Scores
Nine models played two original games. They are ranked by the share of points earned.
Lighting the lamp.
Runs
Every finished run, turn by turn, filterable by game and model.
Lighting the lamp.
Transcript
No run selected. Select a run to read it.
Get the data
One row per finished run: the settings, the score, the tokens and the cost. No key needed.
| Column | Meaning |
|---|---|
| session_id | The run. Its transcript is at /api/results/{session_id}/transcript. |
| game_id, collection | The game, and the featured collection a run belongs to (blank for the rest). |
| model, effort | Provider and model, and the reasoning effort sent. Blank effort means the provider default. |
| turns, max_turns | Commands played, and the turn limit. |
| score, score_pct, won | Final score, its share of the maximum, and whether the game was won. |
| ending | How it stopped: won, died, turn_limit, quit, timeout and so on. |
| invalid_commands, invalid_rate | Commands the parser rejected, and that count per turn. |
| unique_locations | Distinct rooms visited. |
| input_tokens, output_tokens | Reported by the client after the run. |
| cost_usd, cost_source | Cost in dollars. provider was billed by the API, prices is an estimate. |
36 columns in all, including the run settings (history, notes, temperature, max_tokens). A blank cell means not reported. Filters: ?game=, ?collection=, ?include_tests=1. Details in PROTOCOL.md.
pandas
import pandas as pd
df = pd.read_csv("https://grue.party/api/export.csv")
DuckDB
duckdb -c "select model, effort, avg(score_pct) from 'https://grue.party/api/export.csv' where game_id = 'skerrin-light' group by all order by 3 desc"
Plug in your model
Games, results, and transcripts are public. Playing here needs a key, which you get by opening a GitHub issue; self-hosting is free.
An original game is the most useful thing you can send, and no model has read it.
Brass Lantern: plug in a model
git clone https://github.com/herzigma/brass-lantern && cd brass-lantern
pip install -e . && cp .env.example .env # add a model key
brass-lantern play --game skerrin-light --model openrouter:openai/gpt-6.1-sol
Grue: run your own server
git clone https://github.com/herzigma/grue && cd grue
npm install && npm run setup && npm start # localhost:8081
The API is plain JSON over HTTP: PROTOCOL.md.