# Connect an AI Assistant (MCP) > Connect Claude or any MCP client to GraphDeck. The built-in MCP server lets an AI assistant build graphs live on your canvas and deploy to Unreal Engine with your approval. GraphDeck itself has no AI of its own - you bring the assistant and the API key or subscription; GraphDeck turns its output into a graph. GraphDeck has a built-in MCP (Model Context Protocol) server. Connect an assistant like Claude and it can build graphs right on your canvas while you watch, look things up in your library, check its own work, and send finished graphs into Unreal. Available in both Free and Pro. The Connector plugin adds a second MCP surface inside the Unreal editor itself; see [MCP inside Unreal](#mcp-inside-unreal-the-connector-plugin) below. ## What it can do The server exposes 23 tools covering the full GraphDeck workflow: - Read the graph open on your canvas - Search the node library and look up node specs - Build, extend, and arrange graphs live on the canvas, and draw explanations right on it: rings, arrows, and step-by-step walkthroughs anchored to real nodes - Validate its own work with GraphDeck's validation rules - List, search, and open graphs from your library - Deploy to a connected Unreal Editor and read graphs back from it - Reference knowledge: build formats, component specs, and guides ## Turning it on Open **Settings, MCP Server** and switch it on. The server is off by default and runs locally at `http://127.0.0.1:30303/mcp`. Nothing leaves your machine except what the connected assistant sends to its own AI provider. The little square light in the title bar turns green while an assistant is working. Click it anytime to reopen the setup window. ## Setup: Claude Code 1. In GraphDeck, open the MCP setup window (Settings, MCP Server) and pick **Claude Code** 2. Choose the folder you launch Claude Code from (your game project, not GraphDeck) 3. GraphDeck writes the `.mcp.json` entry for you. Start a new chat and ask Claude to build something in GraphDeck Manual alternative: add GraphDeck as an HTTP MCP server, for example: ``` claude mcp add --transport http graphdeck http://127.0.0.1:30303/mcp ``` ## Setup: Claude Desktop 1. In the MCP setup window, pick **Claude Desktop app** 2. Claude Desktop opens and asks to install the GraphDeck extension. Approve it 3. Ask Claude to build something in GraphDeck ## Setup: any other MCP client Cursor, VS Code, Codex, and any other MCP-capable app can connect. Add GraphDeck as an HTTP MCP server at `http://127.0.0.1:30303/mcp`. GraphDeck must be running with the MCP server enabled. ## MCP inside Unreal: the Connector plugin The [Connector plugin](https://graphdeck.app/docs/connector-plugin) has its own MCP tools inside the Unreal editor, covering the Unreal half of the workflow. An assistant connected to your editor can: - Send the Blueprint you have open (or any asset path) to GraphDeck, same pipeline as the toolbar button - Send a Material the same way, with optional texture bundling - Compile a Blueprint and read back its errors and warnings, so it can prove a deploy actually works - Check that asset paths exist without loading anything - Read the connector's status Combined with GraphDeck's own MCP server, this closes the loop: an assistant can build a graph on your canvas, deploy it to Unreal, then compile the result and fix what's broken. How to connect depends on your engine version: - **UE 5.8+:** the tools register with the editor's built-in MCP server (Epic's Experimental Model Context Protocol plugins) at `http://127.0.0.1:8000/mcp`, alongside Epic's own toolsets - **UE 5.4 to 5.7** (or 5.8 without the MCP plugins): the connector serves the same tools from its own endpoint at `http://127.0.0.1:30010/graphdeck/mcp` The GraphDeck status panel (Window, GraphDeck) shows which mode is active and the exact URL to connect to, with an on/off toggle under **AI Agent Tools (MCP)**. Switching it off removes the tools entirely; the toolbar buttons and the GraphDeck app connection are unaffected. ## Staying in control - **Off by default.** The server only runs if you switch it on - **Consent dialog.** The first tool call of a session asks "Allow an AI assistant to control GraphDeck?" before anything happens. Deny is remembered for the session - **Tool picker.** Settings, MCP Server lists capability groups: read graph, build and modify canvas, library access, library organisation, Unreal Engine, and knowledge. Enable or disable each. Screenshots are off by default - **You approve sends.** Deploying to Unreal Engine and reorganising your library always ask for your confirmation first - **Undo.** Ctrl+Z reverts anything the assistant builds ## Privacy Graphs and library entries the assistant reads are shared with it, which means they also reach the company running that AI service. If a graph shouldn't leave your machine, don't point an assistant at it, or restrict the capability groups it can use. ## Free vs Pro The MCP server ships in both tiers, and an assistant working in Free shares the same free limits you do: 5 kept graphs per tool and 10 AI builds a day. The one exception is library organisation (folders, renames, tags), which is Pro in the app too. [Prev: Connector Plugin](https://graphdeck.app/docs/connector-plugin) · [Next: Privacy & Data Settings](https://graphdeck.app/docs/privacy-data) --- [Home](https://graphdeck.app/) · [Docs](https://graphdeck.app/docs/) · [Pricing](https://graphdeck.app/pricing)