Agent access (MCP)
Let AI assistants connect to the running app over a local, loopback-only MCP endpoint.
Ganter Lab embeds an MCP server so an AI assistant can connect to the running app — read your configuration, help configure it, operate logic or drive the UI — all on loopback only, never exposed to the network.
Nothing is offered to an agent until you say so, and what you say is read in two steps: the master switch decides whether the endpoint exists at all, and one toggle per area decides which tools it advertises. This area is the complete reference for that endpoint, so you can decide exactly what to allow.
In this area
| Page | What it covers |
|---|---|
| Turning it on | The master switch, the port, and the four things that keep the endpoint local. |
| What each area opens | Every capability toggle, its tier, its default, and the tools it advertises. |
| Connecting a client | Copy-ready setup for Claude, Codex, Gemini CLI, GitHub Copilot and anything else that speaks MCP. |
| The tools, area by area | What every tool does, grouped by the toggle that advertises it. |
| Guardrails and journaling | The funnels and locks every write shares, and the trail an agent leaves. |
Capability tiers
Every toggle belongs to one of four tiers, and the tier tells you the blast radius before you open it:
| Tier | What it can do | Default |
|---|---|---|
| Read | Browse and inspect. Changes nothing. | on |
| Configure | Create, update and delete saved configuration. Journaled on the Events page and reversible from a configuration backup. | mostly on |
| Operate | Reach live equipment: write values, run actions, start runs, control the embedded server, drive the UI. | off |
| Restore | Schedule a configuration restore — replaces the whole configuration at the next start. | off |
The tiers are also the shape of the defaults: reading and configuring are open once the master switch is on, and everything that reaches live equipment stays off until you open it deliberately.