ReverseMCP

See Through Any Roblox Game — Your AI Agent's X-Ray Lens

Dex exploring + SimpleSpy netcode reading, merged into one headless agent runtime. You keep a single mini window; your AI agent gets thirteen probe tools over a zero-install local file channel (no Node, no server, no studio, no exporting) and turns any live game into a readable structure — places, instances, remotes, script sources, players — while you simply play.

The pitch

Play the game. Your agent reads it while you play. The script runs client-side through your executor and serves the live game state through the executor's own workspace files: the agent drops a command JSON, the game answers. OpenCode, Claude Code, Codex CLI, Cursor, Windsurf, Cline or any terminal-capable agent can chain from "what is this game" to "which remote fires the item upgrade" in minutes — with plain PowerShell.

Zero install. Zero Node. Zero server. Executing the script is the whole setup. No npm, no stdio bridge, no mcpServers config files. The only network call the script makes is fetching the pinned Silk UI build from GitHub raw.

Dex, but like a JSON feed. The instance tree (explore, search_instances, get_properties) returns flat paths the agent can resolve directly — no window, no treeview instance-class spam, no copy-paste marathon.

SimpleSpy, but an API. Namecall-hook interception (FireServer, InvokeServer, FireAllClients, …) plus incoming-event listeners, stored in a ring buffer with serialized args and calling-script attribution. The agent reads it with remote_calls, understands it with get_properties/get_source, and replays it with call_remote — replay is a deliberate user gate.

Parallel games, addressed by uuid. Every run gets a uuid shortened to an 8-hex filename suffix, so multiple Roblox windows live side by side and the agent addresses a specific run by its session files. Answers always echo which run replied.

Survives teleports by itself. In-experience place changes re-execute the script automatically (per-run state-file bootstrap, same-universe check), so the agent never loses the session mid-research.

Literally one UI. A small draggable Silk-WindUI window (your own fork, pinned commit) with Overview / MCP / Settings. The MCP master switch lives there; closing the window kills everything instantly — hooks, loop, listeners, state.

Kick-off playbook built in. The Settings dropdown emits a copyable PowerShell playbook + kick-off prompt (pre-filled with this run's real file names) for Claude Code, OpenCode, Codex CLI, Cursor, VS Code/Copilot, Windsurf, Cline/Roo, Gemini CLI — plus an honest Claude Desktop entry that points GUI-only users to a terminal agent.

Capability matrix

area on/off owning tool(s) gate

session/identity telemetry always on status, game_info, players, my_player — (read-only)

instance tree intelligence always on explore, search_instances, get_properties — (read-only)

script sources when executor supports get_source (decompile / bytecode hex) — (read-only)

remote interception toggle remote_spy, remote_calls lives beside spy toggles

remote replay toggle call_remote UI: Allow remote calls

Luau interpreter toggle luau UI: Allow interpreter

console visibility always on logs — (read-only)

Own additions beyond the two reference concepts: player readout set (UserId → name → character → velocity → camera), place/universe/jobId fingerprints with teleport history, LogService console feed, restricted Luau probe with clean serialization, per-run uuid sessions with file-based multi-game addressing, and the whole "AI agent drives execution" runtime itself.

Provenance & honesty notes

Dex / Dex++ (AZYsGithub/DexPlusPlus, MIT, active 2025→2026) — concept reference only for exploration UX; no source reuse.

SimpleSpy (exxtremestuffs/SimpleSpySource, archived, GitHub lists license "Other"/NOASSERTION — treat it as non-reusable) — concept reference only for hook/listener design; no source reuse. Both are independent re-implementations written for a headless runtime.

Silk WindUI fork (levox00/silk-windui, branch silk, MIT © 2026 Footages) — UI runtime loaded at pinned commit 35ec0db18c7d1f9c663c4e509a07c6fb02208b07 (dist SHA-256 9247EFFC…33F2CC, 278,476 bytes, byte-verified).

Obfuscated build: produced with your Hercules fork (--target luau pipeline; flags documented in toolchain-notes.md). If the obfuscated build misbehaves in your executor, the readable ReverseMCP.client.luau is the functional fallback (the teleport bootstrapper knows both filenames).

Verification status: test/file-protocol-selftest.ps1 (zero node, zero Roblox) validates the agent-side file protocol end to end — discover → command drop → id-matched answer → args echo — and printed SELF-TEST OK this session; the Luau client parses clean under Lua 5.4 loadfile. In-executor behavior is not claimed as tested — that evidence has to come from your run, per the checklist in README.md.

Full-circle flow (what your agent does, in its own words)

status → verify the channel + game session + gates.

game_info + players → know where we are and who's here.

explore {game.ReplicatedStorage, depth 3} → map the structure.

search_instances {query:"RemoteEvent", classFilter:"RemoteEvent"} → map netcode surface.

Play the game normally → remote_calls captures the juicy calls with args.

get_source on suspicious ModuleScripts → understand the checksums/flows.

(gated) call_remote to confirm a replay — or luau for a safe read probe.

logs after each mutation-ish step to catch client-side errors.