What is the Offline Protocol MCP server?

The Offline Protocol MCP server is the way a coding agent gets accurate, current information about Offline Protocol instead of relying on what its model remembers. MCP, the Model Context Protocol, is an open protocol that lets AI applications call tools a server exposes. This server exposes the Offline Protocol package registry, integration guides (called skills), workflows, starter templates, an architecture planner and, in local mode, scaffolding.

With it connected, an agent can find packages that fit your platform, read the guidance for the workflow you chose, plan an implementation and generate project files. It works on your code. The mesh itself runs in your app through the SDK you choose.

The server comes in three modes, documented on the CLI and MCP page:

The local server is part of the offline CLI. Your MCP client starts it with one command; the first launch downloads the release from npm and later launches use the npm cache.

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

With the CLI installed globally, offline mcp serve starts the same server. The hosted endpoint rejects organization keys, and limits requests to 60 a minute per key and 600 a minute per network address. The public endpoint uses streamable HTTP and allows 120 requests a minute per network address. A limited request gets 429 with Retry-After in both cases.

The hosted and public endpoints offer 15 read-only tools for package, workflow, guide and template lookup, package resolution, architecture planning and scaffold preview:

• list_packages, get_package, search_packages, resolve_packages
• search_capabilities, search_examples
• list_workflows, get_workflow, search_workflows
• list_skills, get_skill, search_skills
• get_template, generate_architecture, preview_scaffold

The local server provides 18 tools covering packages, capabilities, workflows, examples, skills, templates, planning and scaffolding. Three of them write local project files and exist only in local mode: scaffold_project, integrate_packages and init_project. Run offline mcp tools to see the exact list your installed version serves. Each CLI release resolves the SDK release that was current when it shipped, and CLI 0.2.9 resolves SDK 0.31.0.

The MCP server works on your code, and your app works on your devices. The agent plans and writes the integration with the server’s tools; the built app runs the mesh, so you confirm radio delivery by compiling the changed app and testing it on the target devices.

Scaffolding, installing packages and editing a project run through the local server, while the hosted endpoints serve lookup and planning. Local tools write files and run package-manager commands such as npm install with your permissions, subject to your client’s approval settings, and they are not sandboxed, so keep approval prompts on. The local server and the CLI send no usage analytics or crash reports. Hosted MCP receives the tool arguments your agent sends and keeps rate-limit counters, keyed by a hash of the credential or by network address, in memory for about 90 seconds.

A separate docs MCP server at https://www.offlineprotocol.com/docs/mcp only searches and reads the documentation. It is useful alongside this server and does not replace it.

Four AI coding tools have a one-click install from the “Start building” section of the Offline Protocol homepage. Each has its own article with the manual setup as well:

• Claude. The Offline Protocol connector in Claude’s directory adds the public read-only endpoint to Claude.ai and Claude Desktop. See adding Offline Protocol to Claude.
• Cursor. An install link adds the local server, which Cursor starts through npx. See adding Offline Protocol to Cursor.
• VS Code. The GitHub MCP Registry listing has an Install MCP server button, and GitHub Copilot uses the tools in agent mode. See adding Offline Protocol to VS Code.
• ChatGPT and Codex. An Offline Protocol plugin installs from ChatGPT. See using Offline Protocol in ChatGPT and Codex.

Claude Code adds the local server with claude mcp add, and other MCP clients such as Gemini CLI and Windsurf take the same local server entry in their MCP configuration file. offline mcp install <client> prints configuration for claude-code, claude-desktop, codex, cursor and vscode without editing your files.

The MCP server works best with a concrete brief. The docs recommend starting the agent from the live agent setup prompt, then giving it your repository, target device models and OS, installed SDK version, intended workflow, available transports and storage or backend requirements, along with what success looks like and which device tests it must run.

A first request can be as specific as adding an encrypted local handoff to a React Native app, with the device requirements explained and the receiving app persisting the handoff. The docs’ Build with an AI agent page lists reading paths for sending between two phones, sharing a checklist, delivering readings to a backend and calling a nearby service.

Offline Protocol’s server is an integration and scaffolding server: it helps an agent write code against the SDK. Other vendors publish servers for different jobs, such as searching documentation or driving radio hardware. Which mesh networking SDKs have an MCP server? compares them and lists the questions to ask before you rely on any vendor’s server.