ShellYard

Connectors — MCP server

Run ShellYard as a local Model Context Protocol server so Claude Desktop, Claude Code, or a local Ollama client drives your tools on its own model subscription. Pro tier.

Settings → Connectors turns ShellYard into a local Model Context Protocol server. An MCP client — Claude Desktop, Claude Code, or an MCP-capable client pointed at a local Ollama model — attaches to it and drives ShellYard’s tools directly. Pro tier and above.

This is the inverse of Magellan. Magellan is a chat panel inside ShellYard that you supply an API key to. Connectors puts ShellYard behind the assistant you already pay for, so the model runs on its own subscription and ShellYard never needs an Anthropic key of its own.

Why this exists

Your inventory, your credentials, and your device access live in ShellYard. Your assistant lives somewhere else. Without a connector you end up copy-pasting show interface output into a chat window and pasting the answer back.

With the connector attached, the assistant can ask ShellYard directly — “which switches in the Acme space have an interface flapping?” — and ShellYard answers from real device reads, with your approval gate in front of anything that changes state.

Starting the server

Settings → Connectors → Start server. You get:

  • A streamable-HTTP endpoint, spec-compliant, bound to 127.0.0.1 only. It is not reachable from your LAN, and there is no configuration that makes it reachable.
  • A per-install bearer token, required on every request. Rotate it any time from the same panel — rotation invalidates every attached client immediately.
  • A live client list showing what is currently attached and what it has called.

The server stops when you quit ShellYard. It does not run as a background daemon and does not survive a restart unless you start it again.

Tool surface

The exposed surface is curated, not a dump of every internal binding. It covers:

  • Inventory — list spaces, folders, connections, and their metadata
  • Documents and IPAM — search and read notes, runbooks, subnets, hosts
  • Network lookups — DNS, WHOIS, subnet math, MAC/OUI vendor lookup
  • Device reads over SSH — run read-only show/status commands against saved connections
  • Session and streaming wrappers — attach to an open session’s output

The approval gate

Tools split into two classes.

Read-only tools run straight through. The assistant lists your connections or resolves a DNS name without interrupting you.

Action tools pause for an in-ShellYard approval modal before anything executes. The modal shows the exact tool, the exact arguments, and the target. Nothing runs until you click Approve in the ShellYard window — not in the assistant’s UI, where a compromised or confused model could render whatever it wanted.

Action tools include:

  • Bulk command execution
  • Port-mirror create and remove
  • Enable SCP on a device
  • Remote packet capture start and pull
  • Create or modify a document

What never crosses the boundary

Two hard rules, enforced server-side rather than by prompt instruction:

  • Credentials are projected to metadata only. The assistant can see that a connection uses credential “acme-core-switch — admin,” and can ask ShellYard to use it. It cannot read the secret. There is no tool that returns a password, key, or token.
  • Document lists come back as summaries. Titles, paths, and sizes. Full body text requires an explicit read of a named document, so a broad “list everything” call can’t quietly exfiltrate a runbook full of pasted secrets.

Client setup

Per-client snippets live in the Connectors panel — copy the one for your client. See Connecting a client for the full walkthrough of each.

What it isn’t

ChatGPT is not supported, and this is not an omission we can fix. ChatGPT’s connector infrastructure is cloud-hosted; it reaches out from OpenAI’s servers to a public URL. It cannot reach 127.0.0.1 on your laptop. Exposing the ShellYard MCP server to the public internet to work around this would defeat the entire security model, so we don’t offer a way to do it.

It is also not a replacement for Magellan. Magellan sees the live UI state — the active tab, the tool output on screen, the query you just ran. The MCP surface is inventory and device access. Most people who use both run Magellan for “explain what I’m looking at” and a connector for “go find out.”