big-brain 🧠

A second brain for you and your AI assistants. Plain-markdown notes, projects, and tasks — Obsidian-compatible, owned by you, readable and writable by any LLM through the Model Context Protocol.

Your AI tools each keep their own memory of you, siloed and invisible. big-brain inverts that: one knowledge vault you own, that every assistant reads and writes. Claude Code, Claude Desktop, ChatGPT, Cursor — they all see the same projects, the same tasks, the same notes. Switch models freely; your context comes with you.

• Plain markdown files. No database, no lock-in. Open the vault in Obsidian, grep it, put it in git.

• MCP server with 23 tools: search, capture, daily logs, project and task management, link graph, related notes, vault health.

• CLI for humans: big-brain status, big-brain capture, big-brain tasks.

• Claude Code skills — /brain, /capture, /weekly — installed with one command.

• Projects as the unit of work — each is one file with a goal, checkbox tasks, and a running log.

• Deterministic retrieval first: full-text search (fuzzy, title-boosted), [[wikilink]] graph with backlinks, tags, frontmatter queries. No API keys, works offline.

• Optional local hybrid search: flip embeddings.enabled and a small on-device model (via @huggingface/transformers) adds semantic matching, fused with full-text by Reciprocal Rank Fusion — paraphrases match, exact terms still count through the full-text half of the ranking, and nothing leaves your machine. Powers related_notes similarity too. See docs/vault-spec.md.

• Optional git auto-commit + push — flip a config flag and every note write, from any tool, commits only the note paths it touched and pushes automatically, so unrelated work is not swept in and your other machines stay in sync. Best-effort: a git failure never blocks a save (the note is saved, just not committed), and pushes run in the background so a slow network never delays one.

• Edits that don't damage notes. Writes are atomic and locked per note, keep each note's line endings, don't land inside code blocks or frontmatter, refuse to rewrite frontmatter they can't parse, and refuse to write anywhere that resolves outside the vault (such as through a symlink). See docs/vault-spec.md for the edge cases these don't cover.

Requires Node 20+.

Install

big-brain isn't on npm yet, so install from source (one-time):

git clone https://github.com/jzhao234/big-brain.git
cd big-brain
npm install
npm run build
npm link          # puts the CLI and both MCP server commands on your PATH

npm link skips -g install quirks and lets you git pull && npm run build to update later. (Once published, this becomes npm i -g big-brain.)

The default install keeps semantic-search native dependencies out of the base package. To enable local embeddings, install the opt-in runtime add-on alongside big-brain, then set embeddings.enabled to true in brain.config.json:

npm install --no-save --package-lock=false @huggingface/transformers@^3.8.1

For a future global npm install, install both packages into the same global prefix: npm i -g big-brain @huggingface/transformers@^3.8.1.

Related MCP server: ai-brain

Quickstart

big-brain init ~/brain --name "My Brain"   # scaffold a vault
cd ~/brain
big-brain status                           # look around
big-brain install-skills                   # add the /brain, /capture, /weekly skills

Then connect an AI tool (below) and, optionally, seed the vault from your existing AI history.

Connect your AI tools

For local clients, run big-brain-mcp; the vault is chosen by --vault <dir> or the BIG_BRAIN_VAULT environment variable.

Claude Code

claude mcp add --scope user big-brain -- big-brain-mcp --vault ~/brain
claude mcp list        # expect: big-brain ... ✔ Connected

Claude Desktop / Cursor / other MCP clients — add to the MCP config (e.g. claude_desktop_config.json). Use an absolute command path: GUI apps are launched without your shell's PATH, so a bare big-brain-mcp may not resolve (run which big-brain-mcp to get it).

{
  "mcpServers": {
    "big-brain": {
      "command": "/absolute/path/to/big-brain-mcp",
      "args": ["--vault", "/absolute/path/to/brain"]
    }
  }
}

Other local MCP clients — any client that speaks MCP over stdio works the same way. For clients without MCP, the CLI's --json output makes the vault scriptable.

Remote MCP preview — Big Brain can expose the same 23 tools through an authenticated Streamable HTTP endpoint:

export BIG_BRAIN_MCP_TOKEN="$(openssl rand -hex 32)"
big-brain mcp-http --vault ~/brain
# endpoint: http://127.0.0.1:3333/mcp

It binds to localhost by default and checks Authorization: Bearer <token> before parsing JSON. Bodies are capped at 1 MB (larger requests get HTTP 413 with a JSON-RPC error). Requests without Origin are allowed; requests with one require an exact match in --allowed-origins / BIG_BRAIN_MCP_ALLOWED_ORIGINS, which defaults to empty and refuses browser origins. Keep it behind an HTTPS reverse proxy or secure tunnel; never expose the plain HTTP listener directly. The static token is the first self-hosted transport milestone, not an OAuth implementation—browser connectors that require OAuth still need an OAuth-capable gateway. See docs/mcp-setup.md.

Browser fallback (claude.ai / ChatGPT) — the zero-infrastructure option is to connect the AI to your vault's GitHub repo so it can read the markdown directly, and write it where the connector allows (some GitHub connectors are read-only; you also lose the computed overview/search/task tools, since it's raw file access). See docs/browser-github-connector.md and paste prompts/browser-github-instructions.md.

