Open source · Built on the GreyhoundAPI

The AI lab for greyhound racing.

Six open projects that read the same feed you can. Predictions made before the off and scored against the result, questions answered from the real archive, and the models underneath — all of it public, all of it runnable on your own key.

6
Projects
2
Live now
490
Races called
48.0%
Winner found
MIT
Licence
Traps 1 2 3 4 5 6 Model Ranked call 1 Trap 1 31% 6 Trap 6 22% 2 Trap 2 18% Scored after the off

The lab

Six projects, two of them running today

Each one is a small, single-purpose tool built on the public API. Open the repo, or read how it works.

greyhound-ai

Live
Ranked previews · Python

Reads a racecard and returns a ranked prediction with its reasoning — made before the off, then scored against the real result. Nothing cherry-picked.

490 
races called
48.0% 
winner
76.5% 
podium

greyhound-insights

Live
Plain-English Q&A · Python

Ask the archive a question in ordinary English. It picks its own endpoints, fetches the real races and answers with what it actually found — showing which tools it used.

tools
questions
100% 
grounded

greyhound-mcp

Coming soon
Model Context Protocol · TypeScript

An MCP server for the feed, so Claude, ChatGPT or Cursor can answer "what is Romford's trap-1 rate this year" straight from the live data — no client to write.

 
Repo opens at launch

greyhound-backtest

Coming soon
Strategy harness · Python

Describe a rule, run it across the whole archive, get honest out-of-sample numbers with the staking maths done properly. Built to make a bad idea obvious quickly.

 
Repo opens at launch

greyhound-trapbias

Coming soon
Draw model · Python

Trap advantage modelled per track, distance and going, with confidence intervals — so you know when a bias is real and when it is 40 races of noise.

 
Repo opens at launch

greyhound-form

Coming soon
Form summariser · Python

Turns a dog's race-by-race record into a readable paragraph a human would actually write — for previews, alerts and anywhere raw form is unreadable.

 
Repo opens at launch

Why any of this is public

The data is the product. The models are the proof.

We sell a feed, not a tipping service. Publishing what can be built on it — including when it does not work — is a better argument than a marketing page.

01

Scored, not selected

Every prediction is stored before the race runs and scored afterwards. The accuracy figure includes the bad days, because a number you can only see after the fact is worthless.

02

Runnable by you

Nothing here needs our infrastructure. Clone the repo, point it at your own key, and you get the same output — which is also the fastest way to see whether the feed suits you.

03

Not betting advice

These are research tools. They are built to be inspected and argued with, not followed. No tips, no ratings, no service.

Two minutes

Run the lab on your own key

Install, point it at a race, and read the ranked preview in your terminal. The free sandbox key is enough to try it.

greyhound-ai
# install
pip install greyhound-ai

# rank tonight's card at Romford
greyhound-ai preview --track romford --date today

Romford 19:04 · A3 · 400m
  1. Trap 1  Swift Cassandra   0.31
  2. Trap 6  Coolavanny Pin    0.22
  3. Trap 2  Newlawn Fifi      0.18
  reasoning: T1 rails runner, +4.7pt draw edge at 400m…

What's next

Four more in progress

An MCP server so any AI assistant can query the feed directly, a backtesting harness, a proper draw-bias model and a form summariser. Each one ships the same way — open repo, runnable on your key.

Read this bit. The lab is for research and education and is not betting advice. Accuracy figures are true out-of-sample — each race predicted before its result was known — but they are still small samples on a hard problem. Treat them as evidence the pipeline works, not as an edge.