@wiselaw-ai/provider-catalog

WiseLaw provider catalog — a pure-JSON data package: verified AI provider endpoints (Anthropic-messages / OpenAI-responses / chat-completions) plus curated model metadata (context limits, per-MTok pricing) snapshotted from models.dev.

Subscription-only products such as Kimi Code and GLM Coding Plan omit metered token prices and may carry explicit model metadata when their request model IDs do not exist in models.dev.

The npm artifact contains no executable code — only catalog.json and its schema.json. Consumers fetch the latest published version at runtime, so the catalog updates by npm publish alone, decoupled from any software release.

Contents

• catalog.json — the catalog (see schema.json for the JSON Schema)
• schema.json — generated from the zod source of truth (src/types.ts)

baseUrl convention

Each endpoint's baseUrl is the base to which a fixed per-protocol suffix yields the wire URL:

| protocol | wire URL | harness usage | | --- | --- | --- | | messages | baseUrl + /v1/messages | Claude Code: ANTHROPIC_BASE_URL=baseUrl | | responses | baseUrl + /v1/responses | Codex: base_url = baseUrl + "/v1", wire_api = "responses" | | chat-completions | baseUrl + /chat/completions | Codex: base_url = baseUrl, wire_api = "chat" |

Maintenance (monorepo)

• src/overrides.json — hand-maintained layer: provider list, verified endpoints (verifiedAt per endpoint), and either a models.dev whitelist or explicit subscription-product models
• bun run refresh — pulls models.dev, merges whitelisted model metadata, validates, rewrites catalog.json/schema.json, bumps the patch version
• publish with npm publish --access public; roll back bad data by moving the latest dist-tag, never by unpublishing

Live verification

Put credentials in the monorepo root .env.local; the file is ignored by Git. Use one variable per catalog provider:

VERIFY_KEY_KIMI_CODE=
VERIFY_KEY_ZHIPUAI=
VERIFY_KEY_ALIBABA_CN=
VERIFY_KEY_OPENROUTER=
VERIFY_KEY_MINIMAX=

From apps/provider-catalog, verify both layers:

# Every published protocol with one known model
bun --env-file=../../.env.local src/verify-endpoints.ts \
  kimi-code zhipuai alibaba-cn openrouter

# Every model through one preferred verified protocol
bun --env-file=../../.env.local src/verify-models.ts \
  kimi-code zhipuai alibaba-cn openrouter

An endpoint receives a new verifiedAt date only after a minimal authenticated request succeeds. A model is published only after its exact request ID also succeeds. The scripts never print keys and redact them from provider errors. VERIFY_KEY_MOONSHOTAI remains an accepted local fallback for the renamed kimi-code product preset, so existing ignored env files continue to work.

Kimi credentials require special care:

• api.kimi.com/coding is the Kimi Code membership product, not an overseas alias for the Open Platform.
• China Open Platform uses api.moonshot.cn; international Open Platform uses api.moonshot.ai. Accounts, balances, and keys are region-isolated.
• Only protocols represented by verified endpoint entries are published.

Official references:

• Kimi Code models and protocol Base URLs
• Kimi API product and region troubleshooting
• GLM Coding Plan integration
• Alibaba Cloud Model Studio
• OpenRouter API documentation