol.
4. Build the search index and register it as a tool

   12brain index --vault ~/brain
   claude mcp add brain -- brain mcp --vault ~/brain

   brain index builds a local hybrid search index so the agent can query the vault instead of reading files by hand; the MCP server exposes it as brain_search / brain_read / brain_links / brain_graph / brain_recent / brain_facts tools. The index is keyword-only until an embedding provider is configured (BRAIN_EMBED_*) — set one to add semantic search.

   .brain/ is machine-local and gitignored, so it never syncs — each device builds its own index. Because it's built only from this compiled slice, it can never surface anything the person can't already read.

The daily flow

• In: transcripts and notes reach People/<you>/Inbox/ — dropped in as a file, sent through an email/chat/voice/upload channel backed by brain ingest, or pushed automatically by an outside service via brain webhook. The agent summarizes, extracts decisions and action items, and routes them per the vault protocol. Articles, links, PDFs, and screenshots get distilled into the shared Company/Intel/ wiki, never archived whole.
• Back up: edits sync to the server, which runs a single scheduled command — brain cycle — that writes back every person's edits, provisions any self-service client requests, routes share requests (queuing shares, auto-applying revokes), sweeps agent drafts into the pending queue, and recompiles everyone, in that order. (A rejected edit reverts on the recompile and the command exits non-zero — as does an owner-mismatch client or share request — so cron alerts.)
• Sharing: the agent drafts a promotion; an admin approves it with brain promotions approve <id> --master … --approver admin — the approver is an admin-role identity from org.yaml (a dedicated operator account, not an employee). A Teams/<team>/ target can also be approved by that team's lead instead; anyone else is refused. See Who approves. A space's owner can also share or revoke a colleague's access to that space directly — a separate, owner-initiated flow from the admin-curated promotions gate — via brain shares; share requests still queue for approval by their decider — the recipient for a person-share, a team lead for a team-share, an admin for company-wide — while revoke requests apply immediately. See Sharing a space.
• Search: after each pull, the agent runs brain index --vault ~/brain (cheap — only changed files re-embed) and queries the vault through the brain_search MCP tool. See Retrieval.