Open-Brain

Open-Brain

Personal knowledge graph with 16 MCP tools. Auto-links, deduplicates, tracks themes, synthesizes insights.

Category
访问服务器

README

Open Brain

Your personal knowledge infrastructure. Capture thoughts from anywhere, search by meaning.

Open Brain stores your thoughts with AI-generated embeddings so any AI assistant can search your memory by meaning -- not keywords. Thoughts flow in from Telegram, automated pipelines, or any MCP-compatible AI client (Claude, ChatGPT, and others) and are instantly searchable. A knowledge graph auto-links related thoughts, extracts entities, and tracks co-occurrence patterns. Automated maintenance keeps the graph healthy -- deduplicating near-duplicates, tracking theme evolution, synthesizing insights, and archiving stale content.

Everything lives in your own database, so you own your data.

How It Works

flowchart LR
    TG[Telegram Message] --> TGBot[telegram-bot\nEdge Function]
    MCP[AI Client\nClaude/ChatGPT] --> MCPServer[open-brain-mcp\nEdge Function]
    Pipeline[RSS/HF Papers/\nEmergent Mind] --> RunPipeline[run-pipeline\nEdge Function]
    TGBot --> OR1[OpenRouter\nEmbedding + Metadata]
    MCPServer --> OR2[OpenRouter\nEmbedding + Search]
    RunPipeline --> OR3[OpenRouter\nTriage + Embed]
    OR1 --> DB[(Postgres\n+ pgvector)]
    OR2 --> DB
    OR3 --> DB
    MCPServer --> DB
    TGBot --> TGReply[Telegram Reply\nwith Metadata]

Capture Path

When you send a message to the Telegram bot, the telegram-bot Edge Function picks it up via webhook. It sends the message to OpenRouter in parallel for two things: generating a vector embedding (a numerical representation of meaning) and extracting metadata like topics, people mentioned, action items, theme, quality score, and named entities. The thought is checked for semantic duplicates, stored in your database with auto-linked connections to related thoughts, and the bot replies with a summary of what it captured.

Pipeline Path

The run-pipeline Edge Function automatically ingests ideas from RSS feeds (AI newsletters), Hugging Face daily papers, and Emergent Mind (trending arXiv papers). Each item is triaged for relevance, embedded, deduplicated, and stored. Runs on a schedule via GitHub Actions (Supabase deployment) or a built-in cron container (Docker deployment).

Retrieval Path

Any AI client connected via MCP (Model Context Protocol) can search your thoughts by meaning using semantic search, browse by filters (type, topic, person, time), get aggregate statistics, or request a weekly review of themes. The open-brain-mcp Edge Function handles these requests, authenticated with your personal access key.

Knowledge Graph

Every thought is automatically linked to related thoughts via vector similarity. Connections above 0.80 similarity are classified by an LLM into typed relationships (extends, contradicts, is-evidence-for, supersedes). Named entities (people, tools, projects, organizations) are extracted and resolved into a shared entity graph. Co-occurrence edges track which thoughts are retrieved together over time, strengthening connections based on actual usage patterns.

Storage

Everything lives in Postgres with pgvector for fast similarity search. Thoughts are stored with their embeddings (1536-dimensional vectors), metadata, typed connections, and entity references. You can deploy on Supabase (managed hosting) or self-host with Docker Compose.

Deployment Options

Choose how you want to run Open Brain:

Supabase (hosted) Docker Compose (self-hosted)
Setup Link project + run scripts cp .env.example .env + ./start.sh
Infrastructure Managed by Supabase Runs on your machine/server
Scheduling GitHub Actions Built-in cron container
Cost Supabase free tier + OpenRouter Just OpenRouter
Guide Continue below Docker guide

Supabase Deployment

Prerequisites

  1. Supabase account -- Supabase is a hosted Postgres database with built-in APIs, authentication, and Edge Functions (serverless code). Create a free account at supabase.com. Create a new project -- you will need the project URL (looks like https://abcdef.supabase.co) and the service role key (a long string found under Settings > API).

