MCP automation loop with 167 tested tools for Godot 4
Valid MCP server (1 strong, 1 medium validity signals). No known CVEs in dependencies. Package registry verified. Imported from the Official MCP Registry.
5 files analyzed · 1 issue found
Security scores are indicators to help you make informed decisions, not guarantees. Always review permissions before connecting any MCP server.
Add this to your MCP configuration file:
{
"mcpServers": {
"io-github-beremaran-godot-agent-loop": {
"args": [
"-y",
"@beremaran/godot-agent-loop"
],
"command": "npx"
}
}
}From the project's GitHub README.
Build it. Play it. Prove it.
An MCP automation loop for Godot 4.
Other integrations give agents tools. Godot Agent Loop gives them a tested feedback loop to author, run, observe, playtest, and independently verify Godot games:
author → validate → run → observe → playtest → verify → refine
Watch the 65-second demo · Read the exact run evidence · Inspect the resulting project
claude mcp add godot-agent-loop -- npx -y @beremaran/godot-agent-loop
Then point the agent at a project directory—or an empty directory—and describe
the playable result. The compact default surface exposes 39 tools for the loop
above; the godot_tools meta-tool searches, describes, and dispatches the full
167-tool catalog on demand. Runtime and editor bridges are installed
transiently and cleaned up automatically.
Using Cline, Cursor, or another MCP client? See Configuration.
Support is deliberately bounded: Godot 4.4 is the compatibility floor and 4.7 the primary target; editor UI/rendering depth is verified on Linux, while Windows and macOS receive the documented portable acceptance path. Full debugger automation, native extension builds, and unbounded engine control are not claimed. Details in the verified support boundary.
validate_script,
validate_scripts), test runners for native/GUT/GdUnit4
(run_project_tests), bounded runtime evidence (verify_project), export
checks (verify_export_readiness), and static integrity analysis
(analyze_project_integrity).game_eval executes GDScript with await support (privileged, opt-in).editor_control applies reversible edits through
EditorUndoRedoManager, with a live Agent Activity dock and a human
Pause Agent lock.Godot.NET.Sdk matched
to your installed Godot, generate idiomatic scripts, and restore/build/run
via verify_dotnet_project.The full inventory of 167 tools — runtime interaction, scene authoring, project management, verification, 2D/3D rendering, audio, UI, networking, and more — lives in docs/tools.md. Per-tool verification status and test references are in the generated coverage report.
create_project's dotnet: true
flag or create_csharp_scriptDevelopment targets the latest stable Godot release. The project also keeps a tested compatibility floor while the same implementation remains cleanly portable; currently, CI covers Godot 4.4 and 4.7. The floor may be raised when it blocks useful features or creates meaningful maintenance cost. In that case, the last compatible release remains available, and an older-version maintenance branch will be created only when user demand justifies maintaining it. Such a branch would receive critical fixes rather than new features.
| Area | Status | Evidence or limitation |
|---|---|---|
| Linux headed (desktop or Xvfb), Godot 4.4 and 4.7 | Verified in CI | Full MCP E2E under Xvfb, direct runtime, subprocess operations, and strict script parsing |
| GDScript project and running-game workflows | Verified for advertised tools | See the generated coverage report |
| Privileged runtime commands | Opt-in only | Disabled by default; intended for trusted localhost development |
| Godot .NET/C# | Scaffold, compile, and editor-load verification | Godot .NET 4.4 and 4.7 with .NET SDK 8 |
| Linux exports | Release/debug template export and smoke-run verification | Godot 4.4 and 4.7 installed templates; other targets are not claimed |
| Rendering and screenshots | A headed rendering context is required | Compatibility and Forward+ on Linux software rendering; display-less sessions fail fast with desktop/Xvfb remediation |
| Windows and macOS | Portable acceptance verified | Godot 4.7 process, Unicode path, runtime input, window query, and teardown workflows |
| Windows/macOS editor UI, rendering, and exports | Not claimed | Verified on Linux only; Windows/macOS support is bounded to the portable acceptance suite |
| Editor state and undo/redo bridge | Verified in MCP E2E | The headed editor bridge is authenticated and uses EditorInterface plus EditorUndoRedoManager |
| Full debugger control | Not claimed | Breakpoints, stack inspection, and frame-local evaluation remain outside the supported boundary |
| Profiler, leak, asset, localization, and accessibility audits | Verified in MCP E2E | game_performance and analyze_project_integrity return bounded live/static evidence; native extension builds remain unsupported |
| GDExtension builds | Not claimed | analyze_project_integrity inspects declarations and libraries without invoking arbitrary native toolchains |
The quickstart npx command is all most setups need. The
sections below cover other clients and a source checkout.
