How agents work with genMC

genMC is built so that a person and an AI agent can work on the same skin or build at the same time. The agent is your agent – Claude, ChatGPT, Codex, Claude Code, Cursor or another MCP client. genMC runs no AI of its own and never needs your API key.

What an agent can do#

Through genMC's tools an agent can, depending on the studio:

• read your skin or build, your marks, requests and photos,
• paint pixels, place and replace blocks or generate whole structures with a construction program,
• render images from any angle to check its own work,
• organise a build into named parts, plan a room tour and a spawn point,
• navigate the timeline, export files and tell you what it is doing in the agent column,
• wait for your next request and continue.

Every studio registers only its own tools; the Agent tool reference lists them all. The agent can also read this knowledge base with the tool genmc_docs.

Two ways to connect#

Site tools in the browser (WebMCP). Some agents browse the web themselves and can use tools a page offers – for example the built-in browser of the Codex app. Open the studio there and the agent sees genMC's tools directly. No code needed. The start page and this knowledge base offer a few tools too: genmc_about explains genMC, genmc_docs searches these articles and genmc_open opens a studio, the gallery or an article.

The MCP relay (everything else). Agents outside the browser – Claude, ChatGPT, Claude Code, Codex CLI, Cursor – add genMC once as an MCP server (https://<this site>/api/mcp). Then you give the agent the pairing code shown in the studio's agent column, and its calls run in exactly that tab. See Connect an agent.

Either way, the agent works in the tab you have open, and you see every change live.

You stay in control#

• Every change is undoable. Agent edits are timeline steps like yours (skins, builds).
• Conflicts are refused, not overwritten. Each agent edit carries the revision it was based on. If you changed something in between, the edit is rejected and the agent must look again.
• You decide when it's done. Finish stops all agent edits until you send a new request (Requests, marks and Finish).
• Publishing needs you. An agent can prepare a gallery publication, but only your click in the dialog publishes.
• Your words are data to it. Gallery names and texts from other people reach the agent as data, never as instructions.
• Disconnect anytime under Connect agent in the agent column.

What the page cannot do#

A web page cannot start, wake or keep an agent running. If your agent stopped (its app was closed, it hit a limit or an error), your request waits in the queue; copy the wake-up prompt from the notice and paste it into the agent's chat.