Choosing a model and AI settings
Automatic play needs an LLM. Skip this page if you only play manually.
Where the settings are
Section titled “Where the settings are”Settings has four tabs: LLM, Automatic play, Advanced, and Actions. Every field and its default is covered in Panel reference.
Pick a backend first
Section titled “Pick a backend first”There are four, not just OpenRouter:
| Backend | Needs an API key? | Good if |
|---|---|---|
| Codex CLI | No — reuses the CLI you’re already signed in to | You already use Codex |
| Claude CLI | No — same | You already use Claude Code |
| OpenRouter | Yes | You want the best cost/performance — see the measurements below |
| OpenAI-compatible | Depends on the endpoint | Local models (the default points at Ollama) or self-hosting |
If you already have the Claude or Codex CLI, just pick that backend — no separate API key to set up.
To run entirely locally, choose OpenAI-compatible and put your Ollama address in Base URL. No API cost and nothing leaves your machine — but local models are usually slower and may not keep up with the game’s turn cadence, so raise Request timeout under Advanced.
Which OpenRouter model
Section titled “Which OpenRouter model”The bundled scripts/bench-models.js scores candidate models using the prompt
that actually ships, on two things: do they keep track of what’s really in
the bag, and do they give up on out-of-range targets. Measured results
(cost is an estimate for 8 hours of continuous play):
| Model | Median response | Inventory | Distance | Cost / 8h |
|---|---|---|---|---|
qwen/qwen3.7-flash (default) | 0.7s | 3/3 | 3/3 | $0.60 |
openai/gpt-oss-20b | 0.5s | 3/3 | 3/3 | $0.86 |
anthropic/claude-haiku-4.5 | 1.5s | 3/3 | 3/3 | $14.34 |
mistralai/mistral-nemo | 0.3s | 0/3 | 0/3 | $0.23 |
inclusionai/ling-2.6-flash | 0.7s | 0/3 | 1/3 | $0.14 |
The default is qwen/qwen3.7-flash: it keeps both rules, answers in well under
a second, and costs a fraction of a frontier model — which is exactly what buys
the 8-second turn cadence the game wants.
Look at the bottom two: cheap and completely failing. They misremember the bag and chase targets they can’t reach. What that looks like in play is a character standing around doing nothing useful.
Running the benchmark yourself
Section titled “Running the benchmark yourself”To measure other models:
OPENROUTER_API_KEY=sk-or-... node scripts/bench-models.jsIt prices against the live catalogue, so the numbers above move with provider pricing.