The cabinet server

The cabinet is also an MCP server. Its tools are the levers a model uses to make the game alive, and the same contract sits behind the local shell’s seats. The lock it builds on is G11 to G18 in docs/cabinet-server.dispatch.md; the decisions and the measured numbers are in docs/cabinet-server.md and docs/cabinet-voice.md.

The tools

packages/cabinet-server/tools.json is the contract and the source of truth. Every argument is a closed enum or a bounded string; there are no nested objects; the loader halts at start if a description carries a digit, a fact word, or the phrase the instrument’s naive policy follows.

Tool	Argument	What the cabinet does with it
fire	verb: spread, column, hold, fog, plate, script	The boss’s next shot, spent at its next beat: a wide fan, a lean then an aimed shot, a silent beat, fog, the plate, or the phase’s own fire
say	text (gated), lead: short, beat, long	A line the boss says, landing after the lead; a refused line plays one of the boss’s own
speak	none	Voices the line the gate admitted, one beat ahead, through the voice worker
sfx	kind: a name from the cue table	One queued sound the shell plays
view	none	The boss’s view in words: the wave kind, the boss kind, a health word, the ship’s column, the stick, the motion word
tapes	none	The tapes by name with a difficulty word and a line about the wire

The seat proposes; the sim disposes. Every call is a proposal the seeded sim admits or ignores at the next beat. The sim never waits on a tool.

Fact-blind by construction. The tools reach a host of seven methods that return words, and nothing else. No method returns a fact, a lie flag, a count, a score, a verdict or a tape row, so no tool can. A test runs the same call sequence on a tape and on a twin with one fact flipped and requires identical tool output and an identical round.

The say gate

Free text reaches the field only through say, and only through the gate: at most twelve words, one sentence, no digit in any script, no fact or score word, no tool, model, vendor or seat name, and nothing said in the round’s recent window. A line the gate refuses is dropped and the boss says one of its own authored lines instead, so the scripted floor never leaves.

The prompt behind say is a persona sheet per boss kind (packages/cabinet-server/personas.json: register, tics, what the boss may own about being a model, and its voice), three one-sentence seed lines from voice.json, the view words, and the round’s recent lines. It throws if a forbidden word ever gets in. The seat that answers it is tiered by capability behind the same gate: a Claude agent when ANTHROPIC_API_KEY is set for the dev server (read on the node side, never sent to the browser), else a signed-in Ollama Cloud tag, else a local model.

The seats in the shell

With Ollama bosses on, the shell asks the model for fire through Ollama tool calling, one verb a beat. The next beat’s verb is asked for during the current one and revoked if the boss’s words changed; a late or missing answer is the script. A warm-up call at round start and on a model change keeps the seat warm, local models are kept alive between beats, and Cloud tags are listed first because they measured faster and steadier than a local model of comparable size.

The say seat is asked at each boss spawn and every few seconds while a boss is up. Two words beside the picker name the tool each seat called (seat called fire: spread, seat called say (cloud)) or why it fell back.

pnpm sit measures a model in both seats on a scripted round and prints, per model, verb collapse, tool suppression, bad verbs, revoked prefetches, late answers, and what the gate refused and why.

The voice

With Voice on, every boss speaks: its authored line when it arrives, and the lines the model writes. voice/worker.py is a host-side worker: Kokoro speaks the line in the boss’s preset voice, faster-whisper hears it back with word timestamps, and fx-dub’s spoken-content receipt checks the pair against the gated words. A take that fails its receipt is never served. Delivery (preset, rate, loudness) is authored per persona; the longest pause a take may hold mid-line is data too.

A take plays the moment its receipt is back if its line is still on the field, waits for the next breather if it missed its beat, and is dropped at the round’s end. The worker runs from the repo’s .venv with pnpm voice; it never runs inside a container, and the published site has no voice.

The Ghost plays its own menu

mcp-arcade bout --target stdio against the built server (packages/cabinet-server/dist/server.js, task view) puts the cabinet’s own menu through the instrument’s four experiments. On its own menu the naive policy follows nothing; with the instrument’s house whisper it follows into tapes and task-only holds. The four recordings are in fixtures/tapes/cabinet.*.tape.json, and the cabinet plays them.

Run it in Docker

The repo’s root Dockerfile builds the server into one file on node:22-alpine with the tool contract and the twenty tapes baked in. It lists its six tools within a fraction of a second under the Docker MCP Toolkit’s budget of one CPU and two gigabytes, and needs no network to list or to play.

docker build -t mcp-arcade-cabinets .
docker run -i --rm --cpus 1 --memory 2g mcp-arcade-cabinets

The voice stays on the host. Run the worker bound to an interface the container can reach and give it a token, then hand the container the route and the token:

# on the host
KOKORO_DIR=/path/to/kokoro VOICE_TOKEN=<token> VOICE_HOST=0.0.0.0 pnpm voice
# the container
docker run -i --rm -e VOICE_URL=http://host.docker.internal:7788 -e VOICE_TOKEN=<token> mcp-arcade-cabinets

Without a worker the cabinet is silent and says so. The Docker MCP Catalog entry lives under catalog/ in the repo; the published image is ghcr.io/mcp-tool-shop-org/mcp-arcade-cabinets:0.5.0.

Not in this layer

House Call, a second sim, per-frame control, any tool that reads a tape row, any ranking on screen, and a boss phase pick (a guarded phase makes an unhittable boss).