How do I add Offline Protocol to Claude?

Connect Offline Protocol to Claude and Claude can plan your app’s architecture, pick compatible packages, read the integration guide for your workflow and, with the local server, scaffold a starter app. Which route you take depends on where you use Claude. The Offline Protocol MCP server comes in three forms, and two of them matter here:

• The directory connector, for claude.ai and the Claude desktop app. It connects the public read-only endpoint, https://mcp.offlineprotocol.com/public/mcp, with no account, key or headers.
• The local MCP server, for Claude Code, and for Claude Desktop when you want files written. It runs on your machine from the published Offline Protocol CLI over standard input and output, and needs Node.js 18 or later but no login.

The third form, the keyed hosted endpoint at https://mcp.offlineprotocol.com/mcp, takes an app API key and its App ID in request headers. For claude.ai and Claude Desktop, the docs point to the public endpoint, which needs no headers.

Configure one mode per client under the offline-protocol name, not two side by side.

The Claude directory route takes one click. Open the Offline Protocol listing in Claude’s directory and connect it. Claude’s documentation describes the button as Connect to Claude, and says a connector that doesn’t use sign-in connects as soon as you select it, with no page to visit. The Offline Protocol endpoint has no sign-in.

After that, the connector works in your conversations on the web, desktop and mobile. Claude’s docs describe three things to know:

• It is switched per chat. Select + in the message box, then Connectors, and check that its toggle is on.
• Claude may ask before each tool call. Choose Allow once, or Always allow to skip the prompt for that tool.
• You manage it in one place. Tool permissions, disconnecting and removing all live on the connector’s page under Customize > Connectors.

If you would rather not use the directory, add https://mcp.offlineprotocol.com/public/mcp as a custom connector by its URL. No sign-in is needed.

For Claude Desktop with the file-writing tools, use local mode instead: choose Settings, Developer, Edit Config and add this to claude_desktop_config.json:

{
  "mcpServers": {
    "offline-protocol": {
      "command": "npx",
      "args": [
        "-y",
        "@offline-protocol/cli@0.2.9",
        "mcp",
        "serve"
      ]
    }
  }
}

Claude Code takes the local server with one command, run once:

claude mcp add offline-protocol -- npx -y @offline-protocol/cli@0.2.9 mcp serve

The first launch downloads that CLI release from npm and later launches use the npm cache. Claude Code’s documentation says claude mcp add writes to local scope by default, so the server loads only in the project where you added it. claude mcp list shows whether it connected, and /mcp shows it inside a session.

To share the setup with your team, commit this as .mcp.json in your project instead:

{
  "mcpServers": {
    "offline-protocol": {
      "command": "npx",
      "args": [
        "-y",
        "@offline-protocol/cli@0.2.9",
        "mcp",
        "serve"
      ]
    }
  }
}

Claude Code asks each person to approve a project’s .mcp.json servers the first time they run claude in it, so a cloned repository cannot switch the server on by itself.

If you signed in to Claude Code with your Claude account, connectors you added on claude.ai also appear in Claude Code, so the directory connector’s read-only tools can show up next to the local server. The keyed hosted endpoint works in Claude Code too, with OFFLINE_API_KEY set in your environment:

claude mcp add --transport http offline-protocol https://mcp.offlineprotocol.com/mcp --header "Authorization: Bearer $OFFLINE_API_KEY" --header "x-app-id: app_your_application_id"

That command writes the key into your client configuration, so keep it private.

With either route, Claude can plan the architecture of an app, pick compatible packages and read the integration guide for a workflow. The local server adds scaffolding a starter app on disk.

The 15 read-only tools cover package, workflow, guide and template lookup, package resolution, architecture planning and scaffold preview. Among them are generate_architecture for a plan, resolve_packages to choose compatible packages, get_skill to read an integration guide and preview_scaffold to show what a scaffold would contain without writing it.

Three ready workflows scaffold a React Native app: local-handoff, backend-delivery and nearby-service. If you are building something like a chat app, How to build an app like Bitchat gives the background to brief Claude with.

A concrete request with your platform and devices gets a better answer than a general one. The Offline Protocol docs suggest a first request like this:

Add an encrypted local handoff to this React Native app. Explain the device requirements, choose compatible packages, and show how the receiving app will persist the handoff.

Tell Claude your repository, target device models and OS, installed SDK version and intended workflow. The docs keep a live agent setup prompt as the canonical setup instruction; give Claude its URL rather than a copied version.

Once Claude has planned and scaffolded the integration, the next steps happen on your machine and your devices:

• Compile and test on the target devices. The docs ask for a device test because radio delivery is confirmed on hardware, not by a typecheck or mock.
• Review what the local tools write. They write files and run package-manager commands with your permissions, so keep Claude’s approval prompts on.
• Keep keys out of committed files. The local and public modes need none; a hosted key belongs in your environment.

The same server works in other tools: see Cursor, VS Code, and ChatGPT and Codex. For a comparison with other vendors’ servers, read Which mesh networking SDKs have an MCP server?