big-brain

# big-brain 🧠

[![npm](https://img.shields.io/npm/v/@jzhao234/big-brain)](https://www.npmjs.com/package/@jzhao234/big-brain)
[![CI](https://github.com/jzhao234/big-brain/actions/workflows/ci.yml/badge.svg)](https://github.com/jzhao234/big-brain/actions/workflows/ci.yml)
[![License: MIT](https://img.shields.io/badge/license-MIT-blue)](LICENSE)

**One memory for every AI assistant you use: your notes, projects, and tasks, kept as plain markdown you own.**

Each AI tool keeps its own private memory of you, and none of them share it. big-brain flips that around. Point it at a folder of markdown (a new vault, or the Obsidian vault you already have) and Claude Code, Claude Desktop, Cursor, Codex, or any other [MCP](https://modelcontextprotocol.io) client can search it, read it, capture into it, and keep your projects and tasks current. Switch models whenever you like; your context comes with you.

<p align="center"><img src="docs/demo-ai.svg" alt="Claude Code using big-brain: it reads the overview, captures a note to the inbox, and marks a task done" width="780"></p>
<p align="center"><sub>Replay of a real Claude Code conversation with a demo vault (<a href="docs/demo-ai.json">transcript</a>, recorded by <a href="scripts/demo-ai.mjs">scripts/demo-ai.mjs</a>); timing is illustrative.</sub></p>

## 60-second start

```bash
npx @jzhao234/big-brain setup
```

That one command installs big-brain, creates a vault at `~/brain`, connects Claude Code and Codex if you have them (MCP server plus the `/brain`, `/capture`, and `/weekly` skills), and prints the config to paste into Claude Desktop or Cursor. Pass a folder to use a different one, such as your Obsidian vault: `npx @jzhao234/big-brain setup ~/Obsidian`. It's safe to run again (anything already set up is left alone), and `--dry-run` shows the plan without changing anything.

Then restart your assistant and ask *"what am I working on?"* or *"capture: call the dentist on Friday"*. Other clients, and browser chats like claude.ai and ChatGPT, are covered in [Connect your AI tools](#connect-your-ai-tools).

Prefer to do it by hand? These three lines are what `setup` does for Claude Code:

```bash
npm install -g @jzhao234/big-brain
big-brain init ~/brain
claude mcp add --scope user big-brain -- big-brain-mcp --vault ~/brain
```

**Already use Obsidian?** big-brain reads an existing vault as is: no config file, no migration, and nothing is written until you ask for a change (`setup` writes nothing into it either). Add a `brain.config.json` later to map your folder names, one entry per folder (for example, `"folders": {"projects": "Projects", "inbox": "Inbox"}`), so projects and captures land where you keep them. See [docs/vault-spec.md](docs/vault-spec.md).

## What you get

- **Plain markdown files.** No database and no lock-in. Open the vault in Obsidian, grep it, or put it in git.
- **An MCP server with 23 tools:** search, capture, daily logs, projects, tasks, the link graph, related notes, and a vault health check.
- **A CLI for humans:** `big-brain status`, `big-brain capture`, `big-brain tasks`, and more.
- **Skills for Claude Code and Codex** (`/brain`, `/capture`, `/weekly`), installed with one command.
- **Projects as the unit of work.** Each project is one file with a goal, checkbox tasks, and a running log.
- **Deterministic retrieval first:** fuzzy full-text search with title boosting, `[[wikilink]]` backlinks, tags, and frontmatter filters. No API keys, and it works offline.
- **Optional local semantic search.** Turn on `embeddings.enabled` and a small on-device model (via `@huggingface/transformers`) adds meaning-based matches, fused with full-text search by Reciprocal Rank Fusion. Paraphrases match, exact terms still count, and nothing leaves your machine. It also powers `related_notes`. See [docs/vault-spec.md](docs/vault-spec.md#search--embeddings-hybrid-retrieval).
- **Optional git auto-commit and push.** Turn on `git.autoCommit` and every note write from any tool commits only the files it touched; add `git.autoPush` to push in the background so your other machines stay in sync. A git failure never blocks a save: if the commit fails the note is still saved, and if the push fails the commit stays local until the next push.
- **Edits that don't damage notes.** Writes are atomic and locked per note, keep each note's line endings, never 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). [docs/vault-spec.md](docs/vault-spec.md#known-limits) lists the edge cases these don't cover.

Requires Node 20 or newer. big-brain is pre-1.0: it's used daily, but minor versions may still change behavior; see the [changelog](CHANGELOG.md).

## Install

```bash
npm install -g @jzhao234/big-brain
```

That puts three commands on your PATH: `big-brain` (the CLI), `big-brain-mcp` (the MCP server over stdio), and `big-brain-mcp-http` (the remote preview below). To try it without installing, `npx @jzhao234/big-brain status --vault ~/brain` works too.

The base install leaves out semantic search and its native dependencies. To enable local embeddings, install the add-on into the same place, then set `embeddings.enabled` to `true` in `brain.config.json`:

```bash
npm install -g @jzhao234/big-brain @huggingface/transformers@^3.8.1
```

To work on big-brain itself, install from source instead: `git clone https://github.com/jzhao234/big-brain.git && cd big-brain && npm install && npm run build && npm link`. See [CONTRIBUTING.md](CONTRIBUTING.md).

## Quickstart

```bash
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
```

<p align="center"><img src="docs/demo.svg" alt="Terminal demo: create a vault, capture a thought, start a project, add tasks, and print the overview" width="780"></p>

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**

```bash
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).

```json
{
  "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:

```bash
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, so browser connectors that require OAuth still need an OAuth-capable gateway. See [docs/mcp-setup.md](docs/mcp-setup.md#remote-streamable-http-preview).

**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](docs/browser-github-connector.md) and paste [prompts/browser-github-instructions.md](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`](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:

| You want to… | In Claude Code | CLI |
| --- | --- | --- |
| "What am I working on?" | `/brain` | `big-brain status` |
| Find notes about X | *"search my brain for X"* | `big-brain search X` |
| Read one note | *"read the Auth project"* | `big-brain show "Auth"` |
| See tasks / projects | *"what's due?"* | `big-brain tasks`, `big-brain projects` |
| Pull surrounding context | *"what's related to X?"* | `big-brain related X` |

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

| You want to… | In Claude Code | CLI |
| --- | --- | --- |
| Stash a stray thought | `/capture <text>` | `big-brain capture "…"` |
| Log a decision / progress | *"log that I shipped X"* | `big-brain daily --log "…"` |
| Add a task | *"add a task to project Y"* | `big-brain task add "…" --note Y` |
| Finish a task | *"mark that done"* | `big-brain task done <id>` |
| Start a project | *"new project: …"* | `big-brain project new "…"` |
| Add to a