Control Roblox from any MCP-compatible AI (Claude, Gemini, Cursor, Continue) through your exploit executor's WebSocket.
RobloxMCP bridges an MCP server to a running Roblox client via WebSocket. The Roblox side is a single Lua script you run inside an executor (Solara, Potassium, Synapse, etc). Once the script connects back to the local MCP server, your AI assistant can read game state, run arbitrary Lua, see what's on screen, and drive the player — all through a stable tool interface.
AI (Gemini / Claude / Cursor) <--stdio or HTTP--> MCP Server <--WebSocket--> Roblox Executor
The AI gets a clean tool API. The executor handles every privileged action. The MCP server just routes.
execute_lua— run arbitrary Lua in the executor environment. Full power. Every call is auto-recorded.dry_run_lua— compile-check Lua without running. Catch syntax errors before execution.get_recorded_scripts— replay history of everyexecute_luacall (code, result, error, duration). The AI builds its own snippet library.decompile_script— pull Lua source from any LocalScript / ModuleScript / Script. Uses the executor'sdecompileAPI.
describe_view— AI vision substitute: camera state, on-screen players (screen coords + distance + occlusion + team + tool equipped + health), nearby parts sorted by distance, visible UI text with absolute screen positions. Works on every executor, no image data.capture_screenshot— viewport → base64 PNG viaCaptureService. Image content block returned so vision-capable models (Gemini, Claude) see it directly. Falls back to content ID when executor lacks file APIs.
- Player tools —
get_players,get_player_info,teleport_player,kick_player. - World tools —
get_workspace_children,find_parts,spawn_part,destroy_instance. query_instances— search the entire DataModel with name/class/IsA/attribute filters. Way more flexible thanfind_parts.
list_remotes— every RemoteEvent / RemoteFunction / UnreliableRemoteEvent / Bindable in the DataModel.fire_remote— FireServer or InvokeServer any remote by path with arbitrary args.start_remote_spy/stop_remote_spy/get_remote_log/clear_remote_log— hook__namecallviahookmetamethod, log every fired/invoked remote with flattened args (Instances become{__type, path, class}). Works on any exec withgetrawmetatable + hookmetamethod + newcclosure.
take_state_snapshot/diff_state_from— capture flat instance snapshots, diff later: added / removed / moved (>0.01 stud) / reclassed. AI detects events without polling.get_dev_console_logs/clear_dev_console_logs— everyprint/warn/errorfrom the game and script is mirrored to a ring buffer. AI reads its own errors and self-corrects.
- HTTP transport — persistent Node server that survives AI-client restarts. Stdio transport also supported.
- Live dashboard —
http://127.0.0.1:8766/shows connection state, recent tool calls, snapshot count, uptime. - Auto-reconnect — Lua client reconnects with exponential backoff on drop.
- Localhost-only — WS bridge and MCP HTTP endpoint are intended for
127.0.0.1use. Don't expose the WS port across the network without adding your own auth in front.
Live, unedited captures of Claude Code driving Roblox through RobloxMCP.
If the player doesn't render on your browser, download / open
media/demo-full.mp4directly.
Direct link: media/demo-quick.mp4.
execute_lua — printing "Hello, World!" from the AI side:
AI builds an ESP over every entity in the game on request:
- Node.js 18+ (for the MCP server)
- npm
- A Roblox executor with WebSocket support (see list below)
- An MCP-compatible AI client (Claude Desktop, Gemini CLI, Cursor, Continue, etc)
Any executor that exposes a WebSocket client API. Tested / known-good shims included:
- Solara ✅
- Potassium ✅
- Synapse X (
syn.websocket.connect) ✅ - Krnl (
Krnl.WebSocket.connect) ✅ - Wave / Velocity / Fluxus / Trigon — anything exposing global
WebSocket.connect✅
For full screenshot-to-base64 support the executor also needs writefile + readfile + isfile (most modern ones do). Without those, capture_screenshot returns just the content ID and describe_view still works as the vision path.
Three windows total: your server terminal, your AI client, your Roblox + executor. Once set up, daily use = just launch the server.
cd RobloxMCP
npm install
npm run build1. Start the server (in its own PowerShell window — keep it open):
cd "C:\path\to\ExecRobloxMCP"
node dist/index.js --httpYou should see:
[bridge] WS listening on ws://0.0.0.0:8765
[RobloxMCP] HTTP MCP at http://127.0.0.1:8766/mcp (WS bridge port: 8765)
Want different ports? Set MCP_HTTP_PORT and/or ROBLOX_WS_PORT before the command. The Lua client auto-tries ports 8765 and 8767 on both localhost and 127.0.0.1, so default and HTTP-mode setups both work with no edits. If you use a totally custom ROBLOX_WS_PORT, add it to the WS_URLS list at the top of roblox/client.lua.
