by hi-godot

Godot AI review, pricing and limits

Lets agents build scenes, edit scripts and run games in a live Godot editor

Best for
Indie Godot developer using a terminal coding agent
Not for
Teams pinned to Godot 3 or early Godot 4 releases
  • Developer
  • App connectors

Should you use Godot AI?

Use Godot AI if you already code with a terminal agent and want it to see and change your open Godot project instead of guessing from pasted files. It replaces copy and paste between chat and editor. The catch: it needs Godot 4.7 or later, and C# gets file writes but no compiler feedback.

An agent can write a script, run the game and read the resulting error without the developer copying anything between windows.

Its tool reference, read 15 November 2026, lists tools that act on the open editor and the running game, while the Coding-Solo godot-mcp README describes launching the editor, running projects and capturing debug output.

Also for

  • Indie developers who use Claude Code, Codex or Cursor on a Godot 4.7 project
  • Small game teams that want an agent to build scenes and run tests in the editor
  • Godot tool makers who want to expose their own add-on to agents

Main trade-off

Tied to Godot 4.7 and later, with C# limited to writing text files.

Godot AI pricing

Free and open source under the MIT licence, with no paid tier. You still pay for the agent that calls it, through your Claude, ChatGPT or Cursor plan or your own API key.

Install Godot AI

Before you start

  • Godot 4.7 or later, with the Godot AI add-on installed from the AssetLib tab or a GitHub release into addons/godot_ai/
  • uv installed once (it runs the Python server)
  • The add-on enabled under Project, Project Settings, Plugins
  • An MCP client such as Claude Code, Codex or Cursor, configured from the Godot AI dock with Configure
  • Claude Code only: a one-time approval from claude run inside the project

How to invoke it

Exposed as 46 MCP tools (editor_state, node_create, script_patch, project_run, logs_read and 41 more) that the agent calls by name once the Godot AI dock has configured the client.

Works with

  • Claude Code
  • Claude Desktop
  • OpenAI Codex
  • Cursor
  • VS Code
  • Gemini CLI
  • Hermes Agent
  • any MCP client over stdio

Required runtime

  • Godot 4.7 or later
  • uv (runs the Python server)
  • Python 3.11 to 3.14 (fetched by uv)

Where it runs

  • Desktop
  • CLI

Data it can reach

  • Godot scenes and nodes
  • GDScript and C# files inside the project
  • editor, game and plugin logs
  • editor and game screenshots
  • project settings and input map

The permissions this connector asks for, as its own documentation states them.

What Godot AI can do

10 actions an agent can call. Open one for its parameters and an example.

editor_stateReports the editor version, project name, current scene, readiness, play state and whether the game is live.

Example call

editor_state()

Parameters

ParameterTypeNeededDescription
session_idstringOptionalPin the call to one connected editor when several are open.

Returns

FieldTypeDescription
game_statusobjectLiveness of the running game: live, not_live, no_helper, launching, break or stopped.
is_playingbooleanRaw editor play state, kept for compatibility.
scene_get_hierarchyWalks the open scene tree and returns a paginated flat list of nodes with name, type, path and child count.

Example call

scene_get_hierarchy(depth=3, limit=50)

Parameters

ParameterTypeNeededDescription
depthnumberdefault 10OptionalHow many levels of the tree to walk.
offsetnumberdefault 0OptionalWhere to start the page.
limitnumberdefault 100OptionalHow many nodes to return in one page.

Returns

FieldTypeDescription
nodesarrayOne entry per node with name, type, path and child count.
node_get_propertiesReturns the editor-visible properties of one node, or only the fields you name.

Example call

node_get_properties(path="/Main/Player", fields=["position", "velocity"])

Parameters

ParameterTypeNeededDescription
pathstringRequiredScene path of the node.
fieldsarrayOptionalProperty names to return instead of the full set of 50 to 150 entries.

Returns

FieldTypeDescription
propertiesobjectProperty names and their current values.
total_countnumberHow many editor-visible properties the node has.
node_createCreates a node of a given type under a parent, or instantiates a saved scene there.

Example call

node_create(type="CharacterBody2D", name="Player", parent_path="/Main")

Parameters

ParameterTypeNeededDescription
typestringOptionalGodot class of the new node; ignored when scene_path is given.
namestringOptionalName of the new node.
parent_pathstringOptionalScene path of the parent node.
scene_pathstringOptionalA PackedScene to instantiate instead of a bare node.

Returns

FieldTypeDescription
pathstringScene path of the created node.
node_set_propertySets one property on a node, with the change recorded in the editor's undo history.

Example call

node_set_property(path="/Main/Player", property="position", value={"x": 120, "y": 64})

Parameters

ParameterTypeNeededDescription
pathstringRequiredScene path of the node to change.
propertystringRequiredExact property name, best confirmed with node_get_properties first.
valuestringRequiredNew value: a string, number, boolean, object or list.

Returns

FieldTypeDescription
undoablebooleanWhether the change can be undone in the editor.
script_createWrites a script file, registers it with the editor and parse-checks GDScript before import.

