TaskFabric
A file-based task manager operated through an MCP server, enabling agents to create, search, and manage tasks with git versioning and hybrid semantic+keyword search.
README
TaskFabric
A file-based task manager operated entirely through an MCP server over Streamable HTTP. Every task is a markdown file with YAML frontmatter, indexed via QMD for hybrid semantic + keyword search, and git-synced for versioning. No UI — agents are the only interface.
How it works
Tasks live as plain markdown files organized by status:
/tasks/
├── inbox/ # New, untriaged tasks
├── active/ # Currently being worked on
├── waiting/ # Blocked or waiting
├── done/2026-03/ # Completed (monthly subdirs)
└── archived/2026-03/ # Old tasks (monthly subdirs)
Each task has YAML frontmatter (id, title, status, priority, tags, due, assignee, completed_at, waiting_on, dependencies) and a markdown body with free-form content and a ## Log section for timestamped entries. Filenames include the task ID to prevent collisions on duplicate titles.
Every mutation auto-commits to git and pushes to a remote, so all changes are versioned and recoverable.
MCP Tools
| Category | Tools |
|---|---|
| CRUD | task_create, task_get, task_update, task_delete, task_list |
| Search | task_search (keyword/semantic/hybrid + filters), task_expand_query, task_structured_search |
| Workflow | task_move, task_log, task_link, task_batch |
| Views | task_dashboard, task_timeline, task_graph, task_summary, task_recent, task_completion_report |
| Maintenance | task_auto_archive, task_reindex |
| Sync | sync_status, sync_pull, sync_history, sync_diff, sync_restore |
| Settings | settings_get, settings_update |
Setup
Environment Variables
| Variable | Required | Description |
|---|---|---|
TASKS_DIR |
Yes | Path to the tasks directory |
API_KEY |
Yes | Bearer token for MCP authentication |
GIT_USER_NAME |
Yes | Git commit author name |
GIT_USER_EMAIL |
Yes | Git commit author email |
TASKS_REPO_URL |
No | Git remote URL (clones on first start) |
GIT_TOKEN |
No | GitHub PAT for private repos (fine-grained PAT with Contents read/write) |
PORT |
No | Server port (default: 8181) |
Run locally
bun install
TASKS_DIR=./tasks API_KEY=your-secret GIT_USER_NAME="Your Name" GIT_USER_EMAIL="you@example.com" bun run src/server.ts
Run with Docker
Create a .env file:
API_KEY=your-secret
GIT_USER_NAME=Your Name
GIT_USER_EMAIL=you@example.com
TASKS_REPO_URL=https://github.com/you/your-tasks-repo.git
GIT_TOKEN=ghp_your_github_pat
docker compose up --build
The server starts at http://localhost:8181 with:
/mcp— MCP endpoint (requiresAuthorization: Bearer <API_KEY>)/health— Health check (returns{ "status": "ready" })
Connect MCP Clients
Claude Code (CLI)
claude mcp add --transport http task-fabric http://localhost:8181/mcp \
--header "Authorization: Bearer your-secret"
Or add to .mcp.json in your project root:
{
"mcpServers": {
"task-fabric": {
"type": "http",
"url": "http://localhost:8181/mcp",
"headers": {
"Authorization": "Bearer your-secret"
}
}
}
}
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS):
{
"mcpServers": {
"task-fabric": {
"command": "npx",
"args": [
"mcp-remote",
"http://localhost:8181/mcp",
"--header",
"Authorization: Bearer your-secret"
]
}
}
}
Cursor
Add to ~/.cursor/mcp.json:
{
"mcpServers": {
"task-fabric": {
"type": "streamable-http",
"url": "http://localhost:8181/mcp",
"headers": {
"Authorization": "Bearer your-secret"
}
}
}
}
OpenAI Codex CLI
Add to ~/.codex/config.toml:
[mcp_servers.task-fabric]
command = "http://localhost:8181/mcp"
http_headers = { "Authorization" = "Bearer your-secret" }
Or use an environment variable for the token:
[mcp_servers.task-fabric]
command = "http://localhost:8181/mcp"
bearer_token_env_var = "TASK_FABRIC_API_KEY"
Tests
bun test # All tests (159)
bun test src/__tests__/e2e/ # E2E tests (MCP protocol + HTTP)
bun test src/__tests__/tools/ # Integration tests
Stack
- Runtime: Bun
- MCP:
@modelcontextprotocol/sdk(Streamable HTTP) - Search:
@tobilu/qmd(BM25 + vector + LLM re-ranking) - Git:
simple-git - Frontmatter:
gray-matter - Validation: Zod
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。