  2. Supabase CLI -- The command-line tool for managing your Supabase project (applying database migrations, deploying functions, setting secrets).

    npm install -g supabase
    
  3. OpenRouter account -- OpenRouter routes requests to AI models. It is used here for generating embeddings (vector representations of your thoughts) and extracting metadata. Create an account at openrouter.ai and generate an API key from the dashboard.

  4. Telegram bot (recommended) -- The primary way to capture thoughts on the go. Create a bot via @BotFather on Telegram and run the setup script (see below). If you only want MCP access, you can skip this.

Quick Start

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/open_brain.git
cd open_brain

2. Link your Supabase project

cd supabase
supabase link --project-ref YOUR_PROJECT_REF
cd ..

Tip: Your project ref is the subdomain in your Supabase URL. If your URL is https://abcdef.supabase.co, your project ref is abcdef.

3. Run bootstrap

./scripts/bootstrap.sh

Bootstrap walks you through setting up your environment. It prompts for each secret (Supabase URL, service role key, OpenRouter API key, Telegram tokens, etc.), generates a cryptographic MCP access key automatically, and writes everything to .env.local. If you already have a .env.local, it will show your existing values and let you update specific ones.

4. Run deploy

./scripts/deploy.sh

Deploy applies the database schema (creates the thoughts table with vector search indexes), uploads your secrets to Supabase, and deploys all Edge Functions. It shows a step-by-step checklist as each operation completes. At the end, it prints your MCP connection URL and a ready-to-paste Claude Code command.

5. Run validate

./scripts/validate.sh

Validate runs 8 checks against your live deployment to confirm everything works: database access, RPC functions, Edge Function reachability, authentication, thought capture, semantic search, and thought listing. It prints a checklist with pass/fail for each check and a final summary.

Set up Telegram bot (optional)

Create a bot via @BotFather on Telegram, then run the setup script:

./scripts/setup-telegram.sh YOUR_BOT_TOKEN

The script verifies your token, registers the webhook, sets up command autocomplete, and prints the env vars and secrets to configure. Follow the printed instructions to complete setup.

Connect Your AI Client

Once deployed, connect your AI client to start using Open Brain. You need two values:

  • MCP endpoint URL: https://YOUR_REF.supabase.co/functions/v1/open-brain-mcp/mcp (Supabase) or http://localhost:80/functions/v1/open-brain-mcp (Docker)
  • MCP access key: The key generated during setup (stored in .env.local or Docker .env)

Tip: The deploy script (Supabase) or start script (Docker) prints the exact connection command with your values filled in.

Claude Code (CLI -- recommended)

claude mcp add --transport http --header "x-brain-key: YOUR_MCP_KEY" open-brain https://YOUR_REF.supabase.co/functions/v1/open-brain-mcp/mcp

This registers Open Brain as an MCP server that Claude Code can use in any conversation. Replace YOUR_MCP_KEY and YOUR_REF with your actual values.

Claude Code (project .mcp.json)

Add this to a .mcp.json file in your project root to share the connection with your team:

{
  "mcpServers": {
    "open-brain": {
      "type": "http",
      "url": "https://YOUR_REF.supabase.co/functions/v1/open-brain-mcp/mcp",
      "headers": {
        "x-brain-key": "${MCP_ACCESS_KEY}"
      }
    }
  }
}

Note: The ${MCP_ACCESS_KEY} syntax uses environment variable expansion so your key stays out of version control. Set the MCP_ACCESS_KEY environment variable on each machine that uses this config.

