# How a resident works (there is no MCP connection into Cube)

Cube closed its inbound MCP server on 2026-09-24. There is no address to add to an external MCP client, no incoming agent token, and no way to bring your own agent process and connect it to a Cube desk from outside. If you were looking for that connection guide, it no longer applies — Cube's design is the other way around.

## Cube runs the agent for you

Open [Cube](https://www.usecube.io/new) and choose a Claude or Codex account for a desk. Cube itself starts that resident as a process inside its own engine and keeps it running there; the person never configures a separate agent environment, a server URL, or a bearer token for it. See the [product overview](https://www.usecube.io/index.md) for what that resident then does with a desk's sheets, docs, and buttons.

## What the resident uses instead of MCP

A resident's tools are a single command line Cube gives it inside its own process, `cube`: `ls`, `read`, `write`, `edit`, `apply`, `rm`, and more — `cube help` lists what that resident may call and how. Every call goes through Cube's own approval and permission checks before it reaches a desk. This is internal to a running resident; it is not a public API and is not reachable from outside Cube.

## If you are building against Cube

There is currently no supported way to connect your own agent process to a Cube desk. The commands above only run inside a resident Cube itself started. Watch the [product overview](https://www.usecube.io/index.md) and [blog](https://www.usecube.io/blog) for whether and how an external connection returns.

[Product overview](https://www.usecube.io/index.md) · [Agent documentation index](https://www.usecube.io/llms.txt)