Example call

script_create(path="res://player.gd", content="extends CharacterBody2D\n\n@export var speed := 200.0\n")

Parameters

ParameterTypeNeededDescription
pathstringRequiredres:// path of the script; overwrites an existing file.
contentstringOptionalFull source of the script.

Returns

FieldTypeDescription
diagnosticsarrayParse errors and warnings for the written file.
diagnostics_statusstringchecked, or partial if the validation log overflowed.
script_patchReplaces an exact block of text in a script and re-validates it, refreshing a loaded GDScript in place.

Example call

script_patch(path="res://player.gd", old_text="yield(get_tree(), \"idle_frame\")", new_text="await get_tree().process_frame")

Parameters

ParameterTypeNeededDescription
pathstringRequiredres:// path of the file to edit.
old_textstringRequiredExact text to find; whitespace counts.
new_textstringRequiredReplacement text.
replace_allbooleandefault falseOptionalReplace every match instead of failing on more than one.

Returns

FieldTypeDescription
diagnosticsarrayErrors left after the edit, if any.
project_runPlays the main scene, the current scene or a named scene, then waits briefly for the game to come up.

Example call

project_run(mode="current")

Parameters

ParameterTypeNeededDescription
modestringdefault mainOptionalmain, current, or custom (needs scene).
scenestringOptionalScene to run when mode is custom.
autosavebooleandefault trueOptionalSave edits made through the agent before playing.

Returns

FieldTypeDescription
game_statusobjectWhether the game came up, plus any recent errors found while waiting.
recent_errorsarrayErrors seen during launch, naming the failing script on a parse break.
logs_readReads the plugin, game, editor or combined log, optionally with the error code, source location and stack.

Example call

logs_read(source="editor", include_details=True)

Parameters

ParameterTypeNeededDescription
sourcestringdefault pluginOptionalplugin, game, editor or all.
countnumberdefault 50OptionalHow many lines to return.
include_detailsbooleandefault falseOptionalAdd error type, source location and stack context.
since_run_idstringOptionalRead the game log of an earlier run.
since_cursornumberOptionalReturn only editor entries added after this cursor.

Returns

FieldTypeDescription
linesarrayLog entries for the chosen source.
next_cursornumberCursor to pass on the next editor-log poll.
editor_screenshotCaptures the editor's 3D or 2D viewport, a Camera3D render, or the running game's frame.

Example call

editor_screenshot(source="game", max_resolution=640)

Parameters

ParameterTypeNeededDescription
sourcestringdefault viewportOptionalviewport, viewport_2d, cinematic or game.
max_resolutionnumberdefault 640OptionalLongest side of the returned image in pixels.
save_tostringOptionalAlso save the image to this path.

Returns

FieldTypeDescription
imagestringThe captured image, when include_image is true.

What Godot AI is for

Godot AI is a free, MIT-licensed Model Context Protocol server and Godot editor add-on from the hi-godot project. It gives coding agents such as Claude Code, Codex and Cursor 46 tools to build scenes, write and parse-check GDScript, run the game and read its logs, and it needs Godot 4.7 or later.

  • 46 tools over one MCP connection

    19 named verbs plus one rolled-up manage tool per domain, which keeps the count under the 100-tool caps some clients enforce while exposing over 120 operations.

  • Parse-checked script writes

    Every GDScript file written through script_create or script_patch is validated before import, and the errors come back in that same response.

  • Play and read the game

    project_run starts the game and reports whether it came up, and on a parse break it names the failing script from the debugger.

  • Screenshots of the editor and the game

    The agent can look at what it built, from either editor view or a frame of the game in play, at 640 pixels on the long side unless asked otherwise.

Show all 6 features
  • One-click client setup

    The dock configures 26 MCP clients, including Claude Code, Codex and Cursor, and pins the matching version and ports.

  • Signed, confined updates

    The in-editor updater verifies a signature on every file, and writes are kept inside the project with project.godot and .godot/ off-limits.

What people use it for

  • Build a scene from a description

    An agent creates the nodes, sets their properties, wires the signals and saves the scene without you placing anything by hand.

  • Write a script and catch the error at once

    script_create and script_patch parse-check every GDScript file they write and return the error in the same response, so the agent fixes it before you run the game.

  • Run the game and debug from the log

    project_run presses play and waits for the game to come up, then logs_read hands the agent the runtime errors with file and stack details.

  • Check a UI change visually

    editor_screenshot captures the 2D or 3D viewport or the running game's frame so the agent can compare what it built with what you asked for.

Strengths and limits

Strengths

  • It closes the loop that plain chat lacks: the agent writes a script, gets the parse error in the same response, runs the game and reads the runtime log.
  • Setup is a dock button per supported client, so connecting Claude Code, Codex or Cursor takes no hand-written JSON.
  • The project is active, with four releases between 3 October and 3 November 2026 and a commit on 14 November 2026.

