bg-tasks-mcp
MCP server that enables AI agents to start, monitor, and stop background processes like dev servers, builds, and tests, with a web dashboard.
README
bg-tasks-mcp
Let your AI agent manage background processes. One sentence to start, monitor, and stop dev servers, builds, and tests — without leaving the conversation.
Why
Working on a full-stack project means juggling multiple terminal windows to start dev servers, check logs, and kill processes. Every context switch breaks your flow. You're the process manager, not the AI.
bg-tasks flips this: the LLM manages background processes for you.
Before:
┌──────────┐ ┌──────────────┐ ┌──────────────┐
│ opencode │ │ Terminal #1 │ │ Terminal #2 │
│ write code│ │ npm run dev │ │ go run . │
│ ask Q │ │ check logs │ │ check logs │
└──────────┘ └──────────────┘ └──────────────┘
↑ Alt+Tab, manual PID hunting, context switching
After:
┌──────────────────────────────────────────────────┐
│ opencode │
│ │
│ "Start frontend and backend with bg-tasks" │
│ → bg_start("frontend", "npm run dev") │
│ → bg_start("backend", "go run .") │
│ │
│ "Show me frontend logs" │
│ → bg_logs(id="frontend", tail=20) │
│ │
│ "Stop the backend" │
│ → bg_kill(id="backend") │
│ │
│ Dashboard: http://127.0.0.1:9876 │
└──────────────────────────────────────────────────┘
Use Cases
| Scenario | Example |
|---|---|
| Dev servers | bg_start("fe", "npm run dev") + bg_start("be", "python app.py") |
| Builds | bg_start("build", "npm run build") → bg_wait("build") → bg_logs("build") |
| Tests | bg_start("test", "npm test") → bg_wait("test") |
| Long scripts | bg_start("seed", "node seed-db.js") — check progress with bg_logs |
| Multi-project | Open two opencode windows → dashboard shows both projects' tasks |
Installation
npm (recommended)
{"command": ["npx", "-y", "bg-tasks-mcp"]}
No install needed — npx downloads and runs on first use.
Manual install
npm install -g bg-tasks-mcp
Then configure:
{"command": ["bg-tasks-mcp"]}
Git clone
git clone https://github.com/rookie136/bg-tasks-mcp.git
cd bg-tasks-mcp
npm install
Then:
{"command": ["node", "/path/to/bg-tasks-mcp/index.js"]}
When you need to start, monitor, or stop background processes (dev servers, builds, tests),
use the bg-tasks tools.
Usage
6 MCP Tools
| Tool | Purpose | Example |
|---|---|---|
bg_start |
Start a background task | bg_start(name="frontend", command="npm run dev") |
bg_wait |
Wait for completion or timeout | bg_wait(id="frontend", timeout=30) |
bg_logs |
Read task output | bg_logs(id="frontend", tail=20, stream="stdout") |
bg_status |
Inspect or list tasks | bg_status(id="frontend") / bg_status() |
bg_send |
Send text or signal | bg_send(id="frontend", input="<Enter>") |
bg_kill |
Terminate a task | bg_kill(id="frontend") / bg_kill(id="frontend", force=true) |
5 MCP Resources
| Resource | Description |
|---|---|
bg://status |
All tasks as JSON |
bg://{id}/status |
Single task detail + log summary |
bg://{id}/logs |
Combined stdout/stderr |
bg://{id}/logs/stdout |
stdout only |
bg://{id}/logs/stderr |
stderr only |
Web Dashboard
When the first opencode window starts, the HTTP dashboard comes online automatically:
http://127.0.0.1:9876 English (default)
http://127.0.0.1:9876/?lang=zh 中文
Features:
- Window list (owner PID, remote PIDs, task counts, last active time)
- Real-time running tasks view
- Clickable log links for remote window tasks
- Per-window history with clear buttons
- Bilingual UI (EN/中文 toggle)
Architecture
Multi-Window Design
┌─ opencode window #1 (owner) ──────────────────────────────────────┐
│ bg-tasks process │
│ │
│ serveStdio (MCP stdio) ←─ LLM tool calls │