Claude Desktop

Claude Desktop does not support remote MCP servers via configuration files. Instead:

  1. Open Claude Desktop > Settings > Connectors
  2. Click Add a new connector
  3. Enter the MCP endpoint URL: https://YOUR_REF.supabase.co/functions/v1/open-brain-mcp/mcp
  4. Configure the x-brain-key authentication header with your MCP access key

ChatGPT (Pro/Team/Enterprise/Edu)

  1. Go to Settings > Connectors > Advanced > Developer Mode
  2. Add the MCP server URL: https://YOUR_REF.supabase.co/functions/v1/open-brain-mcp/mcp
  3. Configure the x-brain-key authentication header with your MCP access key

Usage Examples

Telegram Capture

Send any message to your bot and Open Brain automatically processes it:

You: Just had a great meeting with Sarah about the Q3 product roadmap.
     She wants to prioritize the mobile app redesign.

Bot: Captured!
  Type: meeting_note
  Theme: personal
  Topics: q3-roadmap, mobile-app-redesign
  Quality: 0.7
  People: Sarah
  Action items: Prioritize mobile app redesign
  Why: Records a product strategy decision with clear ownership
  Related: "Product planning session notes..." (82% similar)

Every message is embedded as a vector, enriched with extracted metadata, checked for duplicates, auto-linked to related thoughts, and entities are resolved into a knowledge graph.

Semantic Search

Ask any connected AI client to search your brain:

You: Search my brain for anything about product roadmap discussions

Claude: I found 3 relevant thoughts:
  1. (0.89 similarity) Meeting with Sarah about Q3 product roadmap...
  2. (0.82 similarity) Product planning session notes...
  3. (0.76 similarity) Quarterly goals discussion...

Semantic search finds thoughts by meaning -- even if you used different words. Asking about "product planning" will find thoughts about "roadmap discussions" because the meanings are similar.

Weekly Review

Get an AI-generated summary of your recent thinking:

You: Give me a weekly review of my recent thoughts

Claude: Here's your weekly review:
  Themes: Product planning, team meetings, technical architecture
  Open loops: Mobile redesign decision pending, API migration timeline
  Connections: Sarah mentioned in 3 meetings this week, all about mobile

The weekly review analyzes the past 7 days of thoughts and synthesizes themes, open loops, connections between ideas, and gaps in your thinking.

Available Tools

Tool Description
search_thoughts Semantic search with optional graph expansion (1-hop traversal)
list_thoughts Browse thoughts filtered by type, topic, person, theme, quality, time
thought_stats Aggregate statistics: counts, type/theme breakdown, top topics/people
capture_thought Save a new thought from any AI client (with auto-embedding)
get_connections Graph traversal from a thought (typed links: extends, contradicts, etc.)
list_entities Browse extracted entities (people, tools, projects, orgs) by frequency
weekly_review AI-generated summary of themes, open loops, and next steps
analyze Graph analysis: hubs, density, sources, co-occurrence, themes
dedup_review Duplicate candidates with similarity zone histogram
refresh_salience Recompute all salience scores
update_thought Rewrite content (re-embeds, re-extracts metadata)
delete_thought Permanent delete (cascades connections)
serendipity_digest Resurface forgotten high-quality thoughts
pipeline Pipeline monitoring: health status, run history, merge audit
review_stale Review and act on stale thought candidates
migration_guide Instructions for importing memories from other platforms

See docs/cookbook.md for detailed usage patterns, tool compositions, and non-obvious behaviors.

Skills (Claude Code Workflows)

Open Brain includes Claude Code skills -- structured multi-phase workflows that compose the MCP tools above into higher-level analysis. Skills are auto-discovered from .claude/skills/ and invoked as slash commands.

Skill What it does
/discover Incremental pattern discovery across recent thoughts. Builds on previous reports (EVOLVED/NEW/STALE classification), dispatches parallel research agents, correlates with project priorities.
/pulse Pipeline and data health report. 9 parallel MCP calls, rubric-scored (GREEN/YELLOW/RED), cross-run memory for tracking findings over time, 6 cross-metric pattern detectors.
/brain-health Knowledge graph health report. 12 parallel MCP calls covering theme attention, graph density, hub health, co-occurrence alignment, dedup pressure, synthesis output, and entity landscape.