Limits

  • It needs Godot 4.7 or later, so teams on Godot 4.4 or earlier, or on Godot 3, must upgrade first.
  • C# is text only: the agent can write .cs files, but Godot AI does not build .NET or show compiler errors, which leaves C# teams debugging by hand.
  • Windows users can hit a server that dies at start-up: GitHub issue #1165, opened 17 October 2026 and still open when read on 15 November 2026, reports it on hosts without PyYAML wheels.

Godot AI alternatives

The nearest AI agent skills on Hokai for the same job.

SkillWorks withConnects byChecked
Godot AI This pageClaude · Codex · Cursor · Gemini +1MCP15 Nov
AisyncAdvisory file claims and a shared journal for your AI coding agents.Claude · Codex · CursorSkill file11 Nov
CloakroomMasks secrets and personal data before your coding agent sees them.Claude · Codex · Cursor · GeminiSkill file11 Nov
FactIQGives an agent read-only SQL over 100M+ official economic and financial series.Claude · ChatGPT · CodexMCP21 Oct

Godot AI facts

Protocol
MCP
Authentication
Token
Latency class
Fast
Open source
Yes
Rate limits
None documented. All traffic stays on the local machine (server on port 8000, editor WebSocket on port 9500).

What changed

  • Release Godot AI 4.4.1 (v4.4.1) source
  • Release Godot AI 4.4.0 (v4.4.0) source
  • Security Godot AI 4.3.1 (v4.3.1) source

Questions about Godot AI

How much does Godot AI cost?

Godot AI itself costs nothing: the add-on and its Python server are MIT-licensed with no paid tier. What costs money is the agent that calls it. Claude Code, Codex or Cursor bill through their own plans or your API key, and every tool call and screenshot uses that allowance.

Is Godot AI open source?

Yes, it is published on GitHub under the MIT licence, and releases are signed so the in-editor updater can check them. You run it on your own machine: the server and editor talk only over local connections. Anonymous telemetry is on by default and can be turned off with GODOT_AI_DISABLE_TELEMETRY=true or in the editor.

Which AI agents work with Godot AI?

Any client that speaks the Model Context Protocol over standard input and output. The dock sets up 26 of them directly, among them Gemini CLI, Hermes Agent and the three big coding agents. For a client it does not list, the dock prints a command to run by hand.

Does Godot AI work with Godot 4.4 or Godot 3?

No. The README asks for Godot 4.7 or later, so older projects need an upgrade first. Godot 3 projects also need Godot's own 3-to-4 converter run before that, since GDScript syntax changed between the two.

Can Godot AI fix C# errors in a Godot project?

Not by itself. It can create, patch and outline .cs files, but it does not build .NET or report compiler errors, and attaching a .cs script needs a .NET-enabled editor build. GDScript is the language it fully validates.

About Godot AI

Godot AI is an open-source add-on and Model Context Protocol (MCP) server, published by the hi-godot project on GitHub, that connects a coding agent to a Godot editor that is open on your machine. There is no chat window or interface of its own. The add-on hands its tools to an agent you already run, whether that is Claude Code, OpenAI Codex or Cursor, and the agent calls them by name. Its 4.x line shipped through 2026, with version 4.4.1 released on 3 November 2026.

Read the full background

The chain has three hops. Your agent starts a small bridge over standard input and output, the bridge talks to a local Python server, and the server talks to the Godot editor plugin over a WebSocket that only accepts loopback connections with rotating credentials. The plugin then does the work inside the editor: it creates nodes, writes and patches GDScript, wires signals, sets up UI, materials, cameras and particles, and can press play and read back the game's log. GDScript the agent writes is parse-checked on the spot, and the error comes back in the same tool response.

The add-on's dock configures 26 clients by itself, including Claude Code, Claude Desktop, Codex, Cursor, VS Code with GitHub Copilot, Hermes Agent and Gemini CLI, and prints a manual command for the rest. That makes it the practical way to let a terminal agent see the scene tree, take a screenshot of the running game and fix its own errors, instead of guessing your node names from pasted code. On 15 November 2026 we read the maintainers' public cyberpunk-hud-demo repository: its friction log, dated 15 April 2026, records Claude Code building a full HUD and pause menu through the add-on with no hand-edited scene files, and lists 12 snags on the way, including a missing scene-instancing call and a failing game screenshot that the current tool reference now covers. It lists next to other agent connectors in the developer skills section.

It is not for everyone. It needs Godot 4.7 or later, so a team on Godot 4.4 or earlier, or still on Godot 3, cannot use it without upgrading. C# projects get less help, because building and compiler errors stay with you. If you only want the agent to launch the project and read the debug output, the older and smaller Coding-Solo godot-mcp server (6,660 GitHub stars, read 15 November 2026) does that without an editor add-on. If you want an assistant panel inside Godot rather than a terminal agent, Ziva is the paid in-editor alternative.

Sources and guides

More Agent Skills on Hokai

  • KrolikSelf-hosted MCP server that gives coding agents a call graph and review memory
  • spicedb-devPlan, build and test SpiceDB permissions without leaving Claude Code.
  • Skill CreatorDesign and package agent skills as self-contained SKILL.md files