Then teach the assistant how to use the vault: Claude Code reads the vault's CLAUDE.md automatically when you work inside the vault folder; elsewhere, and for other tools, paste prompts/agent-instructions.md into their custom instructions.

Using it: load and save

The whole point is that you never re-feed context. The assistant loads a small, relevant slice on demand and saves small pieces as you work — you just talk to it.

Load — orient at the start, pull details as needed:

Save — as decisions, progress, and ideas happen (append-first; never destructive):

Review — /weekly runs a guided pass: triage the inbox, prune projects, reschedule overdue tasks, fix broken links, and consolidate — big-brain doctor flags bloated notes and near-duplicates, and the review proposes splits/merges/stale-fact pruning for your approval. Capture and search keep a brain useful; consolidation keeps it trustworthy.

Skills

Three skills wrap the tools into commands, for Claude Code (/brain) and Codex ($brain):

• brain — load the overview and switch into brain-aware mode for the session

• capture — zero-friction capture to the inbox

• weekly — a guided weekly review

big-brain agents install claude codex installs them for both agents (see below). big-brain install-skills still copies the Claude versions into ~/.claude/skills (--force overwrites). They're path-agnostic — they call the big-brain CLI / MCP tools and resolve the vault from BIG_BRAIN_VAULT (or --vault), so the same skill works on every machine.

Agent profiles: your agent setup in the vault

Your coding agents' personal setup — skills, global instructions, hooks, the status line, a few settings — can live in the vault too, so every machine gets the same setup from one command and a change made on one machine reaches the others through git.

agents/claude/instructions.md   -> ~/.claude/CLAUDE.md
agents/claude/skills/<name>/    -> ~/.claude/skills/<name>   (one link per skill)
agents/claude/files/**          -> ~/.claude/**              (one link per file: statusline.sh, hooks/…)
agents/claude/settings.json     -> merged into ~/.claude/settings.json
agents/codex/instructions.md    -> ~/.codex/AGENTS.md
agents/codex/skills/<name>/     -> ~/.codex/skills/<name>
agents/codex/config.toml        -> merged into ~/.codex/config.toml

big-brain agents status                    # linked, missing, edited here, or only on this machine
big-brain agents install --dry-run         # show the plan
big-brain agents install                   # link everything (agents with a profile; or name them)
big-brain agents save claude statusline.sh --setting statusLine    # capture a local change
big-brain agents save codex skills/my-skill --setting tui.status_line

How it stays safe:

• Links, not copies. Edit a skill or script in the vault (or through the link) and every machine has it after git pull. Everything else in ~/.claude / ~/.codex — other skills, history, credentials — is never touched.

• Nothing unmanaged is replaced unless you say so. A file in the way is reported; --replace moves it to a timestamped backup under ~/.local/state/big-brain/backups (outside every agent folder, so a backed-up skill is never picked up) with a restore manifest, then links the vault's.

• Settings merge three ways. big-brain remembers what it applied on each machine: a value you edited locally is reported, not overwritten (--replace takes the vault's); a value you never touched follows the vault; a key dropped from the vault is removed only if it wasn't edited. Claude hooks are matched by event, matcher, and command, so changing a timeout updates the hook instead of registering it twice. TOML is edited line by line (comments, trust entries and hook hashes stay put), and any edit that would change more than the managed keys is refused.

• Plan first, then write, and roll back on failure. Every link and settings change is validated before the first write; a refusal leaves the machine untouched, and if a step still fails midway the steps already done are undone (if even that fails, nothing is deleted and the error says where the originals are). Nothing is written through a symlinked folder, and a settings file kept as a symlink (e.g. in a dotfiles repo) stays a symlink. --prune removes links whose vault file was deleted.

• save refuses known secret and state files. Whole settings files, .credentials.json, auth.json, .env files, history, sessions, caches and databases are refused; save individual settings with --setting. It doesn't inspect contents, so review the files and --setting values you save (an env.API_KEY setting would be copied as is).

• Codex hooks stay local. Codex asks you to trust each hook on each machine, so hook registrations are not managed for Codex. Claude Code users: review new hooks with /hooks. Status warns when Codex's AGENTS.override.md overrides AGENTS.md.

CLAUDE_CONFIG_DIR and CODEX_HOME are honored. The folder name is folders.agents in brain.config.json (default agents), and it is never indexed as notes.

Use it on multiple machines

The vault is a git repo, so this is just git: clone it on each machine, and each machine runs its own big-brain-mcp against its own clone. Git keeps them in sync.

• Write side — automatic. Set git.autoCommit (and git.autoPush) in brain.config.json and every note write — from any tool, any machine — is committed and pushed (best-effort). See docs/vault-spec.md.

• Read side — pull before you start. Add a SessionStart hook so a session always opens on the latest:

  // ~/.claude/settings.json
  "hooks": { "SessionStart": [ { "hooks": [ { "type": "command",
    "command": "git -C \"$HOME/brain\" pull --ff-only --quiet 2>/dev/null || true" } ] } ] }

  --ff-only means a pull never clobbers local work. (Or just git -C ~/brain pull by hand.)

