debuff

@debuffjp/mcp

An MCP server that lets AI agents inspect, build, test and play a debuff game. It uses the same evidence a human would want: real state, real frames, real test results.

Setup

New projects from npm create debuff include a .mcp.json, so MCP clients that read it (such as Claude Code) pick the server up automatically. For other clients:

claude mcp add debuff -- npx -y @debuffjp/mcp        # Claude Code

{
  "mcpServers": {
    "debuff": { "command": "npx", "args": ["-y", "@debuffjp/mcp"] }
  }
}

Run it from the game's folder, or pass --root <dir> (or set DEBUFF_ROOT).

As a Claude Code plugin, which also adds a skill that teaches Claude the workflow:

/plugin marketplace add debuffjp/debuff
/plugin install debuff@debuff

As a Codex plugin: codex plugin marketplace add debuffjp/debuff, then codex plugin add debuff@debuff.

For Claude Desktop, mcpb/manifest.json describes a desktop extension that asks for the game folder. Pack it from that folder with npx @anthropic-ai/mcpb pack.

Tools

Tool	What it does
inspect_project	Dependencies, plugins and their contributions, assets, last build size
list_assets	Every asset with its type, size and importer, plus warnings and possibly unused ids
typecheck	Type-check with the project's own tsc; no browser needed. Vite doesn't type-check, so run it after edits
build	Production build plus size report; leaks must be empty
run_tests	Run the logic tests in tests/ (keiko): the real game in Node, no browser; failures carry the tick, scene and state
run_scenarios	Run scenarios/*.json headless; per-step results, errors, captures
fuzz	Play with random input; each crash, stuck state or back-and-forth loop comes with an explained error and a scenario that replays it
game_start	Attach to the live game, or open a private headless copy, paused (see below). Headless takes a viewport: [w, h]; reducedMotion forces it on or off
game_step	Advance fixed simulation steps (pausing first); returns state (quiet leaves out your inspect() data)
game_pause	Pause or resume a live game
game_frames	Recent frame breakdown: update/render ms, catch-up steps, quads, draw calls, long tasks; summary gives percentiles and the 3 slowest
game_notes	Moments the human flagged in the overlay, with the state at that tick
game_rewind	Live mode: take the human's game back (or forward) to a tick, paused, with its exact state then
game_replay	Save the live session up to a tick (for example a note's) as a scenario in .debuff/replays/ that reproduces it
game_input	press / release keys; tap, pointer_down, pointer_move, pointer_up
game_state	Scene, tick, logical width/height, renderer, frame stats, reducedMotion, your inspect() data
game_command	Run a command the game registered with devTools({ commands }), with JSON args
game_params / game_set_param	Read and tune typed params (validated)
game_capture	The current frame as a PNG image; path also saves it in the project (thumbnails, store art)
game_errors	Runtime errors, explained: whose problem (your code, a rejected call, a debuff bug), the source line, a hint
game_stop	Close the game

Prompts

Workflows clients can offer as slash commands: playtest, fix-errors, tune-feel, add-mechanic and explain-stutter.

Live and headless

If you have the game open under debuff dev, game_start attaches to that tab (mode live). The overlay shows an agent badge, and every pause, step and param change the agent makes appears in its activity log, where you can undo it. Type a note in the overlay ("jump feels floaty here") to flag a moment; the agent reads it, along with the game state at that tick, through game_notes. game_stop resumes your game and disconnects.

Otherwise, or with mode: "headless", the agent gets a private copy in a headless browser. Time only advances through game_step, so what it sees is reproducible, and a scenario can replay it exactly.