GETTING STARTED / 使用说明

A clear first handoff.

Web, installable PWA and local MCP connector all use the same workspace. This is a draft-only pilot. No email, buying, browsing, file execution or publishing tools are enabled.

01 / Start on the web

Choose Open workspace and create a workspace. Save the access token in a password manager. There is no email recovery. The token grants full workspace access.
Try Demo first. Its three synthetic stages do not call models or spend API credits. They demonstrate handoffs, not actual AI reasoning.
For real work, open Manage connections. Select OpenAI API or Meta Model API, enter your exact available model ID and provider API key, then save. Saving does not verify access; the first live task does.
Add your brief and source excerpts. URLs alone are not fetched. Choose the provider for each role. You can use one or both providers.
Select Live, authorize data sharing and click Run the handoff. Each stage receives the same brief, source text and earlier drafts. API charges are separate from consumer subscriptions.
Review the three artifacts. Accept deliverable marks your approval only; it sends nothing externally. Export Markdown or JSON for your own use.

You can stop active tasks. An in-flight provider request may still incur charges. Retry is available once for failed/interrupted tasks and resumes after completed stages. It may repeat a billed request whose response was not saved.

02 / Install the PWA

iPhone / iPad: open the HTTPS site in Safari → Share → Add to Home Screen. If you are in an in-app browser, open Safari first.

Android / desktop: use the page's Install the app button when offered, or your browser's app-install menu. Installation support depends on browser and device.

PWA is an installed website, not an App Store binary. It uses the same server and workspace. If a new app window asks you to sign in, use your saved workspace token.

The app shell and guide are available offline after a successful first load. Tasks, connections, history and downloads need a network connection. Tasks are never queued offline. API responses and secrets are not cached by the service worker.

03 / Connect local agents

Download the local connector ZIP ↓

Download the full local workspace ZIP ↓. Install Node.js 24+, unzip, then use start-windows.cmd on Windows, or run node scripts/launch.mjs from the extracted folder on macOS/Linux. The launcher opens a browser on localhost; keep its terminal open. macOS/Linux shell launchers are included and may need chmod +x. Runtime data is saved under data/; keep its development.key private.

For the smaller MCP-only bridge: requires Node.js 24+. Unzip and point a compatible MCP client at connector.mjs. The connector runs on Windows, macOS and Linux; no native installer or code signature is included.

{
  "mcpServers": {
    "agentsfair-connect": {
      "command": "node",
      "args": ["/ABSOLUTE/PATH/connector.mjs", "mcp"],
      "env": {
        "AF_URL": "https://YOUR_CONNECT_ORIGIN",
        "AF_TOKEN": "YOUR_WORKSPACE_ACCESS_TOKEN"
      }
    }
  }
}

Use Settings → Download MCP config template to insert this server's origin automatically. Replace the local path and token yourself. Do not share the filled-in file. Client setup formats vary.

Available tools: task_create, task_get, task_list, task_cancel. Start with this instruction to your client:

Use Agents Fair Connect to create a demo task:
mode: demo
goal: Turn my supplied notes into a pilot plan and review the assumptions.
Then read the returned task ID until its status is awaiting_review.

The server also offers a stateless MCP HTTP endpoint at /mcp. Requests require Authorization: Bearer YOUR_WORKSPACE_TOKEN. It supports task tools over POST. It does not provide OAuth discovery; consumer applications requiring OAuth need an additional integration. Personal Muse and ChatGPT end-to-end connection has not been verified.

Data & permissions

The server stores workspace names, hashed access tokens, encrypted API keys, task briefs, source text, model drafts and usage counts in a local SQLite database. API keys are encrypted using AES-256-GCM; the server operator controls the master key and can decrypt them. Encryption is not protection from the server operator.

Live tasks send their brief, source text and preceding drafts to each assigned provider. Review those providers' current retention policies before supplying sensitive information. This app requests store:false; that is not a guarantee of zero provider retention.

Remove a provider key in Settings. Delete workspace removes its active database rows, keys, tasks and usage records. Deleted SQLite content may remain in database pages, WAL files or operator backups until those are rotated or securely erased. Exports and provider records remain separate.

Workspace tokens function as bearer credentials. Anyone holding one has full access, including deletion and spending available API credits. Use separate workspaces for different customers. Rotate a token to revoke existing sessions and local connectors. There are no collaborator roles in this version.

No third-party analytics, remote fonts or tracking scripts are included.

Current limits

Three fixed stages: analyst → creator → reviewer. No open-ended agent conversations.
One active task per workspace. One manual retry per task. Maximum 30 provider call attempts per workspace over 24 hours.
256–4,096 maximum output tokens per stage. A token limit is not a dollar spending cap; use your provider's billing limits.
Each request has a 90-second timeout. Restarted tasks require a manual retry. Keep the server running for background work.
Source input is text only, up to 12,000 characters. Local .txt/.md files are copied into that field; links are not fetched.
API model adapters and MCP task tools are implemented. Real provider access, consumer-client compatibility and installation on physical devices require your acceptance test.
No A2A implementation, OAuth login, native desktop installer, payment system, local file agent or automatic site deployment is included.
Single Node process and persistent local disk. Do not deploy this SQLite version as stateless serverless functions or multiple independent replicas.

中文快速说明

网页版：启动后台或由上线 AI 部署到 HTTPS 后，打开网页。点击 Open workspace 创建工作空间，立即保管好访问令牌。令牌相当于该空间的全部权限；没有邮箱找回。
先试演示：选择 Demo。三段内容是标记清楚的固定示例，不是三个真实 AI 在思考，也不会调用收费接口。
连接真实模型：点击 Manage connections，选择提供商，输入你账户可用的准确模型 ID 与 API Key。保存会加密，但不代表验证成功；真实任务成功才证明该模型可用。API 计费与 ChatGPT/Muse 个人订阅不同。
提交任务：写目标，粘贴资料和证据摘录，给三个角色分配提供商。切换 Live 并勾选资料共享授权。模型无法浏览网页，因此只贴链接不足以提供证据。
查看成果：依次查看整理、创作与复核稿。Accept 仅表示你验收，不会发邮件或上线。可下载 Markdown 或 JSON。
安装到手机/电脑：支持的浏览器点击 Install the app；iPhone 用 Safari 的分享菜单 → 添加到主屏幕。这是 PWA，不是应用商店安装包。任务需要联网，后台服务器必须持续运行。
本地连接器：下载 ZIP，解压，安装 Node.js 24+，将 connector.mjs 路径、网站地址和工作空间令牌填入兼容 MCP 客户端。它不会自动控制电脑上其他 AI，也不读取你的个人历史。
费用和资料：Live 会把任务资料与前面阶段的稿件传给所选提供商；每空间每天最多 30 次 API 尝试，每任务最多重试一次。停止不保证撤销已产生的费用。不要把访问令牌放进公开配置或截图。

本版本适合自用与受控试点。公开商业运营仍需由上线方配置 HTTPS、持久化磁盘、加密主密钥和准入规则，并使用你自己的 API 密钥完成真实测试。没有修改或部署现有网站。