│ HTTP :9876 │
│ GET / → dashboard HTML (bilingual) │
│ GET /status → JSON API │
│ GET /logs/{id} → log page │
│ POST /api/report → receive remote window tasks │
│ GET /api/health → liveness probe │
│ │
│ orphan check (30s) → parent alive? → 3 fails → self-terminate │
└─────────────────────────────────────────────────────────────────────┘
┌─ opencode window #2 (non-owner) ───────────────────────────────────┐
│ bg-tasks process │
│ │
│ serveStdio (MCP stdio) ←─ own LLM tool calls │
│ mini HTTP :9877 │
│ GET /logs/{id} → log page (read-only) │
│ │
│ reporter → every 10s POST /api/report (:9876) push task state │
│ health check → every 5s GET /api/health (:9876) → owner alive? │
│ → takeover :9876 → become new owner → close mini HTTP │
└─────────────────────────────────────────────────────────────────────┘
Owner/Takeover Lifecycle
Window #1 starts → binds :9876 → owner → dashboard online
Window #2 starts → EADDRINUSE → non-owner → mini HTTP :9877 → reporter push
Window #3 starts → EADDRINUSE → non-owner → mini HTTP :9878 → reporter push
Window #1 closes → health check fails → Window #2 jitter 500-1500ms → binds :9876
→ new owner → closes mini HTTP :9877 → tasks become local
Window #2 closes → Window #3 takes over
All windows closed → orphan check triggers → self-terminate
Data Flow
non-owner window owner window
───────────────── ─────────────
bg_start("server", ...) ──┐
bg_start("client", ...) ──┤
│
reporter (10s interval) ◄────┘
│
│ POST /api/report { pid, port, tasks: [...] }
▼
owner dashboard ──► aggregates remote + local tasks
│
├─ GET / → dashboard HTML (window list, running, history)
├─ GET /logs/{id} → local task logs (from MemoryLogStore)
└─ remote task links → http://127.0.0.1:{port}/logs/{id} (mini HTTP)
Logging System
JSONL format written to <install-dir>/.log:
{"ts":"20260809T023839","evt":"owner_ok","pid":30508,"port":9876,"ppid":11972}
{"ts":"20260809T023900","evt":"bg_start","pid":30508,"id":"d26c6305","name":"backend","cmd":"python..."}
{"ts":"20260809T024200","evt":"bg_exit","pid":30508,"id":"d26c6305","exit":0,"ms":60000}
{"ts":"20260809T024503","evt":"orphan_check","pid":30508,"attempt":1,"alive":false}
{"ts":"20260809T030000","evt":"shutdown","pid":30508,"reason":"SIGTERM"}
Auto-rotation: 1MB per file, keeps last 2 backups (.log.0, .log.1).
i18n
Translations are JSON files in lib/i18n/. To add a new language, copy en.json to <lang>.json and translate the values.
lib/i18n/
├── en.json # English (default, fallback)
└── zh.json # 中文
File Structure
bg-tasks-mcp/
├── index.js # entry: MCP stdio + owner detection + orphan check
├── lib/
│ ├── store.js MemoryLogStore (ring buffer, 4MB per key)
│ ├── registry.js TaskRegistry (state machine, onTaskEvent hook)
│ ├── platform.js Windows/Unix adaptation (shell, kill, signal)
│ ├── session.js session ID generation
│ ├── status-http.js HTTP dashboard + routes + multi-window aggregation
│ ├── reporter.js non-owner task state push (10s interval)
│ ├── logger.js JSONL file logger (1MB rotation)
│ ├── i18n.js i18n loader (in-memory cache, en fallback)
│ └── i18n/
│ ├── en.json English translations
│ └── zh.json 中文翻译
├── tools/ 6 MCP tools (bg_start, bg_wait, bg_logs, etc.)
├── resources/ MCP resources (bg://status, bg://{id}/logs, etc.)
└── test/ Node.js native test runner
Environment Variables
| Variable | Description | Default |
|---|---|---|
BG_TASKS_NO_ORPHAN_CHECK |
Set to 1 to disable orphan detection |
off |
Platform
Node.js 18+, Windows/Linux/macOS. Windows uses taskkill instead of POSIX signals.
License
MIT
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。