@vintcessun: 原来多个 AI agent 组队干活,比单个通用 agent 高明在这:每个角色绑自己最擅长的模型,记忆和技能跨聊天累积。不是轮流调用,而是通过 handover 传一句简报就把任务交出去。本地跑,文件状态全在 ~/.crew44 里,免…
摘要
Crew44 is a local-first orchestrator that turns coding agents like Claude Code and Codex into a coordinated team of specialists, each bound to its best model, with persistent memory and skill accumulation across sessions. It runs entirely on your machine with no cloud dependence and is free under MIT license.
查看缓存全文
缓存时间: 2026/06/09 08:50
原来多个 AI agent 组队干活,比单个通用 agent 高明在这:每个角色绑自己最擅长的模型,记忆和技能跨聊天累积。不是轮流调用,而是通过 handover 传一句简报就把任务交出去。本地跑,文件状态全在 ~/.crew44 里,免费 MIT。凌晨看这个实现路径,感觉团队协作的思路终于落到 agent 层面了。
getcrew44/crew44
Source: https://github.com/getcrew44/crew44
Crew44
A local-first orchestrator for running specialist AI agents on your own machine.
Crew44 turns the AI agents you already run — Claude Code, Codex, Gemini, Cursor, and more — into a coordinated team. Instead of one generalist agent re-explaining context to itself all day, you assemble specialists, bind each to the model that wins at its job, and let them hand work off to each other inside one shared workspace.
Everything runs on your machine. State is plain files under ~/.crew44/. No cloud account, no subscription, no telemetry — the only network traffic is whatever your underlying coding agent already makes.
https://github.com/user-attachments/assets/ce6e5293-6c58-4c37-8e00-74d654d2277c
Download
Grab a signed desktop build for your platform — free, no account required:
| Platform | Format |
|---|---|
| macOS (Apple Silicon) | .dmg |
| Windows (x64) | .exe installer |
| Linux (ARM64) | .AppImage / .deb |
Prefer to build it yourself? See Getting started.
Why Crew44
| The old way — one agent | With Crew44 — a crew of specialists |
|---|---|
| A new contractor every morning: never seen your repo, never learned your conventions. | Per-project memory. Monday’s fix is in the crew’s muscle memory by Thursday. |
| Skills never compound — whatever it figured out is gone by next task. | Skills that compound. Capture a workflow once as SKILL.md; every agent you attach it to gets it on every turn, across providers. |
| One generalist plans, builds, and reviews — no deep expertise in any role. | Specialists in parallel. Planner drafts while builder codes while reviewer checks. Handovers ship the baton, not the whole context. |
| Locked to a single model: pays Opus rates for a rename, runs a fast model on a hard call. | The right model per role. Opus plans, GPT-5.5 codes, a local model reviews — swap per task, not per app. |
Supported runtimes
Crew44 detects and routes to any of these installed on your machine:
Claude Code · Codex · Cursor Agent · Gemini CLI · Hermes · Kimi · OpenCode · OpenClaw · Pi · Qoder · Qwen Code
The same skills folder runs across every provider, so you’re never locked in. Have another CLI? The runtime layer is a small Go interface — add an adapter under daemon/internal/backendagent/ and it shows up in the picker.
Core concepts
| Concept | What it is |
|---|---|
| Runtime | A coding-agent CLI on disk (Claude, Codex, Cursor, …) discovered by scanning. |
| Agent | A named persona bound to one runtime + model, with an instruction and attached skills. |
| Skill | A file-based capability (SKILL.md + assets) injected into the runtime session when its agent runs. |
| Project | A working directory plus the chats that belong to it. Stored under Documents/Crew44/ or a folder you pick. |
| Chat | A turn-by-turn thread. One in-flight response at a time; events are an append-only events.jsonl. |
| Worktree | An optional isolated git checkout for a chat. Toggle it on a new task and the crew works on its own crew/… branch without touching your working tree. |
| Handover | A marker an agent emits to pass the turn to a teammate, with a one-line brief. |
The default crew ships with a Partner, an Engineer, a Product Lead, and a Designer — each owning a role, a model, and its own skills folder.
How it works
┌─────────────────┐ WebSocket JSON-RPC ┌──────────────────┐ spawn ┌─────────────────┐
│ Electron / UI │ ◄───────────────────► │ Go daemon │ ──────────► │ agent CLIs │
│ React 19 │ │ 127.0.0.1 │ │ (claude, codex,│
└─────────────────┘ └──────────────────┘ │ cursor, …) │
│ └─────────────────┘
▼
~/.crew44/ (plain-file state)
- Daemon — a single Go process at
daemon/. Owns runtime discovery, agent/skill/chat state, and the JSON-RPC + event-stream surface. Auth is a per-launch bearer token; only/healthis unauthenticated. - Renderer — React 19 app in
src/. Routes for Crew (agents, skills, runtimes), Tasks (chat threads with live streaming), New Task, Auto (suggestions), and Onboarding. - Mobile — Expo app in
packages/mobile/pairs over an end-to-end encrypted Noise tunnel through a small relay, so you can read or nudge a running crew from your phone.
