Armadra 是一块本地优先的 AI Coding 画布。Claude Code、Codex、OpenCode、Pi、Oh My Pi 与 GitHub Copilot 以真实终端节点的形式放在无限画布上，保留各自的账户、模型与权限策略。你通过连线共享必要的上下文，通过消息和交接组织分工，在同一张画布上查看资料、执行过程与成果。

从使用一个 Agent，到组织多个 Agent

单独使用一个 Agent 时，上下文都在一个会话里。工作一旦跨越多个会话或多个 Agent，就会遇到这些问题：

• 接手的 Agent 不知道已经确认的决定、失败过的方法和结果的位置；
• 不同难度的步骤用同一个模型，预算没有花在最需要判断的地方；
• 多个终端窗口之间的关系只存在于你的脑子里。

Armadra 用画布把这些关系变成看得见、能操作的工作空间：

• 工作可以接续：相连的 Agent 按范围读取摘要、转录与相关资料，交换消息并进行阶段性交接。
• 模型可以分工：为节点选择不同模型，让经济模型承担边界明确的工作，让更强的模型处理复杂判断。当前分工由你或任务指令安排，自动难度路由尚未实现。
• 过程可以查看：Agent、便签、编辑器、Git 差异、文件、浏览器和白板共同构成工作面。

节点里运行的是完整的 Agent 产品

Armadra 不重写 Agent。每个 Agent 节点启动的就是你已经在用的 CLI，复用它的多轮执行、工具调用和会话机制；Armadra 只负责节点身份、连线、消息与可读上下文。

适合哪些工作

• 软件工程：需求分析、代码开发、测试、问题诊断与交叉审查。
• 算法与数据科学：算法实现、数据处理、模型训练与评测、实验结果比较。
• 科研：论文复现、计算实验、数据分析与可视化、研究结果复核。

运行方式

• 桌面应用：macOS、Windows、Linux 安装包，数据保存在本机 SQLite 与工作区的 .armadra/。
• 无窗口服务器壳：在服务器或 NAS 上常开，浏览器与手机经 HTTPS 访问同一份页面，支持账号与按工作空间共享。

准备好了就去安装。

Armadra is a local-first AI coding canvas. Claude Code, Codex, OpenCode, Pi, Oh My Pi and GitHub Copilot run as real terminal nodes on an infinite canvas, each keeping its own accounts, models and permission policy. You share the context that matters through links, split the work with messages and handoffs, and see the material, the process and the results on one canvas.

From one agent to a team of agents

With a single agent, the whole context lives in one session. Once work spans several sessions or several agents, familiar problems appear:

• The agent picking up doesn’t know what was decided, what already failed or where the results are;
• Every step uses the same model, so the budget isn’t spent where judgment matters most;
• How the terminal windows relate exists only in your head.

Armadra turns those relationships into a workspace you can see and act on:

• Work carries over: linked agents read summaries, transcripts and related material within scope, exchange messages and hand off at milestones.
• Models split the work: pick a model per node so economical models take well-bounded tasks and stronger ones handle hard judgment calls. Today the split is arranged by you or the task instructions; automatic difficulty routing is not implemented.
• The process stays visible: agents, notes, editors, Git diffs, files, browsers and the whiteboard form one work surface.

Each node runs a complete agent product

Armadra doesn’t rewrite agents. Each agent node launches the CLI you already use and reuses its multi-turn execution, tool calls and sessions; Armadra only adds node identity, links, messages and readable context.

What it is good for

• Software engineering: requirements analysis, implementation, testing, diagnosis and cross-review.
• Algorithms and data science: algorithm work, data processing, model training and evaluation, comparing experiments.
• Research: reproducing papers, computational experiments, data analysis and visualization, checking results.

How it runs

• Desktop app: installers for macOS, Windows and Linux; data lives in local SQLite and the workspace’s .armadra/.
• Headless server shell: always on, on a server or NAS; browsers and phones reach the same pages over HTTPS, with accounts and per-workspace sharing.

When you’re ready, head to Install.