GUIDE

Agent Studio, in five minutes.

A local app where three AI agents plan, produce and review a task together, correcting each other automatically. You bring your own API keys; nothing is uploaded anywhere except the model calls you authorise.

1. Install

Install Node.js 24 or newer (free, one installer).
Unzip the Agent Studio download anywhere, e.g. your Documents folder.
Start it: Windows → double-click start-windows.cmd; macOS → double-click start-macos.command (first time: right-click → Open); Linux → ./start-linux.sh.
Your browser opens http://127.0.0.1:3900. Keep the black window open while you use Studio; close it to stop.

No installer, no admin rights, no npm packages. Everything lives in the folder you unzipped, including the data/ folder with your jobs and encrypted keys.

2. Watch the demo

Press Watch a demo. Synthetic agents walk through the loop — plan, draft, a review that sends it back, a corrected draft, approval. No key, no cost. It shows you exactly what a live run looks like.

3. Add API keys

Click API keys. For each provider you want to use, paste the key and keep or change the model name, then Save & verify — Studio makes one tiny test call so you know it works.

Anthropic (Claude) — console.anthropic.com → API keys. Default model claude-sonnet-4-5.
OpenAI (GPT) — platform.openai.com → API keys. Default model gpt-4.1.
Google (Gemini) — aistudio.google.com → Get API key. Default model gemini-2.5-flash.

Model names change over time; if a save fails with “check the model name”, look up the current name in that provider's documentation. Costs are billed by the providers to your own accounts; a typical 3-round job uses 5 to 7 model calls.

4. Run a live job

Write the brief: what to produce, for whom, in what shape. Paste any facts into Material — agents cannot browse, so this is their only source of truth.
Pick the team. Tip: use a different vendor for the Reviewer than for the Worker — models are better at finding other models' mistakes than their own.
Set Mode to Live, tick the consent box, choose max rounds (3 is a good default), press Run the team.
Watch rounds appear. Each review lists concrete issues with severity; each revised draft answers them. When the Reviewer passes it (or the round limit is hit), the Final deliverable appears.
Read it critically, then Accept, Copy or Export (.md keeps the full collaboration record).

5. Use it from Claude, Cursor or any MCP client

Studio is also an MCP server, so your chat assistant can hand it a brief and fetch the result — no copy-paste. Add this to your client's MCP configuration (use the absolute path to studio-connector.mjs):

{
  "mcpServers": {
    "agent-studio": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/agent-studio/studio-connector.mjs"]
    }
  }
}

Studio must be running. Tools: studio_job_create, studio_job_get, studio_job_list, studio_job_cancel. Example prompt to your assistant: “Create an Agent Studio live job to draft the Q1 newsletter with these notes, then fetch the final deliverable when it is ready.”

What Studio does not do

It does not browse the web, read files, send email, pay for anything or publish. Agents only write and critique text.
It does not control ChatGPT, Claude or other apps on your computer, and does not read their history.
It does not guarantee correctness. A reviewer catches many errors, not all. Verify facts before using the result.
It stores keys encrypted on this computer only; the operator of your computer can still read them. Do not run it on a shared machine.

中文说明

这是什么：一个在你自己电脑上运行的小程序。你给一个任务，三个 AI 角色自动协作——规划者定标准，执行者写出成果，审核者挑错并退回修改，循环到通过为止。你只看最终版，不用在几个 AI 之间复制粘贴。

安装：① 安装 Node.js 24 以上（nodejs.org，免费）；② 解压下载包；③ Windows 双击 start-windows.cmd，macOS 双击 start-macos.command（第一次需右键→打开）；④ 浏览器自动打开 http://127.0.0.1:3900。黑色窗口保持打开，关掉即停止。

先看演示：点 Watch a demo，不需要任何 Key，看完整流程。

加 API Key：点右上角 API keys，粘贴 Anthropic / OpenAI / Google 任意一家或多家的 Key，点 Save & verify，程序会发一条极小的测试请求确认可用。Key 加密保存在本机 data/ 目录，只会发给你选的那家厂商。

正式运行：写清任务和要求；把资料贴进 Material（AI 不能上网，这是它唯一的事实来源）；分配角色——建议审核者用与执行者不同厂商的模型，互审比自审更能抓错；Mode 选 Live，勾选同意，设置最多几轮（默认 3），点 Run the team。每一轮的草稿、审核意见、修改都会显示出来；审核通过或到达轮数上限后出现最终成果，可复制、导出 Markdown（含完整协作记录）。

接入 Claude / Cursor 等：Studio 同时是一个 MCP 服务，把上面第 5 节的配置加进你的客户端，就能对助手说"把这个任务交给 Agent Studio 做，做完把结果拿回来"。

不做什么：不上网、不读文件、不发邮件、不付款、不发布；不控制你电脑上的 ChatGPT 或 Claude 应用；不保证结果绝对正确——审核能抓很多错，但不是全部，事实请自行核对。费用由各厂商按你自己的账号计费，一次 3 轮任务通常 5–7 次模型调用。