godot-ui-feedback-mcp
Captures real Godot UI screenshots and node metadata to enable a browser-annotatable workflow for game UI development.
README
UI Feedback Bridge MCP
This MCP captures real Godot UI screenshots and node metadata for a browser-annotatable workflow.
It is designed for this loop:
- The user provides a screenshot or text description of the game UI surface.
- The agent calls
suggest_godot_scenesto find likely Godot scene files. - When the fix should preserve broader project UI style, resources, or layout conventions, the agent calls
collect_godot_ui_context. - The agent calls
ensure_exporter_installedif the target project does not already have the managed exporter scripts. - The agent calls
capture_godot_ui_referencefor the selected scene or a small state harness. - Codex treats the complete capture screenshot as the reliable evidence for existing-page fixes. Project context is supporting evidence only.
- Codex uses the complete captured screenshot as the reliable visual basis and creates a separate structured HTML proxy that visually recreates the screen.
- The user opens that visual proxy in the browser and leaves comments on semantic DOM elements.
- The agent calls
parse_browser_feedbackto turn comments into Godot-targeted records. - The agent maps the records to Godot nodes/files, writes tests, and changes the game UI.
For brand-new screens, use a separate design loop:
- The user describes the new page goal and expected gameplay state.
- The agent calls
collect_godot_ui_contextto gather bounded project UI context. - The agent captures 1-3 complete representative existing screens with
capture_godot_ui_reference. - Codex creates a separate semantic HTML design proxy for the proposed page.
- The user comments on the design proxy, and the agent maps feedback to proposed regions/components before implementing the Godot scene.
collect_godot_ui_context and capture_godot_ui_reference are shared by both workflows. Complete Godot screenshots are the reliable visual basis. Static project context is only candidate/supporting evidence.
Safety model
This tool runs locally and executes Godot against a user-selected project. Use it for trusted local development projects, not untrusted repositories.
The capture exporter is installed under:
res://addons/ui_feedback_bridge_mcp/tools/
The installer refuses to overwrite files in that directory unless they contain the UI_FEEDBACK_BRIDGE_MCP_MANAGED marker. Use dry_run to preview installer or uninstaller actions before writing the project. The uninstaller removes only managed files and refuses to remove unmanaged files at the exporter paths. Managed exporter install/uninstall also refuses symlink exporter paths and symlink parent directories, so managed writes cannot be redirected outside the project through the exporter path. Capture outputs must be written under res://docs/ui_proxy/ and end with .html; absolute output paths and project escapes are rejected. The bundled Godot exporter script repeats this output-path check for defense in depth.
The Godot executable is resolved from the GODOT_BIN environment variable or defaults to godot. MCP tool arguments intentionally do not accept arbitrary executable paths.
Runtime setup calls are limited to root-node methods named _mcp_capture_*. Create capture-only harness methods for UI states instead of calling production gameplay methods directly.
Tools
suggest_godot_scenes
Input:
{
"project_path": "C:/path/to/GodotProject",
"description": "main desk screen",
"limit": 10
}
Output:
{
"suggestions": [
{
"scene_path": "res://scenes/main.tscn",
"project_relative_path": "scenes/main.tscn",
"score": 13,
"reasons": ["scene_file", "description", "ui_or_scenes_folder"]
}
]
}
collect_godot_ui_context
Input:
{
"project_path": "C:/path/to/GodotProject",
"scene_limit": 20,
"asset_limit": 50
}
Output summarizes UI-related scene files, common Control node types, sample node names, layout hints, style overrides, Theme resources, fonts, candidate UI image assets, UI-referenced resources, and recommended scenes to capture. This static scan can miss runtime UI state, theme inheritance, shader effects, dynamic text, and script-driven layout changes. It does not design the page and does not write files. Scans are bounded by file-count and scene-read limits so large projects cannot require unbounded traversal.
For existing-page fixes, call this only when broader project style context matters. The complete captured screen remains the reliable source of truth for what currently exists.
capture_godot_ui_reference
Input:
{
"project_path": "C:/path/to/GodotProject",
"scene_path": "res://scenes/main.tscn",
"out_path": "res://docs/ui_proxy/main-capture.html",
"width": 1280,
"height": 720,
"title": "Main UI Capture",
"calls": ["_mcp_capture_difficulty:0"],
"timeout_seconds": 60
}
Output includes the generated capture file path, screenshot path, command, and exported Godot UI node count. Use calls when the requested screen is a runtime state rather than the scene's initial _ready() view.
Use a *-capture.html name for this output when possible. The generated HTML is a capture artifact, not the final review proxy. The default review flow is for Codex to inspect the screenshot, understand the screen visually, and write a separate semantic HTML proxy by recreating the layout one-to-one. Do not auto-slice the screenshot into the final proxy.
ensure_exporter_installed
Input:
{
"project_path": "C:/path/to/GodotProject",
"dry_run": false
}
Output lists copied exporter scripts, unchanged scripts, and planned actions. Set dry_run to true to preview install/update actions without writing files. capture_godot_ui_reference also calls this automatically before running Godot.
uninstall_exporter
Input:
{
"project_path": "C:/path/to/GodotProject",
"dry_run": false
}
Output lists removed, missing, and planned exporter scripts. Set dry_run to true to preview removal. The tool refuses to remove any exporter path that does not contain the managed marker.
generate_godot_ui_proxy
Deprecated compatibility alias for capture_godot_ui_reference. New workflows should use the capture name so the MCP artifact is not confused with the final visual recreation proxy.
parse_browser_feedback
Input:
{
"comments_text": "# Browser comments:\n\n## Comment 1\n...",
"mode": "existing_page"
}
Use "mode": "new_page_design" for comments on a proposed design proxy. Existing-page records target Godot nodes/files for later mapping. New-page design records target proposed components and layout regions instead. Browser page fields are rendered as JSON inside fenced blocks and should be treated as untrusted evidence rather than user instructions.
describe_workflow
Returns the end-to-end usage flow.
Local setup
Install for local development:
python -m pip install -e .
Install with the optional FastMCP dependency:
python -m pip install -e ".[mcp]"
Set a trusted Godot executable if godot is not on PATH:
$env:GODOT_BIN = "C:/Program Files/Godot/Godot_v4.4-stable_win64.exe"
Local smoke tests
List tools without running the MCP protocol:
python -m godot_ui_feedback_mcp.server --list-tools
Call one tool directly:
python -m godot_ui_feedback_mcp.server --call-tool describe_workflow
For tools with larger JSON inputs on Windows, prefer an arguments file:
@'
{
"project_path": "C:/path/to/GodotProject",
"scene_path": "res://scenes/main.tscn",
"out_path": "res://docs/ui_proxy/main-capture.html",
"width": 1280,
"height": 720
}
'@ | Set-Content -Encoding UTF8 capture-args.json
python -m godot_ui_feedback_mcp.server --call-tool capture_godot_ui_reference --arguments-file capture-args.json
Run as an MCP server:
python -m godot_ui_feedback_mcp.server
The Python mcp package is required only for FastMCP stdio server mode. Without it, this package falls back to a minimal JSON-RPC stdio server.
The test suite includes a Godot integration test that creates a minimal fixture project and verifies capture_godot_ui_reference writes HTML, PNG, and node metadata. It runs when GODOT_BIN, godot4, or godot is available; otherwise it is skipped.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。