2. Register with your AI client (only the first time, or after a config wipe):
-
Claude Code (CLI / VS Code):
claude mcp add roblox --scope user --transport http http://127.0.0.1:8766/mcp
Then
/mcpinside Claude Code to confirmroblox · √ connected. -
Claude Desktop — edit
%APPDATA%\Claude\claude_desktop_config.json:{ "mcpServers": { "RobloxMCP": { "url": "http://127.0.0.1:8766/mcp" } } }Then fully quit + reopen Claude Desktop (system tray, not just the window).
-
Cursor — Settings → MCP → Add new MCP server:
{ "mcpServers": { "roblox": { "url": "http://127.0.0.1:8766/mcp" } } }(Or edit
~/.cursor/mcp.jsondirectly with the same content.) Restart Cursor. -
Gemini CLI — edit
~/.gemini/settings.json:{ "mcpServers": { "roblox": { "httpUrl": "http://127.0.0.1:8766/mcp" } } }New shells of
geminipick it up automatically. -
OpenAI Codex CLI — edit
~/.codex/config.toml:[mcp_servers.roblox] url = "http://127.0.0.1:8766/mcp"
Restart
codex. -
Continue (VS Code / JetBrains) — add to
~/.continue/config.jsonunderexperimental.modelContextProtocolServers:{ "transport": { "type": "streamable-http", "url": "http://127.0.0.1:8766/mcp" } }Reload the IDE window.
-
Anything else MCP-compatible — point it at
http://127.0.0.1:8766/mcpusing whichever HTTP transport key it expects (url,httpUrl,endpoint, etc.). If your client only speaks stdio, see Advanced: stdio mode at the bottom.
3. Inject the Lua client into Roblox:
- Join the Roblox experience you want to control.
- Open your executor (Solara, Wave, Krnl, etc).
- Paste the contents of roblox/client.lua and execute.
Roblox dev console (F9) should show:
[mcp-bridge] client loaded, connecting…
[mcp-bridge] connected to ws://localhost:8765
Your server terminal will also print [bridge] incoming connection from ::1.
4. Use it. Ask your AI anything:
- "List all players."
- "What can you see right now?" (uses
describe_view) - "Teleport me next to PlayerName."
- "Take a screenshot."
- "Run this Lua:
return Workspace:GetChildren()" - "Show me the last 50 dev console errors."
Open http://127.0.0.1:8766/ in your browser. Auto-refreshing live view: Roblox connection state, every tool call with duration + status, snapshot count, uptime.
In any shell: curl http://127.0.0.1:8766/health → {"ok":true,"roblox_connected":true|false}. Programmatic JSON via /api/status and recent tool calls via /api/tool-log?limit=50.
- Server: Ctrl+C in the server terminal (or close the window).
- Roblox side: rejoin / re-execute / disconnect — server holds no persistent state.
| Symptom | Likely cause | Fix |
|---|---|---|
Server: EADDRINUSE |
Another node already on 8765/8766 | Kill it: Get-NetTCPConnection -LocalPort 8766 -State Listen | %{ taskkill /F /PID $_.OwningProcess } |
| AI client says "connection refused" | Server not running | Start the server (step 1) |
/mcp shows roblox × failed |
Server crashed or never started | Check the server terminal output |
| Tool call: "Roblox client not connected" | Lua not running in executor | Re-paste + execute roblox/client.lua |
Roblox console: all ports failed: ... retry in Xs |
Server not running, or on a non-standard port | Start the server. If using a custom ROBLOX_WS_PORT, add that ws://localhost:<port> to WS_URLS in the Lua |
Roblox console: WebSocket timed out after Ns |
Lua reached a port with nothing listening | Server isn't up on that port — start it (default WS 8765) |
Skip step 1 entirely and let your AI client manage the server lifecycle:
claude mcp add roblox node "C:\path\to\ExecRobloxMCP\dist\index.js" --scope userTradeoff: when Claude Code closes, the server dies and Roblox disconnects. HTTP mode (above) avoids that.
MCP → Roblox: {"id":"uuid","action":"...","...":...}
Roblox → MCP: {"id":"uuid","ok":true|false,"result"|"error":...}
Unsolicited: {"event":"log","level":"info|warn|error","message":"..."}
Add a tool = add an entry in src/tools.ts plus a handler in actions inside roblox/client.lua. That's it.
Made by SkieHacker.
If you fork, build on top, or include this in another project, keep the copyright line in LICENSE. That's the only string attached.
MIT — see LICENSE.
You can use, modify, distribute, and sell it. You only must preserve the copyright notice and the license text. Don't strip the credit.