The repository ships one neutral bundle that starts the matching npm MCP server and provides the same build, debug, verify, and ship skills to Claude Code, Codex, OpenCode, and Pi. For Claude Code:
/plugin marketplace add beremaran/godot-agent-loop
/plugin install godot-agent-loop@godot-agent-loop
For a local checkout, use claude --plugin-dir ./agent-plugin. See the
portable agent bundle guide for verified Claude Code,
Codex, OpenCode, and Pi install paths.
Add to your Claude Code MCP settings:
{
"mcpServers": {
"godot": {
"command": "npx",
"args": ["-y", "@beremaran/godot-agent-loop"],
"env": {
"GODOT_PATH": "/path/to/godot",
"DEBUG": "true"
}
}
}
}
Add to your Cline MCP settings (cline_mcp_settings.json):
{
"mcpServers": {
"godot": {
"command": "npx",
"args": ["-y", "@beremaran/godot-agent-loop"],
"disabled": false
}
}
}
Create .cursor/mcp.json in your project:
{
"mcpServers": {
"godot": {
"command": "npx",
"args": ["-y", "@beremaran/godot-agent-loop"]
}
}
}
For a source checkout, use node as the executable and pass the built server
path as a separate argument:
{
"command": "node",
"args": ["/absolute/path/to/godot-agent-loop/build/index.js"]
}
git clone https://github.com/beremaran/godot-agent-loop.git
cd godot-agent-loop
npm install
npm run build
No setup is required when the game is started through run_project: the server
installs the interaction autoload automatically by generating an override.cfg
(which Godot merges over project.godot at startup) and copying the runtime
scripts into the project, then removes them again on stop_project, game exit,
or server shutdown. project.godot is never modified. If an earlier server
crashed or was killed before cleaning up, the next server detects and removes
the leftover files on first contact with the project; an installation you
manage yourself (declared in project.godot) is never touched.
To run the interaction server without run_project, copy
build/scripts/mcp_interaction_server.gd to your project and register it as an
autoload:
build/scripts/mcp_interaction_server.gd to your project's scripts folderMcpInteractionServerThe server listens on 127.0.0.1:9090. Each MCP server launch generates a
cryptographic runtime secret, passes it only to the Godot child process, and
authenticates it during capability negotiation before any runtime command is
accepted. A manually managed runtime should set the same
GODOT_MCP_RUNTIME_SECRET value in both processes; leaving it unset retains
legacy unauthenticated behavior and is suitable only for a trusted machine.
Commands that execute arbitrary GDScript, invoke arbitrary node properties or
methods, mutate scripts, call multiplayer peers, or make HTTP/WebSocket
connections remain disabled by default even after authentication. Grant only
the required group with GODOT_MCP_PRIVILEGED_GROUPS: reflection enables
arbitrary property/method access, code-execution enables eval/script control,
and network enables RPC, HTTP, and WebSocket. The legacy
GODOT_MCP_ALLOW_PRIVILEGED_COMMANDS=true grants all three groups. Use either
only for a trusted local developer workflow. Authentication and policy denials
never echo secrets, source, property values, URLs, headers, or engine errors.
Authentication success/failure emits a structured audit event containing only
the event name, runtime component, numeric session ID, and timestamp.
