impulse-labs-sas/brainstack

人类和AI助手共享第二大脑。自我托管。MCP原生。

BrainStack 在 Postgres 中以 Markdown 格式保存你的笔记。它会将笔记暴露给任何兼容 MCP 的 AI 助手：Claude Code、Claude Desktop 和 claude.ai、Cursor、Codex、Gemini CLI、Continue、Cline 等。人们使用这个网页应用，涵盖了 Obsidian 的功能：文件树、带维基链接的 Markdown 编辑器、全文搜索、图表，以及导入你已有的 .md 文件。助手通过 MCP 服务器阅读和写出相同的笔记，可以在一次调用中收集所有问题所涉及的内容;图表的哨兵视图会重放他们所读的内容，并将你自己的问题和所需笔记转化为提示，供你粘贴到助手或集成开发环境（IDE）中。

每个账户都有自己的保险库，任何文件夹都可以与他人共享，无论是只读还是写权限。

状态： 预备测试。由Impulse Labs制造。

为什么

如果你用多个AI助手，你会不断在各处重打相同的上下文：仓库里的设计文档、Claude Projects里、Cursor Rules里、Gemini里、脑海里。BrainStack是唯一的真相来源。

Shared second brain for humans and AI assistants. Self-hostable. MCP-native.

BrainStack keeps your notes as Markdown in Postgres. It exposes them to any MCP-compatible AI assistant: Claude Code, Claude Desktop and claude.ai, Cursor, Codex, Gemini CLI, Continue, Cline and others. People use the web app, which covers what Obsidian did: a file tree, a Markdown editor with wikilinks, full-text search, a graph, and import of the .md files you already have. Assistants read and write the same notes through the MCP server, and can gather everything a question refers to in a single call; the graph’s Sentinel view replays what they read, and turns a question of your own into a prompt, with the notes it needs, to paste into an assistant or an IDE.

Every account has its own vault, and any folder can be shared with other people, read-only or with write access.

Status: pre-alpha. Built by Impulse Labs.

Why

If you work with several AI assistants, you keep retyping the same context everywhere: design docs in the repo, in Claude Projects, in Cursor rules, in Gemini, in your head. BrainStack is the single source of truth. Your assistants pull what they need through MCP and write back when you tell them to. One brain, every AI.

What’s inside

apps/
  server/     Hono + tRPC + MCP server (Node 24+), with its own OAuth 2.1 provider for MCP clients
  web/        Next.js 15 web app: react-aria-components + CodeMirror 6
packages/
  core/       Markdown parsing, wikilink resolution, Postgres store and full-text search
  skill/      Canonical AI instructions + adapters for every major client

Storage is any Postgres. Sign-in is email and password, with optional Google sign-in and optional TOTP. MCP clients connect through OAuth or with an API key.

Quick start (self-host)

You need Docker. The compose file runs Postgres, the server, the web app and Caddy, which serves them all on one origin with automatic TLS.

git clone https://github.com/Impulse-Labs-SAS/brainstack.git
cd brainstack
cp .env.example .env
# edit .env: at least DOMAIN, PUBLIC_ORIGIN and POSTGRES_PASSWORD
docker compose up -d

Then open your PUBLIC_ORIGIN and create your account right away. The first account on an instance is its owner and needs no email verification. After it, sign-up is closed.

The schema is created on first start and updated by every new version on its own. Already have a Postgres? Set DATABASE_URL in .env and BrainStack uses it instead of the bundled one.

Email: optional for one person, needed to add others