grok-game-mcp

A standalone MCP server for Grok-powered game-asset generation — sprites, directional animation sheets, voiceover, SFX, and chiptune music. Extracted from the claude-mermaid-collab dogfood into a self-contained project (only dependency on Grok is an API key; rendering is pure-TS via jimp + ffmpeg, both bundled).

Setup

bun install
export XAI_API_KEY=xai-...        # or put it in ./config.json or ~/.grok-game-mcp/config.json

Run the MCP server (stdio):

bun run src/server.ts            # artifacts → ./out (override with GROK_GAME_OUT)

Register it with any MCP client (e.g. Claude) as a stdio server running bun run src/server.ts.

Related MCP server: ComfyMCP Studio

Tools

Music exports OGG/MP3/WAV (default OGG) + a .mid sidecar. Run bun run preview for a local gallery (animate sheets, play audio, A/B the DSP presets).

How it works (the patterns)

• Grok decides, our code renders. Grok generates the image/video/voice or, for SFX/music, the parameters/score (sfxr params, chiptune pattern); pure-TS synthesizes the audio.

• Sprite sheets use the turntable trick: a grid of frozen poses on cyan pedestals → one Grok-video orbit rotates every cell in place → frames keyed + sliced + packed.

• Shared audio DSP (ffmpeg filterchains) applies uniformly to voice, SFX, and music (epic-announcer, ice-demon, robot-8bit, bitcrush, …).

Architecture

• tooling/imagegen — image/video/sprite pipeline (jimp + ffmpeg). Provider-agnostic, xAI-first.

• tooling/audiogen — TTS provider, sfxr synth, chiptune synth, shared DSP.

• tooling/config.ts — secret access (env → config.json).

• src/core.ts — standalone generators that write artifacts to out/.

• src/server.ts — the MCP server.

Status

Full parity with the collab dogfood — all 14 tools ported to standalone core (write to out/): image, sprite sheet, voiceover (+DSP), SFX, music, apply-audio-dsp, list-dsp-presets, VFX, tileset, background, define-character, character-animations, estimate-cost, replace-sheet-cell. tsc clean; image/sprite/TTS/SFX/music/tileset paths verified live.

Roadmap done: preview gallery, project-style cohesion, generate_game_pack meta-tool, music OGG/MIDI export. Remaining: neural image→3D (the visual-hull spike — port the pure-TS hull as an optional tool; hosted neural only if a 3D game needs it) and --out packaging (per-asset folders, a pack index/zip).