@rovecode-labs/models

Offline model catalog for terminal agents: context windows, pricing (including prompt-size tiers), capabilities and provider env-var names — trimmed from models.dev and shipped with the package, so the first lookup costs ~1.7 ms and no network.

import { ModelCatalog, resolveContextWindow, keyNameFor } from "@rovecode-labs/models";

const catalog = new ModelCatalog();

catalog.lookup("anthropic", "claude-opus-5");
// → { contextWindow: 1_000_000, maxOutput: 128_000,
//     pricing: { inputPerMTok: 5, outputPerMTok: 25, cacheReadPerMTok: 0.5, cacheWritePerMTok: 6.25 },
//     supportsTools: true, supportsReasoning: true, source: "models.dev" }

catalog.modelsFor("anthropic");          // every model id the catalog knows for a provider
catalog.supportsImages("openai", "gpt-5.4");

resolveContextWindow(catalog, spec, provider, model);
// config (per-model, then provider-wide) → catalog → an explicit 128k assumption, marked as assumed

keyNameFor("anthropic"); // → "ANTHROPIC_API_KEY"  (the env var a key is read from)

What makes it different

• The data ships with the package. 213 providers / 7,527 models, trimmed to the six fields an agent actually reads, split into a 138 KB hot file (the providers every built-in maps to) and a lazily loaded extra file (the other 196 — reachable: a hand-registered provider id is tried as its own models.dev key).
• Honest unknowns. A model that is not in the catalog resolves to undefined, never to a guess; the context-window fallback is explicit and marked source: "assumed".
• Pricing tiers models.dev has no field for — per-request vs marginal repricing past a prompt-size threshold (xAI, OpenAI, Gemini), each with the vendor page and the day it was read.
• A local overlay with provenance for ids the snapshot lacks (DeepSeek's API aliases, retired xAI slugs) — every row names the page it was read from and when, and source: "local" says so.

Live refresh (opt-in)

const catalog = new ModelCatalog({ fetchFn: fetch, cacheDir: ".cache/models" });
await catalog.refresh(); // layers https://models.dev/api.json over the bundled data, 24 h disk cache

Requirements

Bun ≥ 1.3 (the package ships TypeScript source and JSON data; the runtime imports both directly).

Development

This package is developed inside the rovecode monorepo (packages/models) and mirrored here with its full history. Issues and PRs are welcome on either repository; the monorepo is where the code lands first.

bun test                          # the package's own suite
bun ../../scripts/build-model-index.mjs --check   # (from the monorepo) the index matches @opencode-ai/models