AgentHub
A local MCP server that orchestrates multi-agent coding teams by spawning workers, delegating tasks, and managing session lifecycles through an external MCP client.
README
AgentHub
AgentHub is a local MCP agent-team runtime for coding projects. An external orchestrator (Cursor, Codex, Claude, a custom MCP client, etc.) spawns named workers, delegates scoped tasks, waits for results, and resumes sessions. The human watches a read-only board; they do not drive the team from provider tabs.
North star: docs/MUST-DO-DIRECTION.md
What you get
| Piece | Role |
|---|---|
MCP server (agenthub mcp) |
Primary product contract — session lifecycle, workers, delegation, updates |
| Daemon + HTTP | Same runtime for TUI board and automation |
Read-only TUI (agenthub launch) |
Dynamic worker board; legacy operator UI behind /debug on |
.agenthub/ |
Durable mailbox, sessions, team state on disk |
Implementation: agenthub-daemon (TypeScript) + agenthub-tui (Rust).
Provider support (V1)
| Provider | Level | Resume in board |
|---|---|---|
| Codex | supported |
Yes — adapter-tested codex exec / exec resume, CLI 0.128.0 |
| Cursor | supported |
Yes — adapter-tested headless cursor-agent, stream-json |
| Grok | experimental |
No — use for scoped tasks; live two-turn resume not proven |
| OpenCode | experimental |
No — use for cost/model routing; permission server not wired |
Details: docs/PROVIDER-SUPPORT-MATRIX.md
Per-provider CLI evidence and orchestrator notes: docs/PROVIDER-KNOWLEDGE.md
Experimental providers: Grok and OpenCode may spawn and delegate, but AgentHub does not claim durable resume (runtimeCanResume: false). Prefer Codex or Cursor when restart-safe work matters.
V1 release status
Core V1 is release-ready when the local release gates are green: build, daemon tests, TUI tests, package validation, packed install smoke, and the final dogfood transcript.
- Current status: docs/STATUS.md
- Final dogfood: docs/V1-DOGFOOD-2026-06-06.md
- Release notes: docs/RELEASE-NOTES-V1.md
Install
Requires Node.js 20+. No Rust/MSVC at install time.
Installer channel (irm / curl)
The live stoyanco.com channel currently publishes the Windows x64 artifact.
irm https://stoyanco.com/agenthub/install.ps1 | iex
macOS/Linux use the same channel after their platform artifacts are built and uploaded:
curl -fsSL https://stoyanco.com/agenthub/install.sh | bash
Maintainer build/upload flow: docs/INSTALLER-CHANNEL.md
npm package
npm install -g @agenthub-mcp/cli
agenthub doctor
agenthub init .
agenthub mcp
Connect your orchestrator to agenthub mcp (stdio MCP). Typical flow:
create_session → get_session_state → spawn_worker → delegate_to_worker
→ wait_for / get_updates → recover_session (after daemon restart)
| Target name | MCP tool |
|---|---|
create_session |
agenthub_create_session |
activate_session |
agenthub_activate_session |
get_session_state |
agenthub_get_session_state |
spawn_worker |
agenthub_spawn_worker |
delegate_task |
agenthub_delegate_to_worker |
wait_for |
agenthub_wait_for |
get_updates |
agenthub_get_updates |
recover_session |
agenthub_recover_session |
stop_worker / archive_worker |
agenthub_stop_worker / agenthub_archive_worker |
Also: agenthub_list_sessions, agenthub_get_runtime_manifest (includes providerCapabilities + knowledge doc paths).
Full orchestrator guide: docs/AGENTHUB-SKILL.md
Read-only board
agenthub launch .
Starts the daemon and TUI board (worker tabs, ops queue, settings). Default UX is read-only observation — orchestration stays in MCP.
- Requires a prebuilt TUI binary when installed from npm (
vendor/tui/<platform>-<arch>/). - Legacy / debug (not V1 product path): provider-first tabs (
/debug on), Home slash orchestration, F2/F3 writer lock, conductor MCP tools. See docs/LEGACY-RETIREMENT.md.
Dev checkout
git clone <repo-url>
cd agenthub
npm install
npm run build
agenthub doctor
agenthub mcp
Cursor MCP (dev checkout)
This repo ships .cursor/mcp.json so Cursor loads AgentHub automatically for this workspace (bin/agenthub.js mcp with AGENTHUB_PROJECT=${workspaceFolder}). After npm run build, reload MCP servers in Cursor settings.
Smoke test (mock adapters, no provider credits):
npm run smoke:mcp
Windows TUI builds need MSVC (npm run build:tui). See docs/RELEASE-CHECKLIST.md.
HTTP API (optional): agenthub daemon on port 4367.
Docs map
| Doc | Purpose |
|---|---|
| STATUS.md | Current snapshot, tests, V1 verdict |
| PROVIDER-SUPPORT-MATRIX.md | Machine-readable support contract |
| PROVIDER-KNOWLEDGE.md | CLI evidence, models, permissions, risks |
| AGENTHUB-SKILL.md | MCP orchestrator workflow |
| CLI-FLAGS.md | Provider CLI flags |
| V1-MAILBOX.md | File-backed mailbox protocol |
| RELEASE-CHECKLIST.md | Build/test gates |
Adapter overrides (optional)
$env:AGENTHUB_CODEX_ARGS="-m gpt-5.4-mini"
$env:AGENTHUB_CURSOR_ARGS="--model composer-2.5"
$env:AGENTHUB_GROK_ARGS="-m grok-build-0.1"
$env:AGENTHUB_OPENCODE_ARGS="-m opencode-go/kimi-k2.6"
See docs/CLI-FLAGS.md and provider knowledge files for launch templates and resume behavior.
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。