Compass MCP
A shared operational layer that synchronizes tasks and project contexts across Claude Chat, Cowork, and Code surfaces using local markdown files. It enables persistent state management and cross-session knowledge sharing without the need for a database.
README
🧭 Compass MCP
The missing bridge between Claude Chat, Cowork, and Code.
Compass is an MCP server that gives Claude shared operational state across all three surfaces. Add a task in Chat, check it off in Code, filter by project in Cowork — same data, zero re-explaining.
┌───────────┐ ┌───────────┐ ┌───────────┐
│ Chat │ │ Cowork │ │ Code │
└─────┬─────┘ └─────┬─────┘ └─────┬─────┘
│ │ │
└────────┬───────┴────────┬───────┘
│ │
┌───────▼────────────────▼───────┐
│ Compass MCP │
│ │
│ ┌──────────────────────┐ │
│ │ ~/compass-data/ │ │
│ │ tasks.md │ │
│ │ contexts/ │ │
│ └──────────────────────┘ │
└────────────────────────────────┘
Why?
Claude Chat, Cowork, and Code don't share state. Each session starts from zero. There's no way to ask in Chat "what did I finish in Code?" or tell Cowork "here's the context for this project" without re-explaining everything.
Compass fixes that. One MCP server, visible to all three surfaces, backed by plain markdown files.
This is not another todo app. It's the operational layer that connects how you think (Chat), how you execute (Cowork), and how you build (Code).
Install
git clone https://github.com/richlira/compass-mcp.git
cd compass-mcp
npm install
npm run build
Connect to Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"compass": {
"command": "node",
"args": ["/absolute/path/to/compass-mcp/build/index.js"]
}
}
}
Restart Claude Desktop. All three tabs (Chat, Cowork, Code) will see Compass.
Usage
No commands to memorize. Just talk:
| You say | Compass does |
|---|---|
| "Set up my workspace" | init_workspace → creates ~/compass-data/ |
| "Add task: write copy for landing page, deadline April 5" | add_task → adds to tasks.md |
| "I finished the workshop slides" | complete_task → marks it done |
| "What's still pending?" | get_tasks → returns filtered tasks |
| "Save context for Impact Lab: April 18, hackathon format" | save_context → creates contexts/impact-lab.md |
| "What do we know about the workshop?" | get_context → reads contexts/workshop.md |
Works in English, Spanish, Spanglish — whatever. Claude matches your intent to the right tool automatically.
Tools
init_workspace
Creates ~/compass-data/ with tasks.md and contexts/ directory. Safe to run multiple times.
add_task
Adds a task to tasks.md.
- title (required) — what needs to be done
- tags — project or category tags (e.g.
["cancun", "marketing"]) - deadline — due date
- section —
"active"(default) or"backlog"
complete_task
Marks a task as completed in tasks.md.
- title (required) — task to complete (fuzzy matched)
- notes — optional completion notes
get_tasks
Returns tasks filtered by status, tags, or deadline.
- status —
"active","backlog","completed", or"all" - tags — filter by tags
- deadline_before — show tasks due before a date
save_context
Creates or updates a project context file in contexts/.
- project (required) — project name (becomes the filename)
- content (required) — markdown content with project details
get_context
Reads a project context file.
- project (required) — project name to look up
Storage
Everything lives in ~/compass-data/:
~/compass-data/
├── tasks.md ← all tasks with status, tags, deadlines
└── contexts/
├── cancun.md ← project-specific context
├── workshop.md
└── ...
Plain markdown. Human-readable. Git-versionable. No database.
The workspace path is configurable via the COMPASS_DATA_DIR environment variable.
How the three surfaces use Compass
Chat — your thinking space. Plan the day, add tasks, check status, save project context. Chat is where you decide.
Cowork — your execution space. Cowork reads the same tasks and contexts. It can check what's pending, mark things done as it works, and read project context to understand what to do.
Code — your building space. Code checks for build-related tasks, marks them complete when done, and reads project context for specs and decisions. Code is also where you maintain and improve Compass itself.
All three read and write the same tasks.md and contexts/ — that's the bridge.
Pro tip: combine with Google Calendar
If you have the Google Calendar connector enabled, Claude can combine Compass tasks with your calendar:
"Check my tasks and my calendar for today, and block time for what I can get done"
Claude reads your tasks from Compass, finds free slots in your calendar, estimates duration, and creates focus blocks. No extra code needed — just Claude orchestrating two tools together.
Requirements
- Node.js 18+
- Claude Desktop (macOS or Windows)
- Claude Pro, Max, Team, or Enterprise plan
Development
# Build
npm run build
# Test with MCP Inspector
npx @modelcontextprotocol/inspector node build/index.js
# Watch mode
npm run watch
Philosophy
Compass is intentionally minimal. Six tools, two file types, zero database. Decisions, progress tracking, and daily recaps are left to Claude's memory — that's what it's good at. Compass only handles what Claude can't do alone: queryable task state and structured project context that persists across all surfaces.
If you need more, add it. If you don't, enjoy the simplicity.
License
MIT
Chat, Cowork, Code — and Compass. 🧭
推荐服务器
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。
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。
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。
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。
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。
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。
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。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
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。