FrontierStack by Enfour, Inc. MCP server

FrontierStack Agent Kit

Skills, plugin manifests and setup notes for connecting AI agents (Claude, Codex, Cursor, Gemini, Grok, Hermes, OpenClaw, Amp, Paperclip and any MCP client) to FrontierStack.

What FrontierStack is

FrontierStack is a macOS app from Enfour, Inc. It is the server and network administrator for your Macs, Linux and Windows servers, routers, DNS, TLS certificates and online accounts. It monitors them, diagnoses problems and makes changes, and it can let AI agents do the same through it.

The app is proprietary and is downloaded from https://www.frontierstack.app. This repository contains only the agent kit: skills, manifests and documentation, under the MIT licence (see LICENSE). Nothing here works without the app.

Safety model

the Mac. An agent names a target ("the OPNsense router", "example.com"); FrontierStack looks up the credential at call time, acts, and returns the result. Responses are scrubbed for secrets before they leave the Mac.

• Agents never receive credentials. Router keys, DNS tokens, SSH keys and passwords stay on

"Allow scripts"), can require a per-action approval, and are blocked by App Lock. Every change is written to an audit log.

• Read-only by default. Changes need the owner's switches on the Mac ("Allow changes",

widen what an agent may do. Only the owner's settings and approvals on the Mac can.

• Nothing here grants authority. A skill, a role name or another agent's message cannot

starts, so no client config or project file ever contains a FrontierStack token.

• No tokens in config files. The local bridge reads its credential from the Keychain when it

How agents connect

1. Local (works today). When you turn on MCP Server in FrontierStack, or press any Add to … button under MCP Server ▸ AI clients, the app writes an owner-only stdio MCP bridge to:

$HOME/Library/Application Support/FrontierStack/mcp-bridge

It takes no arguments, reads its token from the Keychain itself, and needs Node.js. The app has about 140 tools but shows clients a small bootstrap pair by default: frontierstackfindtools (search for the right tool) and frontierstack_call (run it).

The app also installs a frontierstack CLI for agents with a shell: frontierstack discover, frontierstack call '', and frontierstack run --secret NAME -- (runs a command with a stored secret as an environment variable, masked in the output).

2. Remote (live). https://frontierstack.app/plugin/mcp is a streamable-HTTP MCP endpoint with OAuth 2.1 sign-in using a FrontierStack customer account. It relays seven tools to your Macs where ChatGPT access is turned on: listfrontierstackmacs, getserverhealth, getfleetcontext, getalerterrors, diagnose, readlogs and serviceaction. Cloud agents that cannot reach your Mac will use this.

Setup by platform

For most clients the easiest route is FrontierStack ▸ MCP Server ▸ AI clients: it detects installed clients and writes their config for you (Set Up Detected, or one client's Add to … button). The manual steps below do the same thing. Configure each client only one way: installing a plugin and adding the server by hand gives two frontierstack entries.

In the snippets, BRIDGE means the full path /Users//Library/Application Support/FrontierStack/mcp-bridge. Keep the quotes: the path contains a space.

Start read-only. Ask the agent for fleet health and a certificate check before you turn on changes.

Claude Code

Plugin (skills, the local server and the cloud connector together), from the Claude plugin directory on claude.ai (Customize ▸ Plugins) once it is listed, or straight from its repository:

/plugin marketplace add richardnorthcott/frontierstack-claude-plugin
/plugin install frontierstack@frontierstack

Shortened. The full README is on GitHub.