On a fresh machine: install big-brain (above), git clone your vault, big-brain agents install, register the MCP server. Four commands and your notes, skills, instructions, hooks, and status lines match your other machines.

Seed it from your existing AI history

An empty brain is useless; the fastest way to a useful one is extracting what your LLMs already know about you. Every assistant you've used heavily — ChatGPT, Claude, Gemini — has months of context about your projects, preferences, and goals sitting in its memory and chat history. Seeding pulls that out, once, into files you own.

1. Run the seed interview in each LLM you use. Copy the prompt from prompts/seed-interview.md into each assistant — ideally the account with the most history, with memory/personalization enabled. It's a structured export interview: profile, current projects, stack, how you like to work with AI, goals, people, routines, standing constraints, open loops, and blind spots. It instructs the model to be concrete, mark inferences (inferred), and write "No data" instead of inventing — so the output is honest enough to build on.

2. Save each output into the vault's inbox/ as one file per model (inbox/seed-chatgpt.md, inbox/seed-claude.md, …). Don't organize anything yet — capture first is the whole inbox philosophy.

3. Triage. Ask a vault-connected assistant to run the process-inbox MCP prompt (or just say "triage my inbox" / run /weekly in Claude Code). It merges duplicates across models — different LLMs know different slices of you, and they mostly complement rather than conflict — then turns the material into real structure: projects with tasks for everything in flight, people/ notes, a preferences note your agent instructions can point at, reference notes for your stack/setup, and open loops as tasks. You confirm per item; raw exports can be kept in archive/ for provenance.

4. Repeat occasionally. Models keep accumulating context about you — re-running the interview every few months and triaging the diff keeps the vault ahead of any one provider's memory.

Two more seeding paths worth knowing: if you already keep an Obsidian vault, big-brain can adopt it in place (add a brain.config.json, map your folder names via the folders config — see docs/vault-spec.md); and a work session with a vault-connected agent seeds as a side effect — the agent instructions tell it to create projects and capture facts as you work, so the brain fills in from real activity even if you skip the interview.

The vault

brain/
├── BRAIN.md          # index & ground rules
├── CLAUDE.md         # agent instructions (picked up by Claude Code)
├── brain.config.json
├── inbox/            # capture now, organize later
├── daily/            # one note per day: focus, log, tasks
├── projects/         # one file per project: goal, tasks, log
├── areas/            # ongoing responsibilities
├── notes/            # evergreen knowledge, densely [[linked]]
├── people/           # one note per person
├── reference/        # external facts, docs, how-tos
├── archive/          # nothing is deleted, only archived
└── templates/        # daily/project/note/person templates

Notes are markdown + YAML frontmatter (type, tags, status, due…). Tasks are - [ ] checkboxes with Obsidian Tasks-style metadata: 📅 2026-03-01 due, ⏫ priority, ✅ done-date. Full details in docs/vault-spec.md.

MCP tools

Plus three MCP prompts: orient, weekly-review, process-inbox.

CLI

big-brain init [dir]            scaffold a vault        big-brain tasks [-p X] [-t tag] [--due-by D]
big-brain status                overview                big-brain task add|done ...
big-brain search <query>        search (hybrid if on)   big-brain projects [--status active]
big-brain show <note>           print a note            big-brain project new|status ...
big-brain new <title>           create a note           big-brain links <note>
big-brain related <note>        related notes + why     big-brain tags
big-brain capture <text>        quick capture           big-brain doctor
big-brain daily [--log "..."]   daily note / journal    big-brain index [--status|--rebuild]
big-brain append <note> <text>  append [--heading H]    big-brain frontmatter <note> k=v ...
big-brain archive <note>        archive (non-destructive delete)
big-brain agents status|install|save   agent setup from the vault (see Agent profiles)
big-brain install-skills        add Claude Code skills  big-brain mcp   run the MCP server

search, related, tasks, projects, and doctor take --json for scripting. task done and task update take a task id or its text: a task's whole text wins over longer tasks that contain it, and an ambiguous fragment lists the candidates. Put -- before append text that starts with - (a list item), or it is read as an option. frontmatter reads each value as YAML, like key: value in the file: 'tags=[work, llm]' is a list, due=2026-11-01 stays a date string, and key=null (or key=) removes the key. Quote any assignment that contains spaces, or the shell splits it into separate arguments.

Design principles

1. Files over databases. Markdown you can read in 30 years beats any app.

2. The human owns the vault; agents are guests. Agents append and capture freely, but rewriting and archiving are deliberate acts.

3. Deterministic beats clever. Search + links + tags retrieves reliably and works offline; the semantic layer is opt-in, fully local, and fused with — never a replacement for — exact search.

4. Model-agnostic by construction. Everything goes through MCP or the filesystem — nothing assumes a particular AI vendor.

Development

npm install
npm run typecheck && npm run lint && npm test
npm run build

MIT © Junhao Zhao