Armadra

让不同的 Agent,Put your agents on one canvas,在同一张画布上协作and let them work together

01软件The app

07

约定Agreement

保持现有 session 接口;改完互相审查。Keep the session API as is; review each other when done.

ama 协调者Coordinator主 · 2 从Lead · 2运行中Working

canvas team claude codex

2 名成员已就位,各带一条 worktree2 members ready, each on a worktree

等待成员回报…Waiting for members…

Claude Code 实现Implement从 @amaMember @ama需要你Needs you
权限请求Permissionpnpm test

读取 src/auth/login.tsxRead src/auth/login.tsx

编辑 src/auth/session.tsEdit src/auth/session.ts

新增 tests/session.test.tsAdd tests/session.test.ts

Codex 审查Review从 @amaMember @ama运行中Working

canvas inbox

读取 claude 的差异(经连线)Read claude’s diff (via link)

建议:cookie 加 sameSiteSuggest: add sameSite to the cookie

session.ts+2 −1

@@createSession(user, opts)

−const maxAge = DAY;

+const maxAge = remember ? 30 * DAY : DAY;

+setCookie('sid', token, { maxAge });

return token;

localhost:5173/login
登录Sign in

01每个 Agent 都是真实终端Every agent is a real terminal

02出发点Why

三个出发点Three starting points

  1. 01

    工作可以接续Work carries over

    相连的 Agent 读取摘要、转录与资料,阶段性交接。Linked agents read summaries, transcripts and material, and hand off at milestones.

  2. 02

    模型可以分工Models split the work

    经济模型做边界明确的事,强模型做复杂判断。Economical models take bounded tasks; stronger ones make the hard calls.

  3. 03

    过程可以查看The process stays visible

    Agent、资料、差异与成果摊在同一张画布上。Agents, material, diffs and results on one canvas.

03工作方式How it works

04

四步组织一次协作Four steps to a collaboration

  1. 01

    放上 AgentPlace agents

    每个节点启动一个真实 CLI。Each node starts a real CLI.

  2. 02

    连线Link them

    连线决定谁能读谁。Links decide who reads whom.

  3. 03

    分派与交接Assign and hand off

    协调者派活,阶段结果交接。A coordinator assigns; results are handed off.

  4. 04

    收回结果Gather results

    差异、测试与结论回到画布。Diffs, tests and conclusions land on the canvas.

04能力Capabilities

06

围绕画布的一整套工作面A whole work surface around the canvas

画布与白板Canvas and whiteboard

  • 九种节点Nine node types
  • 手绘、形状与 MermaidFreehand, shapes, Mermaid
  • 评论与 @ 提及Comments and @mentions

Agent 协作Agent collaboration

  • 按连线读取上下文Context along links
  • 收件箱与投递Inbox and delivery
  • 预览后再交接Handoffs you preview first

协调者与工作流Coordinator and workflows

  • 并行、流水线、汇总Parallel, pipeline, fan-in
  • 每人一条 worktreeA worktree per member
  • 模板定时重跑Templates on a schedule

终端与主机Terminals and hosts

  • tmux 会话跨重启tmux sessions survive restarts
  • 空闲自动休眠Idle agents hibernate
  • SSH 远端执行Remote execution over SSH

代码工作面Code surface

  • 编辑器与语言服务Editor and language services
  • 多仓库 GitMulti-repo Git
  • GitHub PR 评审GitHub PR review

远程与多端Remote and devices

  • 无窗口服务器壳Headless server shell
  • 账号与共享Accounts and sharing
  • 手机扫码配对Pair a phone by QR

05运行方式Where it runs

在桌面,也在服务器上On your desktop, and on a server

06AgentAgents

07

支持的 AgentSupported agents

  • 01Claude Code
  • 02Codex
  • 03OpenCode
  • 04Pi
  • 05Oh My Pi
  • 06GitHub Copilot
  • 07ama协调者Coordinator

07原则Principles

06

设计原则Design principles

01一个 core,两种壳One core, two shells
桌面与服务器装配同一份 core 与页面。Desktop and server assemble the same core and pages.
02真实 CLIReal CLIs
账户、模型与权限都沿用 CLI 自己的。Accounts, models and permissions stay the CLI’s own.
03连线即上下文Links are context
能读到什么,由画布上的连线决定。Links on the canvas decide what can be read.
04本地优先Local first
桌面安装没有服务端,数据留在本机。No server for the desktop; data stays on your machine.
05如实的状态Honest status
未知就显示未知,不显示成 0 或成功。Unknown shows as unknown, never as zero or success.
06最小侵入Minimal footprint
只在画布启动时注入,画布外的 CLI 不受影响。Injected only when launched from the canvas.

08最近更新Recent releases

v0.4.0

最近更新Recent releases

全部更新日志Full changelog

09开始Start

v0.4.0

把第一组 AgentPut your first agents放上画布on the canvas