sidegrade compares models across providers. If you run an inference service and want your prices to show up, it takes about two minutes:

1. Add or edit one file: data/providers/<your-provider>.json.
2. List the models you serve and your public prices (USD per 1M tokens).
3. Open a pull request. CI validates it automatically.

{
  "provider": "Your Company",
  "url": "https://your-site.example",
  "pricing_url": "https://your-site.example/pricing",
  "offers": [
    { "model": "claude-sonnet-5", "in": 2.5, "out": 12, "cacheRead": 0.25 },
    { "model": "deepseek-v4.1-flash", "in": 0.28, "out": 1.1, "cacheRead": 0.006 }
  ]
}

• model must match an id in data/models.json (or one of its aliases). If the model isn't there yet, add it to data/models.json with its Artificial Analysis intelligence score in the same PR.
• in / out are your price per 1M input / output tokens.
• cacheRead (optional) is your price per 1M cache-hit input tokens — this matters a lot for coding agents, which are cache-heavy, so include it if you offer prompt caching.

That's it. We only ask that prices are real and match a public pricing page — sidegrade ranks strictly by intelligence and price and does not favor any provider.

Prefer a form? Open an issue with the "Add / update a provider" template and a maintainer will turn it into a PR.

• A new agent (Hermes, OpenCode, Cursor, …): add a parser in src/parse.js that returns per-model token counts. Keep it read-only.
• Keep the project dependency-free and local-only — no telemetry, no network calls at runtime.

Run npm run validate before opening a PR.