Privacy
- All UI, state, and orchestration happens on
127.0.0.1. Crew44 itself does not call out to any remote service. - The only outbound traffic is whatever the underlying coding-agent CLI you chose (
claude,codex, …) makes on its own — including the on-demand headless browser, which fetches its Playwright package and Chromium via npm and visits the URLs the agent navigates to. - Mobile pairing, when enabled, uses a relay for NAT traversal, but the payload is end-to-end encrypted with Noise, so the relay only sees ciphertext. You can self-host the relay if you prefer; the URL is configurable.
- No analytics, no error reporting, no phone-home.
Getting started
Prerequisites
- macOS, Windows, or Linux
- Node 20+ and Go 1.26+
- At least one coding-agent CLI installed (
claude,codex,cursor, …)
Run the desktop app
npm install
npm run dev
This builds the Go daemon, starts Vite, launches Electron, and connects the renderer once the daemon reports healthy.
Project layout
.
├── electron/ Electron main process, preload, app assets
├── src/ React renderer
├── packages/mobile/ Expo mobile app
├── daemon/ Go module
│ ├── cmd/crew44-daemon daemon entrypoint
│ ├── internal/ app, rpc, httpapi, store, runtime, agent adapters
│ └── test-utils/
├── docs/ manual e2e harnesses + design notes
└── public/ renderer static assets
Development
npm run dev # development app
npm run build # renderer build + local app
npm run web:dev # daemon + bare Vite development server
npm run test # Go tests + renderer tests + mobile tests
npm run clean # remove local build artifacts
Daemon configuration via env vars:
| Variable | Default | Description |
|---|---|---|
HOST / CREW44_DAEMON_HOST | 127.0.0.1 | TCP listen host. |
PORT / CREW44_DAEMON_PORT | 8080 | TCP listen port. |
AUTH_TOKEN / CREW44_AUTH_TOKEN | empty | WebSocket bearer subprotocol token. Empty = dev mode. |
CREW44_STATE_DIR | ~/.crew44 | Root directory for persisted state. |
When auth is enabled, /rpc requires WebSocket subprotocols:
new WebSocket("ws://127.0.0.1:8080/rpc", [
"crew44.rpc.v1",
`crew44.bearer.${token}`,
])
See the JSON-RPC method list and skill-injection walkthrough in docs/.
License
MIT © 2026
相似文章
@justloveabit: 用这个开源工具,我让一群AI替我上班了 事情是这样的,最近一直在折腾各种AI agent。Claude Code开一堆窗口,Codex也在跑,偶尔还要用Cursor。结果呢,乱成一锅粥——哪个agent在干啥,花了多少钱,完全搞不清楚。重…
本文介绍了一款名为Paperclip的开源工具,用于统一管理和调度多个AI Agent。它通过模拟公司组织架构、任务分配与预算控制等功能,解决了多Agent协作时上下文丢失、成本不可控和调度混乱的痛点。
@WWTLitee: 又是一个多 agent 协同工具:agency-agents 他可以直接把前端、社区、创意检查、现实校验这些角色分开跑,每个 agent 都有自己的边界、节奏和交付习惯,看起来更像一支小队在配合 仓库现在已经有 103.5k stars,…
Agency-Agents 是一个开源的 AI 多 agent 协作工具,集合了多个专业化角色(如前端、社区、创意检查等),每个 agent 有独立边界和交付习惯,适合构建多 agent 协作流程。GitHub 仓库已有 103.5k stars。
@yanhua1010: 突然有一个想法: 用 Multica去管理本地的runtimes (Claude Code, Codex…);用 Helio预定义常用Agent,包括skills,connectors;然后用 Obsidian 做Agent 的记忆或上下…
一个构想:用Multica管理本地runtimes(如Claude Code、Codex),用Helio预定义Agent,用Obsidian做记忆/上下文系统,再结合Harness Engineering,探讨本地多agent系统的最佳方案。
@justloveabit: 100个开箱即用的子代理,直接把Claude Code从“助手”升级成“完整AI公司”。 大多数人用Claude Code还停留在“单代理聊天”。 真正的高手,已经在用 100+ 专精Subagents 搭建自己的AI工程团队了。 神级仓…
介绍一个收录了100多个子代理的Awesome仓库,可将Claude Code从单代理助手升级为多代理协作的AI开发平台,并提供了快速上手指南和顶级子代理推荐。
多代理工作流的开源本地控制器——为Cursor、Claude Code、Codex、Antigravity会话提供统一收件箱及代理间群聊
一个名为Claude Command Center (CCC) 的开源本地面板,管理跨多个引擎(如Cursor、Claude Code、Codex和Antigravity)的多个AI编码代理会话,提供跨引擎收件箱、实时状态、git工作树生成和代理间群聊。