godot-mcp-bridge

An AI that works with you in the Godot editor — not instead of you.

Every Godot MCP server lets an AI drive the editor. This one also tells the AI what you just did — the scene you opened, the node you selected, the file you saved — so you can both work in the same project at the same time without stepping on each other. Add a real step-debugger, scope-aware refactoring through Godot's language server, live-tree edits that never clobber your unsaved work, and 231 tools that were each verified against a running editor rather than just written.

Started as a fork of tomyud1/godot-mcp (MIT licensed) and has since diverged substantially — see CHANGELOG.md.

💬 What you'd actually say to it

"The player falls through the floor sometimes. Set a breakpoint in _physics_process and tell me what velocity is when it happens."

"Build me an enemy: CharacterBody2D, circle collider, sprite, patrol script, and put it in the enemies group."

"Run the game, press jump, screenshot it, and tell me if the animation played."

"Which of my resources aren't referenced by anything anymore?"

"Export a Windows debug build and tell me when it's done."

Under the hood that's a breakpoint hit read from a paused frame, a scaffolded scene tree, a runtime input + screenshot loop, a dependency sweep, and an async headless build — but you don't have to know which tool does which.

// debug_launch({scene: "res://scenes/level.tscn"})
{ "state": "stopped", "stopped_reason": "breakpoint", "hit_breakpoint": true }

// debug_stack_trace()
{ "frames": [{ "name": "_physics_process", "line": 17,
               "source": ".../scenes/player.gd" }] }

// debug_variables({variables_reference: 1})   ← the Locals scope
{ "variables": [{ "name": "delta",     "value": "0.01666666666667" },
                { "name": "direction", "value": "<null>" }] }

// debug_evaluate({expression: "velocity"})
{ "result": "(0.0, 0.0)" }

✨ Why this one

It's bidirectional. The AI can poll get_editor_activity to see what you just did in the editor — selection, scene open/close/save, script focus, resource saves, asset reimports, undo/redo, which screen you're on — tagged human vs its own actions. It finds out you moved something without having to ask. Every other Godot MCP is one-directional: the AI drives, and is blind to you. (Checked by reading the source of 12 competing servers in July 2026, not their READMEs.)

It doesn't clobber your work. Many Godot MCP servers edit your .tscn files on disk. If you have that scene open with unsaved changes, they silently overwrite it. Here, when a scene is open, every mutating tool edits the live editor tree instead — your unsaved edits survive and every change goes through Godot's undo system (Ctrl+Z works). Closed scenes still edit on disk as usual.

The claims are tested, not asserted. 526 GDScript checks against the tool handlers, 147 Node tests for the bridge and the tool registry, and 43 that drive a real Godot editor — creating scenes, mutating one that is open, launching an actual game — on every push, on both Godot 4.5 and 4.7. Every one of the tools that writes to your project has automated coverage.

That suite is not decoration; it is where the bugs came from. It caught close_scene_tab being broken on 4.5 (the minimum this README promises), a run_scene that froze the whole editor for its entire timeout on every single call, and a res:// texture path that five different tools accepted and silently threw away. Each of those was found by a test that failed, not by reading the code — and each is in CHANGELOG.md with what it cost.

See Limitations for what it still can't do.

More of what it does:

