GeezerVison

Thanks for building Sim: the shared openai-compat helper and the add-model skill made this a small, clean addition.

Summary

Adds Runcrate as a BYOK model provider for the Agent block. Runcrate is an AI cloud with an OpenAI-compatible inference API at https://api.runcrate.ai/v1, so the new provider goes through the shared openai-compat chat-completions helper, the same way Nebius does.

• Catalog models are namespaced runcrate/<org>/<model>: DeepSeek V3.2, Gemma 3 27B, Llama 3.3 70B, MiniMax M3, Qwen3 235B, Qwen3 Coder 480B and GLM 5.3. Ids, context windows and per-million-token pricing were taken from GET https://api.runcrate.ai/v1/models.
• Strict JSON schema response format is sent natively (checked against the live API on three of the models).
• Forced tool choice is only sent to models that honor it (DeepSeek V3.2, Gemma 3 27B, GLM 5.3). The API answers 400 for a named tool_choice on the others, so those fall back to auto, the same pattern Nebius uses.
• BYOK wiring: key lookup, settings row, API contract, usage-page icon and a new RuncrateIcon.
• Attachments are disabled for this provider.
• Docs: agent.mdx and costs.mdx.

Product update (if relevant)

Workspaces can add their own Runcrate API key and use Runcrate-hosted open models in Agent blocks. On self-hosted Sim, the key goes in the Agent block's API Key field.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• Other: ___________

Testing

Run from apps/sim unless noted, after bun install --ignore-scripts (the isolated-vm native build failed on my machine, which does not affect these checks):

• bunx vitest run providers/runcrate providers/utils providers/models providers/attachments providers/history-adapters lib/api-key: 14 files, 359 tests passed (the 9 new providers/runcrate/index.test.ts tests: wire contract and auth header, case-insensitive model resolution, strict JSON schema, forced-tool downgrade for four models, SSE reasoning and usage with catalog pricing, missing key). The new Runcrate cases in lib/api-key/byok.test.ts pass with the rest of that file.
• bun run type-check (in apps/sim): clean.
• bunx biome check on the touched files: no findings.
• From the repo root: bun run check:byok-providers, check:icon-paths, check:bare-icons, check:file-names and check:source-conventions pass.
• Live API: I sent a handful of small chat requests with a Runcrate key to confirm the tool-choice and JSON schema behavior above. I did not add a live acceptance test or run the app by hand.

Reviewers may want to look at the per-model forcedToolUse flags and the icon, which I drew as a simple SVG of the brand mark.

Disclosure: I work on Runcrate.

This PR was prepared with AI assistance (omp with Claude): the code, tests, docs and this description were written with it, and the commands above were run as part of preparing it.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

Screenshots/Videos