Godot MCP

godot-mcp

[npm version] [CI] [Godot 4.5+] [Node 20+] [License: MIT]

Give your AI assistant eyes and hands in the Godot editor — and a running game it can actually playtest.

Why this one?

There are a few Godot MCP servers out there, and most can open a scene and poke at nodes. This one is built around a harder problem: letting an agent verify its own work. Run the game, drive it like a player, observe what actually happened, and prove the change did what it claimed — without you ferrying screenshots and error logs back and forth.

The pieces that make that possible, and that you won't find elsewhere:

• Deterministic playtesting. Freeze the game clock, step exact slices of game time (or step until a condition holds), with inputs riding inside the window. Observation never races the game.
• Cheap observation. Live entity state — positions, velocities, animation state, your own custom data — as structured JSON. Most "what's happening on screen?" questions get answered without spending vision tokens on a screenshot.
• Real input. Named actions with analog strength, joypad buttons and stick vectors, raw keys with modifier combos, relative mouse-look, text typing. Sequences run with precise timing and can prove they changed game state.
• Scenario setup. Run GDScript inside the running game: grant the weapon, skip to wave 3, spawn a test bot — no debug hooks baked into your game code.

Here's what that looks like when an agent tests a boss fight:

godot_editor_edit   run frozen=true           # boot with game time frozen at frame 0godot_exec          GameState.wave = 3        # set up the scenario worth testinggodot_game_time     step_until "tree.get_nodes_in_group('boss').size() >= 1"godot_runtime_state digest                    # exact positions and state — no pixels, no guessinggodot_game_time     step 500ms + dodge input  # play the moment that mattersgodot_editor_read   screenshot_game           # and a screenshot when it's actually worth the tokens

Less copy-paste, more creating.

Quick Start

1. Configure your AI assistant

Add godot-mcp to your MCP client. See the Installation Guide for config examples (Claude Desktop, Claude Code, VSCode/Copilot, and more). The short version:

{  "mcpServers": {    "godot-mcp": {      "command": "npx",      "args": ["-y", "@satelliteoflove/godot-mcp"]    }  }}

2. Install the Godot addon

npx @satelliteoflove/godot-mcp --install-addon /path/to/your/godot/project

Enable it in Godot: Project > Project Settings > Plugins > Godot MCP

3. Start building

Open your Godot project, restart your AI assistant, and start building. If anything refuses to connect, the Troubleshooting Guide has you covered.

What's in the box

21 tools, 86 actions. Full API docs in the Tools Reference.

A note on shape: tools split along the read/write boundary, so every godot_*_read tool (and the other read-only tools) can be safely auto-allowed in your client's permission settings while writes stay gated. Related operations still live as actions inside one tool, so your agent's context isn't flooded with definitions it won't use. Anything an agent can do by editing project files directly — creating scenes and nodes, attaching scripts, connecting signals — is deliberately not duplicated as a tool; the bridge covers what files can't: editor state, verification, binary-encoded cell data, and the running game. A --read-only flag serves the look-but-don't-touch use case.

Things to ask for

A feel for what an agent can do with this, in plain prompts:

• "Run the game and screenshot the title screen. Does the layout survive 1280x720?"
• "The player can clip through the east wall. Reproduce it, then check the collision shapes and tell me why."
• "Step the game until the second wave spawns and give me every enemy's position and velocity."
• "Hold the left stick at half deflection for two seconds — does the walk animation blend correctly?"
• "Profile ten seconds of gameplay and find what's causing the frame spikes."
• "Add a hit-flash animation to the Player: modulate to red and back over 0.2 seconds."

For richer runtime observation, add key nodes to the mcp_watch group or give them a _mcp_state() method — see the Runtime State Guide.

How it works

┌─────────────────┐   stdio    ┌─────────────────┐  WebSocket   ┌──────────────────┐  debugger   ┌──────────────┐│   MCP client    │ ◄────────► │  godot-mcp      │ ◄──────────► │  Bridge addon    │ ◄─────────► │ Running game ││ (Claude, etc.)  │            │  server (Node)  │   :6550      │  (Godot editor)  │    wire     │  (autoload)  │└─────────────────┘            └─────────────────┘              └──────────────────┘             └──────────────┘

The server talks to an editor addon over a local WebSocket; the addon reaches into the running game over Godot's own debugger protocol, so the game process needs no extra ports or setup. The addon binds to 127.0.0.1 by default. WSL2 is fully supported (auto-detection, host IP discovery, configurable bind modes) — see the Installation Guide.

Curious about connection lifecycles, the single-client policy, or how frozen-time stepping works under the hood? The Architecture Guide goes deep.

Works well with minimal-godot-mcp

minimal-godot-mcp by @ryanmazzolini covers exactly what this server doesn't: LSP diagnostics for fast static GDScript checking, and the running game's console output over DAP. This server covers everything that needs an editor bridge. No overlap, no conflict — run both, and your agent gets static analysis and the game console alongside full editor and runtime control.

One caveat: a Godot editor serves a single godot-mcp client at a time. Extra clients (a subagent inheriting your MCP config, say) wait their turn rather than hijacking your session — details in Troubleshooting.

Documentation

• Installation Guide — MCP client configs for Claude Desktop, Claude Code, VSCode/Copilot, and more
• Migrating to v4 — renamed tools, removed actions, allowlist updates
• Troubleshooting — connection checklist, CLI smoke test, common fixes
• Claude Code Setup Guide — CLAUDE.md template for Godot projects
• Runtime State Guide — expose game state to agents via mcp_watch and _mcp_state()
• Tools Reference — all 21 tools with full API docs
• Architecture Guide — how the server, addon, and game bridge fit together
• Contributing — dev setup, adding tools, release process
• Changelog — release history

Requirements

• Godot 4.5+ (the addon uses the Logger class introduced in 4.5)
• Node.js 20+
• Any MCP client that speaks stdio

Development

cd servernpm install && npm run buildnpm test                # unit + schema-snapshot testsnpm run test:protocol   # wire-level smoke of the built servernpm run generate-docs

There is also an agentic eval harness (npm run eval) that runs realistic tasks through headless Claude Code against a real project — see server/evals/README.md. It never starts Godot; you bring the open editor.

Contributions welcome — this project favors tools that solve real, time-wasting problems. Read CONTRIBUTING.md before building something big, or open an issue and we'll figure out the right shape together.

License

MIT