flash-mod-bridge
Enables AI agents to inspect and modify running Flash SWF files through a patched Ruffle player, including getting/setting display object properties, calling methods, and editing SharedObject data without reloading.
README
flash-mod-bridge
Live Flash / Ruffle bridge for AI agents and tools via MCP (Model Context Protocol).
Inspect a running SWF, walk the display list, get/set properties, call methods, and edit SharedObject data — without reloading for in-memory state — through a forked Ruffle player plus a Python MCP hub.
This is not stock Ruffle (no agent API). It is not AwayFL, Unity, or HTML5/Phaser (see suite links below).
| Piece | Path | Role |
|---|---|---|
| Patched Ruffle (desktop) | Releases → ruffle_desktop.exe |
Play SWF + HTTP mod bridge :8768 |
| Engine patches | engine/ |
modBridgeRpc sources to drop into a Ruffle tree |
| MCP hub | run_mcp.py, translator/ |
FastMCP tools (hub :8767) |
| Chrome inject | inject/ |
Optional page agent for web Ruffle |
| Self-host page | host/ |
Simple local player host |
| Flashpoint pin | flashpoint/ |
Install fork as Flashpoint’s default Ruffle |
| Agent skill | skills/flash-mod-bridge/ |
How agents should use the tools |
Status: Early public release. Core path is desktop + Flashpoint; web inject is optional.
What you can do
| Area | Capabilities |
|---|---|
| Discover | List display tree, find by keywords, list props on a path |
| Read / write | Get/set AVM properties by path (root, stage, root/Child) |
| Call | Invoke methods on display objects with JSON args |
| SharedObject | List SOs, set SO properties (saves / persistent data) |
| Raw RPC | Pass through any player JSON op via flash_mod_rpc |
| Transport | Desktop HTTP bridge, or Chrome inject when using web Ruffle |
Not supported: stock ruffle.rs builds (no modBridgeRpc), AwayFL runtimes, or rewriting the SWF on disk. Prefer live paths over re-decompiling ActionScript unless you have a separate pipeline.
Always start with ping_flash_bridge so you know a player/agent is connected before bulk get/set.
How it fits together
Agent (Cursor / Claude / Grok)
│ MCP stdio
run_mcp.py + translator/ hub :8767
│ HTTP
┌────┴──────────────────────────┐
│ Desktop forked ruffle │ :8768 ← Flashpoint / local SWF
│ Web forked Ruffle + inject/ │ optional browser path
└───────────────────────────────┘
│
AVM display list / SharedObject
| Port | Who |
|---|---|
| 8767 | Python MCP hub (agents connect here) |
| 8768 | Desktop Ruffle mod-bridge HTTP (hub falls back here) |
| 8765 / 8766 | Other suite tools (Unity WebGL / HTML5) — not this repo |
MCP tools
Names as exposed by translator/server.py:
Connection
ping_flash_bridge— hub health, desktop player, connected agentslist_connected_agents— which inject/desktop endpoints are liveflash_ping— ping the player RPCflash_mod_rpc— raw JSON request string to the player
Display list & values
flash_list_display— walk tree (max_depth,limit)flash_find— keyword search (money, score, day, …)flash_list_props— properties on a path (defaultroot)flash_get/flash_set— read/write by path (value_jsonfor set)flash_call—path+method+args_json
SharedObject
flash_list_so— known SharedObjectsflash_set_so_prop— set one SO field (name,prop,value_json)
Paths
| Path | Meaning |
|---|---|
stage |
Stage AS3 object |
root |
Root movie clip |
root/Child/Grand |
Display-list walk |
so:name |
Live SharedObject data |
so:name|prop |
SO property (| when name has / or .) |
Requirements
- Windows x64 for the prebuilt desktop player (typical)
- Python 3.10+ for the MCP hub
- A patched Ruffle build with mod bridge (prebuilt release or apply
engine/yourself) - Optional: Flashpoint for archive SWFs
- Optional: Chrome for
inject/+ web selfhosted Ruffle
Upstream Ruffle quality limits AS1/AS2/AS3 compatibility — this bridge cannot fix content Ruffle cannot run.
Quick start
1. Get the patched player (recommended)
From Releases (tag continuous):
| Asset | Role |
|---|---|
ruffle_desktop.exe |
Patched Ruffle + mod bridge (:8768) |
flash-mod-bridge-mcp.zip |
Python hub + inject/host (optional zip) |
.\ruffle_desktop.exe "D:\path\to\game.swf"
Building from source is optional — see engine/README.md.
