sportsclaw is an open-source agent engine with keyless live data, market odds and real-time game events built in. Run it in your terminal, ship it as a bot — or use it inside Machina, where every project ships with its own sportsclaw relay.
curl -fsSL https://sportsclaw.gg/install.sh | bashmacOS / Linux. Installs the engine from npm and the sports-skills data layer from PyPI.
The data, the live game events and the real-time plumbing ship in the box, so you spend your time on the experience instead of stitching feeds together.
Ask in plain language. Scores, standings, stats, odds and news come from live lookups across 14 sports — the model decides what to fetch, never what the numbers are.
Read moreOne command runs a community bot with rich embeds, buttons, polls, generated graphics and image understanding. Every deployed bot is read-only for markets.
Read more“Alert me about the Lakers” is the whole setup. Followers get a message at tip-off, on lead changes and at the final — no polling code to write.
Read moreESPN, Kalshi and Polymarket in one query, with de-vig, edge, Kelly and arbitrage math built in. Built to track markets, not to trade them.
Read moreOperator mode wakes on a schedule, decides whether anything is worth saying, and publishes on its own — the base for broadcast and studio agents.
Read moreImport the engine into your TypeScript app, run it in Docker or behind the relay API. Anthropic, OpenAI, Google, Azure Foundry or any OpenAI-compatible endpoint.
Read moreAn LLM on its own will happily invent a score. sportsclaw doesn't let it answer from memory for anything it can look up — every number comes from a real data call.
A fast router reads the question and picks the one or two skills that matter, so the model only sees tools it actually needs.
route → football · polymarketsports-skills runs deterministic Python against ESPN, FastF1, Kalshi, Polymarket and more. Identical calls hit a turn cache; huge results are queried, not truncated.
football_get_season_standings(…)A fact-check compares the draft with the raw tool output and records any discrepancy in the run trace. If a tool failed, the answer is held to the data that came back.
verification: kept | flaggedOne sourced answer, wherever the question came from: your terminal, a bot, the relay API, or your own app. Add --json for NDJSON.
sportsclaw "…" --jsonMachina Sports builds and runs sportsclaw. The engine you install is the same one the platform deploys inside every project — so what you prototype on a laptop is what runs in production, with licensed data and durable workflows on top.
curl -fsSL https://sportsclaw.gg/install.sh | bashsportsclaw machina connectalongside the project's Client API, workers, Redis and vault
ESPN · FastF1 · Kalshi · Polymarket · news
real-time feeds, odds, the project's own documents
sportsclaw machina connect, which mints a durable service key through machina-cli. Every project starts with 17 sports and market skills that need no data credentials.
The project’s MCP tools sit next to the keyless skills. When a public feed runs out, an upgrade signal points to the licensed path — once, never as a nag.
With a Machina server connected, memory defaults to project documents — SOUL, fan profile, reflections, strategy — isolated per user and agent.
Long, multi-step work goes to the project’s loop-runner. Every turn is a document; a turn only closes after a code check and an independent evaluator.
/api/decide answers bounded Choice, Score and Noul questions in one authenticated call. Disabled by default.
Play-by-play turned into clip windows: rights-gated, bounded async jobs with SHA-256 receipts for every clip.
| Aspect | Open source | With Machina |
|---|---|---|
| Data | Public APIs, keyless — ESPN, FastF1, Kalshi, Polymarket… | Adds licensed, real-time feeds through the project's MCP server |
| Best for | Personal use, prototypes, community bots | Commercial production, with SLAs and support |
| Memory | Local files, or a Hindsight server | Project documents, isolated per user and agent |
| Long tasks | In-process, one turn at a time | Durable loop on the project, resumable |
| Setup | curl … | bash | sportsclaw machina connect — or nothing, inside a project |
The trade-offs are deliberate. Here's what sportsclaw will and won't do — and where each rule is enforced.
Order, wallet and balance tools are blocked before they reach the model and again before execution. Bots and servers are hard read-only.
The open skills ride public APIs meant for personal, non-commercial use. Production workloads need licensed data.
On our own benchmark, auto-correcting drafts broke more right answers than it fixed. The default fact-check now flags instead of rewriting.
A connected MCP server exposes all of its tools — writes included — unless you pin a tools allowlist.
Writing files, running shell commands and starting durable sessions wait for explicit approval.
Jev routing and verification, OpenShell sandboxing and Hindsight memory are off by default. Cloud transfer needs explicit consent.
20,000-character prompts, 4 concurrent queries, 300 s max timeout — per relay instance, not distributed quotas.
Relay features ship with relay-v* releases. /api/capabilities reports the engine, skills and limits a relay actually runs.
All of it, with the details: Good to know: defaults, limits and trade-offs →
sportsclaw is developed in public. Engine fixes, new bot surfaces, benchmark cases and docs all land through pull requests on GitHub — the same way the Machina team ships.
git clone https://github.com/machina-sports/sportsclaw
cd sportsclaw && npm installnpm run build # tsc, strict
npm test # node:test suiteCI runs both plus a smoke test that needs no API key.
Kebab-case files, explicit .js imports, and no invented feeds — if it isn't in sports-skills, the agent can't claim it.
npm run docs:dev · files in docs/Standing on the shoulders of NanoClaw, OpenClaw, NemoClaw, pi.dev, the Vercel AI SDK and Clack.
Paste this into Claude Code, Cursor or any coding agent. /llms.txt lists every doc page as a fetchable URL, so the agent can read the whole reference on its own.
Build a sports AI app with sportsclaw. First read https://sportsclaw.gg/opensource/llms.txt
for the full doc map, then follow https://sportsclaw.gg/opensource/getting-started/quickstart
to install and scaffold. sportsclaw gives you keyless live scores, standings, odds
and markets via sports-skills, plus one-command Discord and Telegram bots. Use the
docs at https://sportsclaw.gg/opensource/ for the CLI, data coverage, and deployment.