[heydryft]

Summary

Adds Runcrate (inference for open-weight models through an OpenAI-compatible API) as a built-in provider card in the extended catalog, the same way the Opper and API Route cards were added.

• Card id: runcrate, category router, disabled by default, existing OpenAI-compatible adapter.
• Base URL: https://api.runcrate.ai/v1. API keys: https://runcrate.ai/dashboard/api-keys
• Default model: deepseek-ai/DeepSeek-V4.1-Flash (1048576 tokens). Suggestions: zai-org/GLM-5.3, moonshotai/Kimi-K3, Qwen/Qwen3.5-397B-A17B, openai/gpt-oss-120b.
• Vision is off on the card because the default model is text-only; the docs note that it can be turned on for a vision model such as moonshotai/Kimi-K3. Tool calls and Ask streaming stay on.
• No pricing metadata is added. Chrome and Firefox catalogs stay identical.

Files

• src/{chrome,firefox}/src/providers/provider-catalog.js: one new entry.
• src/{chrome,firefox}/icons/providers/runcrate.svg: Runcrate logo mark (no scripts, images or external refs).
• src/{chrome,firefox}/icons/providers/SOURCES.md: icon source row.
• test/run.js: id list, counts (85 extra cards, 120/119 defaults) and a defaults/suggestions check.
• test/provider-catalog-integrations.mjs: endpoint, model, flags and icon URL checks for both browsers.
• docs/providers-and-models.md: id table, counts and a short setup section.
• README.md: card count 119/118.

Testing

• npm run test:provider-catalog: provider catalog integration tests passed.
• npm run test:provider-limits: provider model limit tests passed.
• WEBBRAIN_TEST_FILTER='extended provider catalog' node test/run.js: 1 passed, 0 failed.
• node --check on both catalogs and the integration test: OK.
• Full node test/run.js: 2547 passed, 5 failed. The same 5 tests fail on main without this change (new locale registration, new locale dictionaries, streamed raw tool-call text, Providers page search box, WebGPU worker contracts).
• Live, with a real key, through OpenAICompatibleProvider built from ADDITIONAL_PROVIDER_DEFAULTS.runcrate:
  • testConnection() returned {"ok":true,"model":"deepseek-ai/DeepSeek-V4.1-Flash"}.
  • chat() returned a one-sentence answer from deepseek-ai/DeepSeek-V4.1-Flash with finish_reason stop.
  • chat() with a get_weather tool returned one tool call with arguments {"city": "Paris"}.
  • chatStream() yielded reasoning, text, usage and done events; the streamed text was 1 2 3 4 5.

This PR was prepared with AI assistance (omp with Claude): the catalog entry, tests, docs lines and this description were written with it, and the commands and live calls above were run as described.

Disclosure: I work on Runcrate.

[vercel]

@heydryft is attempting to deploy a commit to the esokullu's projects Team on Vercel.

A member of the Team first needs to authorize it.

[heydryft]

Thanks @esokullu for merging! Glad WebBrain users can run the agent on Runcrate models now. Happy to help if anything comes up.