sportsclaw

Appearance

How It Works ​

You ask a question in plain language. sportsclaw turns that into the right data lookups, runs them against live sources, and writes an answer grounded in what it found.

The loop, in plain terms ​

  1. You ask. "Who's leading the Premier League, and what are the title odds?"
  2. The agent picks tools. It decides it needs current standings and prediction-market prices, and calls the right data skills for each.
  3. Data is fetched live. Those skills pull real results from sources like ESPN and Polymarket — no made-up numbers.
  4. You get one answer. The agent synthesizes everything into a single, sourced response.

Because the data comes from real lookups rather than the model's memory, answers stay current and don't drift into hallucination. If the agent needs a specific team or player and only has a name, it looks up the ID first instead of guessing.

Grounded, not guessed ​

This is the core idea: an LLM on its own will happily invent a score. sportsclaw doesn't let it answer from memory for anything it can look up. Live scores, standings, schedules, player stats, odds, and news all come from real data skills the agent is required to call.

Those skills are provided by sports-skills — the open Python data layer that the installer provisions automatically. It's the source of every sport and market sportsclaw can reach; see the full catalog at sports-skills.sh.

Checked before it's sent ​

Before an answer goes out, a fact-check compares the draft with the raw tool output. Any discrepancy is recorded in the run trace (verification.outcome), and when a tool failed, the answer is held to the data that actually came back. By default the checker flags rather than rewrites — on our benchmark, automatic corrections broke more correct answers than they fixed. See Good to know.

Asking about a sport for the first time ​

sportsclaw ships knowing how to reach 14 sports, but it only installs the ones you use. The first time you ask about, say, cricket, it installs that skill in a couple of seconds and then answers. To load everything at once:

bash
sportsclaw init --all

You can manage installed sports directly with sportsclaw add <sport>, sportsclaw remove <sport>, and sportsclaw list.

Where it runs ​

The same engine powers every surface:

  • CLI — sportsclaw "…" for one-shot answers, sportsclaw chat for a conversation.
  • Bots — long-running Discord and Telegram listeners.
  • Embedded — import the engine into your own TypeScript app.
  • Docker — package the whole thing (engine + data layer) into one image.

Bring your own model ​

sportsclaw is model-agnostic. Point it at Anthropic, OpenAI, Google, or Azure Foundry and the loop works the same way — see Configuration.