| Variable | Description |
|---|---|
GODOT_PATH | Path to the Godot executable (overrides auto-detection) |
DEBUG | Set to "true" for detailed server-side logging. This also runs the headless operations script with --debug-godot, which logs diagnostics and writes a temporary write-access probe file into the project (removed again on every branch). Parameter values are summarized by type and size in both logs, never printed. |
GODOT_MCP_ALLOWED_DIRS | Optional. Restrict run_project to projects under these roots (;, ,, or : separated). When unset, any project path is allowed. |
GODOT_MCP_RUNTIME_SECRET | Optional explicit shared runtime secret. The MCP server generates a fresh 256-bit value when omitted and passes it only to Godot processes it launches. Set the same value manually only when connecting to a separately launched runtime. |
GODOT_MCP_EDITOR_START_PAUSED | Optional, default false. Start the editor addon's cooperative lock in human-editing mode so mutating MCP tools are refused until Resume Agent is pressed. |
GODOT_MCP_TOOL_SURFACE | Optional, default core. Set to full to advertise the complete static tool catalog instead of the compact 39-tool core, which includes godot_tools discovery/dispatch. |
GODOT_MCP_PRIVILEGED_GROUPS | Optional comma-separated least-privilege grants: reflection, code-execution, and/or network. All are denied by default. |
GODOT_MCP_ALLOW_PRIVILEGED_COMMANDS | Optional, default false. Explicitly enable runtime eval, arbitrary property/method access, script control, RPC, HTTP, and WebSocket commands for a trusted localhost developer workflow. |
With DEBUG=true, the MCP server emits JSON request lifecycle events to
stderr. The Godot runtime emits matching events to its captured stdout. Both
use an internal mcp_<number> correlation ID and controlled event fields;
parameters, response values, secrets, source, URLs, and malformed payloads are
never copied into logs. Runtime process output is capped at the latest 1,000
stdout and stderr lines. Stable JSON-RPC error codes remain the authoritative
machine-readable failure classification.
Large responses are bounded rather than allowed to grow with project size.
list_project_files returns deterministic cursor pages of at most 1,000 files;
game_get_scene_tree returns deterministic pre-order trees of 1,000 nodes by
default (configurable up to 10,000) and reports truncation. game_get_logs and
game_get_errors return at most 1,000 unread lines per call with hasMore and
remaining, while retaining the latest 1,000 lines per stream. Runtime JSON
responses are capped at 8 MiB, screenshots additionally enforce pixel and
6 MiB PNG limits, and short-lived subprocess/import commands cap captured output
at 16 MiB. Limit failures are explicit; callers can narrow resource/import
queries instead of receiving partial unlabelled data.
The server uses three bounded execution paths:
Persistent authoring session - The primary scene/resource authoring path. It owns a headed, deterministic Godot main loop and serves authenticated JSON-RPC commands without paying engine startup cost per edit.
Running-game socket - run_project launches the user's game headed and
injects the authenticated mcp_interaction_server.gd autoload through
override.cfg for high-fidelity runtime interaction.
One-shot subprocess fallback - Isolated authoring, validation, import,
and export operations may invoke Godot once and exit. Commands that do not
render may internally use --headless; this is not a supported display-less
agent-loop tier. Authoring sessions, running games, screenshots, and visual
verification require a desktop display, Xvfb, or another reachable rendering
context and fail fast when none exists.
| Path | Description |
|---|---|
src/index.ts | MCP server entry point |
src/tool-definitions.ts | Tool names and JSON schemas |
src/tool-manifest.ts | Per-tool domain, backend, and action declarations |
src/tool-handlers/ | Lifecycle, project, and game handler implementations |
src/scripts/godot_operations.gd | Persistent and one-shot GDScript operations runner |
src/scripts/mcp_interaction_server.gd | TCP interaction server autoload |
tests/ | Vitest unit, E2E, and Godot suites |
The project uses Vitest plus direct Godot and full MCP-to-Godot suites. The source-derived tool, action, command, and suite inventory is published in the coverage report.
npm run check # TypeScript tests, lint, build, and coverage drift
npm run test:e2e # built MCP server through a real client and Godot
npm run test:golden-agent # cold-agent game build acceptance gate
npm run test:godot # strict parsing, subprocess operations, runtime protocol
npm run test:watch # watch mode
"Run my Godot project and check for errors"
"Create a new Godot project called 'MyGame' and write a player script"
"Read the test_level.tscn scene and show me the node tree"
"Check all my changed GDScript files for syntax errors before I run the game"
"Hold down the W key for 2 seconds to test walking"
"Pause the game and take a screenshot"
"Get performance metrics - what's my FPS and draw call count?"
"Set the player's health to 100"
"Connect the enemy's 'died' signal to the game manager's 'on_enemy_died' method"
"Create a new C# (.NET) Godot project and add a CharacterBody2D script"
This project is licensed under the MIT License - see the LICENSE file for details.
Be the first to review this server!
by Modelcontextprotocol · Developer Tools
Read, search, and manipulate Git repositories programmatically
by Toleno · Developer Tools
Toleno Network MCP Server — Manage your Toleno mining account with Claude AI using natural language.
by mcp-marketplace · Developer Tools
Create, build, and publish Python MCP servers to PyPI — conversationally.