2. Run the MCP hub
git clone https://github.com/rkuhn153/flash-mod-bridge.git
cd flash-mod-bridge
python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt
python run_mcp.py
Hub listens on 8767 by default (FLASH_MOD_BRIDGE_PORT).
3. Wire an MCP client
Cursor (~/.cursor/mcp.json):
{
"mcpServers": {
"flash-mod-bridge": {
"command": "C:/Python313/python.exe",
"args": ["C:/path/to/flash-mod-bridge/run_mcp.py"],
"env": {
"FLASH_MOD_BRIDGE_PORT": "8767"
}
}
}
}
Grok Build (~/.grok/config.toml):
[mcp_servers.flash-mod-bridge]
command = 'C:\Python313\python.exe'
args = ['C:\path\to\flash-mod-bridge\run_mcp.py']
enabled = true
[mcp_servers.flash-mod-bridge.env]
FLASH_MOD_BRIDGE_PORT = "8767"
Restart the client (or reload MCP) after config changes.
4. Flashpoint (best for archive SWFs)
Pin the prebuilt player so every Flashpoint .swf launches with the bridge:
cd flash-mod-bridge\flashpoint
.\install-to-flashpoint.ps1 -DesktopExe "C:\Downloads\ruffle_desktop.exe"
# .\install-to-flashpoint.ps1 -FlashpointRoot "D:\Flashpoint" -DesktopExe "..."
Then:
- Restart Flashpoint Launcher
- Play any Flash game → forked Ruffle → :8768
- MCP hub running (
python run_mcp.py) - Agent:
ping_flash_bridge
| Detail | |
|---|---|
| Installed as | %Flashpoint%\Data\Ruffle\standalone\latest\ruffle.exe |
| Pin | Data\Ruffle\.mod-bridge-pin (blocks stock auto-update) |
| Undo | flashpoint\uninstall-pin.ps1 |
Full notes: flashpoint/README.md.
5. Web / inject (optional)
- Build web selfhosted from a patched Ruffle tree; serve
host/ - Load
inject/as an unpacked Chrome extension; hard-refresh the player tab
For Flashpoint standalone you usually do not need inject.
6. First agent session
ping_flash_bridge
→ flash_list_display / flash_find
→ flash_list_props / flash_get
→ flash_set / flash_call
→ flash_list_so / flash_set_so_prop
Player JSON RPC (reference)
The desktop/web player speaks the same ops the MCP wraps:
{"op":"ping"}
{"op":"list_display","max_depth":3,"limit":100}
{"op":"list_props","path":"root"}
{"op":"get","path":"root.someProp"}
{"op":"set","path":"root.someProp","value":999999}
{"op":"call","path":"root","method":"play","args":[]}
{"op":"find","keywords":"money,tip,score","max_depth":5,"limit":60}
{"op":"list_so"}
{"op":"set_so_prop","name":"//example_so","prop":"coins","value":999}
Use flash_mod_rpc with a JSON string when you need an op not wrapped by a dedicated tool.
AI skills
| Skill | Role |
|---|---|
skills/flash-mod-bridge |
Tool order, paths, Flashpoint, limits |
Copy-Item ".\skills\flash-mod-bridge" "$env:USERPROFILE\.cursor\skills\flash-mod-bridge" -Recurse -Force
Related projects (same suite)
| Need | Repo |
|---|---|
| Live Flash / Ruffle (this) | flash-mod-bridge |
| Live Unity get/set/patch | bepinex-mcp |
| Live Unreal runtime (UE4SS) | unreal-engine-mcp |
| Mono C# search | gamecode-rag |
| IL2CPP static decompile | il2cpp-decompiler |
Limits
- Stock Ruffle has no
modBridgeRpc— must use prebuilt or patched build. - Content must run in Ruffle; bridge quality follows emulator support.
- AwayFL / some Coolmath hosts need rehost on the fork.
- Bad paths or spammy set/call can still upset a SWF.
- Nested object writes are more limited than simple primitives.
Layout
flash-mod-bridge/
run_mcp.py # MCP entry
translator/ # FastMCP + HTTP hub
inject/ # Chrome extension agent
host/ # Simple self-host page
engine/ # Ruffle patch sources
flashpoint/ # Install/uninstall pin scripts
skills/flash-mod-bridge/
samples/
requirements.txt
README.md
LICENSE
License
- MCP, inject, host, samples, scripts: MIT — see LICENSE.
- Ruffle engine: MIT / Apache-2.0 (upstream). Patches in
engine/are intended for that tree. - Prebuilt
ruffle_desktop.exeincludes our mod-bridge patches on top of Ruffle.
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。