See docs/skills/README.md for detailed descriptions and usage.

Automated Maintenance

Open Brain runs background maintenance to keep the knowledge graph healthy. These jobs run automatically -- via GitHub Actions (Supabase deployment) or the built-in cron container (Docker deployment).

Job Frequency Purpose
RSS/HF Papers/Emergent Mind ingestion 2x daily Ingest ideas from configured sources
Pipeline monitoring 2x daily Health checks with Telegram alerting on failures
Dream dedup 2x daily Merge near-duplicate thoughts (>0.92 similarity auto-merged, 0.85-0.92 LLM-confirmed)
Graph analysis cache Daily Pre-compute hub, density, and co-occurrence analysis
Dream themes Weekly Track theme velocity, lifecycle transitions (emerging/active/declining), centroid drift
Dream decay Weekly Archive stale thoughts via tiered scoring + LLM confirmation
Dream synthesis Weekly Generate cross-cutting insights from thought clusters
Co-occurrence decay Weekly Decay unused co-occurrence edges

GitHub Actions workflow files are included in docs/workflows/ as reference for customizing schedules.

Benchmark

Open Brain includes a LongMemEval benchmark harness that measures retrieval quality across 500 questions in 6 categories. Each question gets its own isolated brain -- conversations are ingested as thoughts, then questions test whether the right memories can be retrieved and answered.

Baseline results (threshold=0.4, limit=20, graph expansion on):

Category Score What it tests
single-session-assistant 69.6% Recall assistant responses from a single conversation
single-session-user 55.7% Recall user statements from a single conversation
knowledge-update 52.6% Surface the latest version of updated information
temporal-reasoning 30.8% Answer questions about when things happened
multi-session 18.8% Synthesize across multiple conversations
single-session-preference 3.3% Recall casually mentioned preferences
Overall 37.2%

Vector search handles direct factual recall well. Multi-session synthesis and temporal reasoning are the main areas for improvement -- they require retrieval strategies beyond single-query vector similarity.

Run the benchmark against your own instance:

pip install -r benchmark/requirements.txt
cp benchmark/.env.example benchmark/.env  # fill in keys
python -m benchmark longmemeval run --concurrency 5

Results land in benchmark/results/. See benchmark/longmemeval/config.py for tunable parameters (threshold, retrieval limit, models).

Project Structure

open-brain-server/
  supabase/
    migrations/                   # Database migrations (applied with supabase db push)
    functions/
      _shared/                    # Shared modules (supabase-client, openrouter, types, errors, auto-link, entities, dream-*)
      telegram-bot/               # Telegram capture (primary capture path)
      open-brain-mcp/             # MCP server (16 tools)
        tools/                    # Individual tool implementations
      run-pipeline/               # Automated RSS/HF Papers/Emergent Mind ingestion
      monitor-pipeline/           # Pipeline health monitoring with Telegram alerts
      refresh-graph-analysis/     # Graph analysis cache computation
  benchmark/                        # LongMemEval retrieval quality benchmark
    longmemeval/                    # Harness: retrieve → reader → judge → score
    results/                        # Baseline results (summary markdown)
    tests/                          # 60 tests
  docker/                         # Docker Compose self-hosting (6 services)
  pipeline/                       # Python-based local pipeline (Reddit, RSS, briefing)
  scripts/                        # Setup and deployment automation
  tests/                          # Integration tests
  docs/
    cookbook.md                    # MCP tool usage patterns and compositions
    skills/                       # Skill documentation
    workflows/                    # GitHub Actions reference (scheduling)
    writing-a-source.md           # Guide for adding pipeline sources
  .claude/
    skills/                       # Claude Code skills (auto-discovered)
      discover/                   # Incremental pattern discovery
      pulse/                      # Pipeline health report
      brain-health/               # Knowledge graph health report

License

MIT

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选