• Step-debugger — set breakpoints, step, read the real call stack and frame variables, evaluate expressions in the paused frame, over Godot's own Debug Adapter. Stop at the failure and look at actual values instead of inferring them from print() output.
• Real headless export — builds an actual game binary via a shadow-workspace clone, asynchronously, without freezing the editor (export_project → get_export_status).
• Runs your real tests — run_gut_tests executes your GUT unit suite (sync or async) and reports pass/fail, so the AI acts on real results instead of guessing.
• Drives the running game — call methods, set properties, await signals, game_eval a snippet, record/replay input, snapshot the live tree, even a multiplayer peer-spawn harness (spawn_headless_peers) — deterministic playtesting without screenshots.
• Sandboxed paths — every path is guarded against traversal outside the project (the class of bug behind CVE-2026-15522 in another server).
• Writes the boilerplate for you — wire_signal connects a signal and generates the correctly-typed handler; generate_onready_refs emits typed @onready vars for a subtree; scaffold_entity builds a character (body + collision shape resource + sprite
  • movement script) in one call; scaffold_state_machine lays out a working FSM. Physics layers can be set by name instead of bit indices.
• Tells you what's rotting — find_unused_resources, detect_circular_dependencies, analyze_scene_complexity, and analyze_signal_flow (which catches connections whose handler doesn't exist — a bug that otherwise only shows up at runtime).
• Asks what changed, not for everything again — scene_diff takes a snapshot id and then returns only the added, removed and modified nodes (with before/after values), so the agent stops paying for a full read_scene every time it looks back. It catches your edits too, not just its own.
• Catches multiplayer bugs that fail silently — mp_diagnose flags an .rpc() call to a method with no @rpc annotation, a synchronizer replicating nothing, and a spawner whose spawn_path goes nowhere. None of those error when you write them; all of them look like "the client is broken" when a second peer joins.
• Visual regression — compare_screenshots diffs two frames and reports the changed percentage and region, so "did my change actually alter the screen?" has an answer.
• Tests your UI like a human would — click a button by its visible caption (click_control_runtime({text: "Start"}), which refuses and lists candidates if the text is ambiguous), then assert what's on screen with assert_screen_text — reading the live Control tree, so it works headless with no OCR.
• Localization that doesn't half-work — sync_localization registers the .translation files Godot generated from your CSV (the manual step that silently makes a language never load) and reports every key you haven't translated yet, per locale.
• Setup that explains itself — npx godot-mcp-bridge install installs and enables the addon in one command; doctor diagnoses a broken setup; diagnose_connection tells the AI exactly why the editor isn't connecting.
• Fast + robust — batch_execute / batch_scene_edit cut N calls to one; heavy reads (read_scene, scene_tree_dump, classdb_query) take max_depth/filter to stay token-cheap; only 37 tools load by default so the agent stays focused. Measured, not claimed — see below.
• Symbol-accurate refactoring — gd_rename and gd_references go through Godot's language server, so they understand scope: renaming a local speed won't touch an unrelated class's speed the way a text search would. gd_diagnostics surfaces type errors without running the game.
• Multiplayer scaffolding — mp_add_spawner / mp_add_synchronizer build Godot 4's replication nodes (including the SceneReplicationConfig sub-resource that makes them tedious by hand), mp_wire_rpc writes correctly-annotated @rpc methods, and mp_scaffold_lobby generates the host/join plumbing.
• C#, honestly — create_csharp_script scaffolds the boilerplate, and csharp_status tells you up front whether C# can work here at all. The standard Godot build has no C# support: a .cs file saves fine, attaches to nothing, and fails silently. Better to find that out before writing any.
• Preset toolsets at startup — GODOT_MCP_TOOLSETS=runtime,debug (or all) puts those tools in the FIRST tool list. Enabling one mid-session relies on the client re-fetching list_tools, and several clients cache it for the session; presetting sidesteps that entirely.
• Opt-in confirmation gate — set GODOT_MCP_REQUIRE_CONFIRM=true and operations with no undo path (file deletes/renames, script rewrites, mass renames, project settings) require an explicit confirm: true. Edits to an open scene are exempt — those do land on Godot's undo history, so Ctrl+Z (or undo_last) already covers them.
• Pre-flight validation — validate_scripts sweeps every .gd (loading each the way the editor does, so a script using an autoload isn't reported as broken); validate_scene_integrity flags nodes left with an empty required resource — including an instance whose script went missing, which otherwise just stops behaving with no error; validate_meshes catches empty geometry.
• Look at a scene without running it — render_scene_preview renders a 2D scene to a PNG offscreen, auto-framed on its content. Checking whether a level's platforms line up used to mean launching the game; now it doesn't, so it actually gets checked.
• Says when a write didn't land — a property can exist, accept an assignment and still hold something else (Godot clamps and coerces silently: a TextureRect asked for size.y = 6.667 keeps 16). Scene tools read back what they wrote and report the mismatch instead of reporting success.
• Wires exported node slots — set_node_reference points an @export var target: Area2D at another node, the thing you'd otherwise do by dragging in the inspector and which no value-based property tool can express.

⚙️ Running more than one project

One project on one machine needs no configuration. If you run several — or a CI editor alongside your own — set two things, because the addon dials a fixed port and cannot tell which server answered:

GODOT_MCP_PROJECT is the one worth setting even with a single project. Without it the first editor to reach the port is trusted with every tool call, so an unrelated editor that happens to be open can end up receiving edits meant for this one. With it, that connection is refused and the editor says so instead of silently taking the work.

