Edgegap MCP server

edgegap-mcp

An MCP server for Edgegap that lets a coding agent take a developer from "I have a game server container" to "players are connected to it" without the developer reading the API reference.

Ten tools, hand-picked. Not generated from the OpenAPI spec — see Scope for why.

Install

Two ways to run it. Pick based on how much you care about where your token goes — see Where your token goes.

Remote endpoint

Hosted by Edgegap as a Cloudflare Worker. Nothing to install.

{
  "mcpServers": {
    "edgegap": {
      "type": "http",
      "url": "https://mcp.edgegap.dev/mcp",
      "headers": { "Authorization": "token YOUR_API_TOKEN" }
    }
  }
}

Also works as a custom connector in claude.ai: add https://mcp.edgegap.dev/mcp and supply the same token.

Local

Runs on your own machine, spawned by your editor. One line in your MCP client config, nothing to clone, nothing to build.

{
  "mcpServers": {
    "edgegap": {
      "command": "npx",
      "args": ["-y", "@edgegap/mcp"]
    }
  }
}

Works in Claude Code, Cursor, Codex, and VS Code. Pin a version in production (@edgegap/mcp@0.1.5) rather than floating on latest.

Registered in the official MCP registry as dev.edgegap/mcp.

Node version: the local server needs Node 18+. Deploying your own copy of

the Cloudflare Worker needs Node 22+, because wrangler requires it.

Where your token goes

This differs by mode, and the difference is the reason both modes exist.

Local. The server runs as a process on your own computer. The first tool call asks you for a token, shows what it authorises, and requires an explicit acknowledgement before accepting it. Where that token then lives, exhaustively:

• one variable in that process's memory, for the life of your editor session

That is the whole list. Not on disk. Not in a config file. Not in logs. Not on any Edgegap server — the only thing sent to Edgegap is the API call itself, exactly as if you had run curl. Closing your editor revokes this server's access completely.

Remote. Your token is sent to mcp.edgegap.dev on every request and forwarded from there to the Edgegap API. It transits infrastructure Edgegap operates. The worker holds it for the life of the request and does not persist it, but that is a "we don't store it" claim rather than a "we never see it" claim. The two are different, and only local mode makes the second one.

Generate a token at .

In local mode, setting EDGEGAPAPITOKEN takes precedence over the prompt, for CI and for clients that cannot show prompts. Do not pass a token as a command-line argument — arguments are visible to other processes via ps, and the server warns if it detects one.

Which to use. Remote for a first try, a demo, or a supervised session where setup friction matters more than custody. Local for anything unattended, anything in an organization with a live game in it, and anything where you would rather not extend trust you don't have to. The guardrails described below exist only in local mode.

Read this before connecting an agent

The Edgegap API token cannot be scoped. One token authorises every application, every version, every running deployment, and your usage across the whole organization. There is no deploy-only token and no per-application token.

Consequences worth being deliberate about:

test ones it created.

• An agent holding this token can stop production deployments, not just the

page — reaches the token too.

• Prompt injection reaching the agent — from a repo file, an issue, a fetched

token could end up. This server does not log it, but it cannot control what the rest of the agent does.

• Anything the agent logs, echoes, or sends to a model provider is a place the

Edgegap's worker on every call.

• On the remote endpoint, the same unscoped token is additionally handled by

Shortened. The full README is on GitHub.