What is Sports Skills for AI agents?

Sports Skills packages sports-domain instructions in the Agent Skills format. A skill centers on a SKILL.md file that explains when a capability applies, how to call it, what its inputs mean, and how to interpret the returned data. That instruction layer sits beside the executable path, giving a compatible agent more context than a bare endpoint name while keeping the underlying lookup deterministic.

The collection is an open-source interface over public third-party sports data. It does not turn those sources into owned data, and it does not change the permissions attached to them. Its practical role is narrower and useful: make public sports lookups legible to an agent, keep commands inspectable, and return data through a documented path rather than asking the language model to recall a score or table from memory.

The standalone documentation presents two public installation paths. For an Agent Skills-compatible environment, use `npx skills add machina-sports/sports-skills`. This installs the skill instructions so the host agent can discover the relevant SKILL.md material. The documentation and repository remain the authority for prerequisites, supported commands, and any changes to the installation flow.

A Python install path is also available with `pip install sports-skills`. That path is useful when application code needs to call the packaged sports functions directly or when a developer wants to inspect deterministic output outside an agent conversation. Installation does not create new rights to upstream data. In either path, review the skill documentation and the terms of the public source before depending on it.

Read-only means the normal public layer retrieves and analyzes information without changing an external account or submitting a transaction. A request can ask for available sports context, and the selected skill can fetch and structure what its public source returns. The collection should not be treated as permission to take an action merely because related information is visible.

No sports-data API key or signup is required for this public read-only layer. That removes an account step; it does not remove upstream limits, availability changes, attribution duties, or usage restrictions. Source terms apply. Freshness and coverage follow the public source used by the individual skill, so developers should display source context and handle missing, delayed, or changed responses as normal operating conditions.

The public collection is intended for personal, educational, and research use with public source data. It is not proprietary or licensed production data. A repository license can govern the code and written skill files, while the data returned by a third party remains governed by that source. Those are separate rights questions, and both matter when a prototype becomes an application other people depend on.

Commercial use and licensed production are a separate governed Machina platform path. That path should be evaluated against the required data rights, operational controls, and support contract; it is not silently included in the open-source collection. Sports Skills remains useful at the earlier layer: inspecting sports-data shapes, learning agent conventions, testing read-only flows, and deciding which source contracts a production design would need.

Treat each public sports source as a contract the agent must respect: its schema, identity conventions, usage terms, freshness limits, and failure modes travel with the instruction, so a convenient command never becomes an implied data right.

Start with the individual skill, not the collection name. Confirm that the documented sport and query match the intended use, inspect the expected inputs and output, identify the public source, and record what the source says about permitted use. Then test empty results, unavailable fields, stale responses, identifier mismatches, and upstream errors. The useful artifact is not only a successful lookup; it is a clear boundary for what the lookup can and cannot support.

Keep the boundary visible in the product that consumes the data. Cite the source where appropriate, avoid presenting public availability as a service guarantee, and retain a way to disable or replace a failing lookup. If the requirement calls for governed operations or data under a production license, move that requirement to the separate Machina platform evaluation instead of stretching an open-source skill beyond its stated role.