🧮 What it costs per request

Tool definitions are sent on every request, so a large always-on surface is a standing tax on every message — the most common complaint about Godot MCP servers, and one nobody publishes a number for. Here is ours, from scripts/measure-tools.mjs so you can re-run it:

The 44 is 37 Godot tools plus seven the server answers on its own (status, guides, find_tools, the toolset controls). They are not Godot tools, but they ride in every request, so counting only the 38 published a number 1,100 tokens below what a request actually costs — the script counts them now.

So the default surface is 18.7% of the full one, and turning everything on costs roughly 41,000 extra tokens on every request. That is the reason the default is small and the rest is opt-in per toolset (or preset once via GODOT_MCP_TOOLSETS), rather than a judgement that the other 192 tools do not matter. find_tools searches all 231 by what you want to do, so a tool being unloaded never means it is unfindable. The default grew by ~800 tokens when previews (dry_run), subtree reads and the detached launch mode were added to core tools — measured rather than assumed, which is the point of publishing it.

The estimate is chars ÷ 4, which is close enough to compare sets and honest about being an estimate. The most expensive definitions inside core are modify_node_property, add_node and run_scene — verbose because they carry the "use this, NOT that" wording that stops an agent picking the wrong neighbour.

📊 How it compares

Checked in July 2026 by reading each project's source, not its marketing. Star count mostly tracks how early a project shipped, so it's listed last rather than first.

About the Node process. Godot-MCP-Native runs entirely inside the editor and sells that as "no sidecar". It is a real trade, so here is the other half of it. A server living in the editor process is bound to the editor's lifetime and to its main thread. That costs three things: it cannot be spawned over stdio, so stdio-only clients like Codex CLI cannot load it at all; it dies when Godot crashes, taking your AI client's connection with it; and any slow work freezes the editor, because @tool scripts run on the main thread.

That last one is measurable. Asked "which assets does nothing reference?" on a project with a couple of free asset packs in it (24,649 files, 12,201 images), the same analysis takes 1.6 seconds in a separate process and never touches the editor — where an in-editor implementation blocks the UI for 26 seconds. The sidecar is an install step you pay once; the main thread is one you pay every call.

The honest read: undo, a debugger, and runtime control are table stakes now — the good projects all have them. What no one else does is the bidirectional half, and the two most-starred options haven't shipped since April.

📦 Quick Start

0. Install Node.js (one-time setup)

Download and run the installer from nodejs.org (LTS version). It's a standard installer — no terminal needed.

1. Install the Godot plugin

One command, from inside your Godot project folder:

npx godot-mcp-bridge install

That copies the addon into addons/godot_mcp/ and enables the plugin in project.godot (backing the file up first, and keeping every other plugin and setting intact). Add --client claude-desktop or --client cursor and it will register the server in that client's config too.

Something not connecting? Run this and it tells you which step is missing:

npx godot-mcp-bridge doctor

2. Add the server config to your AI client

Claude Desktop — Settings → Developer → Edit Config → open the config file and paste:

Mac / Linux:

{
  "mcpServers": {
    "godot": {
      "command": "npx",
      "args": ["-y", "godot-mcp-bridge"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "godot": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "godot-mcp-bridge"]
    }
  }
}

Cursor — Settings → MCP → Add Server:

Mac / Linux:

{
  "mcpServers": {
    "godot": {
      "command": "npx",
      "args": ["-y", "godot-mcp-bridge"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "godot": {
      "command": "cmd",
      "args": ["/c", "npx", "-y", "godot-mcp-bridge"]
    }
  }
}

Claude Code — run in terminal:

claude mcp add godot -- npx -y godot-mcp-bridge

Codex CLI — run in terminal:

codex mcp add godot -- npx -y godot-mcp-bridge

Or add it to ~/.codex/config.toml by hand:

[mcp_servers.godot]
command = "npx"
args = ["-y", "godot-mcp-bridge"]

On Windows use command = "cmd" and args = ["/c", "npx", "-y", "godot-mcp-bridge"].

Codex CLI speaks stdio only — it cannot use an MCP server that is reachable solely over HTTP. This server is a stdio server, so it works there without a bearer token or a port to keep track of.

Works with any MCP-compatible client (Cline, Windsurf, etc.)

3. Restart your AI client

Close and reopen Claude Desktop / Cursor / your client so it picks up the new config.

4. Restart your Godot project

Hit Restart Project in the Godot editor. Check the top-right corner — you should see MCP Connected in green. You're ready to go.

🧰 What Can It Do?

231 Tools, 38 Loaded by Default

A big always-on tool list makes an AI agent wander between unrelated capabilities and burns context on definitions it never uses. So only core (37 tools) is visible by default — the smallest set that carries a normal session end to end: look around, edit scenes and scripts, run the game, read the errors.

Everything else is grouped by intent and is one call away: enable_toolset({ name: "runtime" }). list_toolsets shows every set, what it's for, and the names of the tools inside it — so the AI can find what it needs without loading the whole catalog first. Nothing is ever unreachable.

For a goal-to-tool index and per-topic usage guides (scene editing patterns, the runtime testing loop, asset generation, troubleshooting), see docs/TOOLS.md — the same content the server exposes live via the get_guide tool, in browsable form.

Interactive Visualizer

Run map_project and get a browser-based explorer at localhost:6510:

• Force-directed graph of all scripts and their relationships
• Click any script to see variables, functions, signals, and connections
• Edit code directly in the visualizer — changes sync to Godot in real time
• Scene view with node property editing
• Find usages before refactoring

🏗️ Architecture

┌─────────────┐    MCP (stdio)   ┌──────────────┐   WebSocket   ┌──────────────┐
│  AI Client  │◄────────────────►│  MCP Server  │◄─────────────►│ Godot Editor │
│  (Claude,   │                  │  (Node.js)   │   port 6505   │  (Plugin)    │
│   Cursor)   │                  │              │               │              │
└─────────────┘                  │  Visualizer  │               │  230 tool    │
                                 │  HTTP :6510  │               │  handlers    │
                                 └──────┬───────┘               └──────────────┘
                                        │
                                 ┌──────▼───────┐
                                 │   Browser    │
                                 │  Visualizer  │
                                 └──────────────┘

⚠️ Limitations

Written to be the section you read before hitting these, not after.

Requires a running editor. This drives a live Godot instance over a WebSocket; it is not a headless CLI. No editor open, no tools. Godot 4.5 or newer — 4.3 and 4.4 were measured against the live suite and the editor-mode scene path does not work there.

Local only, one editor at a time. Port 6505 on localhost, first come first served. Two projects open at once means the second one loses; set GODOT_MCP_PORT on both the server and the addon, or GODOT_MCP_PROJECT so the bridge refuses the wrong project instead of silently driving it.

A closed scene is edited on disk, with no undo entry. When the scene is open everything goes through Godot's undo history and Ctrl+Z works, including over a whole batch_scene_edit. When it is closed there is no history to write to — use version control. Many destructive tools take dry_run: true to preview first.

Enabling a toolset mid-session may not reach your client. Only core (37 tools) is on by default. enable_toolset flips it server-side and the server does send notifications/tools/list_changed, but several clients cache the tool list for the whole session and never re-fetch — and then the newly enabled tools stay invisible until you restart the client. If you know you want them, set GODOT_MCP_TOOLSETS=runtime,debug (or all) so they are in the first list.

game_eval runs your snippet inside the running game. Code that does not compile is caught in the editor before the game ever sees it. A snippet that fails at runtime — dereferencing a freed node, calling a method that does not exist — halts the game under the editor's attached debugger, and the call times out instead of returning an error. That is the debugger, not the bridge: launch with run_scene({attach_debugger: false}) and the same bad snippet comes back as a result in ~18ms with the connection intact (measured). The trade is stated in the tool description — no step-debugging, and get_errors loses its Debugger>Errors source.

What has been measured, and when

Numbers rather than adjectives, all from 2026-09-03 against a 24,880-file project:

Every one of those is re-runnable: mcp-server/scripts/measure-tools.mjs for the schema cost, scripts/measure-runtime-cost.gd for per-tool time and payload, and the rest are assertions in the suites.

C# is scaffolding only. create_csharp_script writes a correctly-shaped file and csharp_status tells you honestly whether this editor can run C# at all (a standard, non-Mono build cannot). The language-server and debugger tools cover GDScript, not C#. The blocker is upstream: Godot has no non-interactive way to generate the .csproj, verified across four CI runs.

The AI still does not know Godot as well as you do. It struggles with complex UI layouts, compositing scenes, and some property manipulation. It cannot build a game on its own — it debugs, writes scripts, runs and inspects the thing, and keeps you company while you do. Feedback welcome.

🔧 Development

To build from source instead of using npm:

cd mcp-server
npm install
npm run build

Then point your AI client at mcp-server/dist/index.js instead of using npx.

📖 Release notes

Narrative write-ups of each release live in release-notes/ — latest is v1.2.1. For the full change history, see CHANGELOG.md.

🤝 Contributing

See CONTRIBUTING.md. Security issues: see SECURITY.md instead of opening a public issue.

📄 License

MIT

Report Issues