Skip to content

ai sdk providers

Jump to bottom
Thibaut Rey edited this page Sep 7, 2026 · 1 revision

Cloud providers through AI SDK

MultiVibe uses the AI SDK provider interface for additional API-key providers. The initial registry includes Anthropic, Google Gemini, OpenRouter, DeepSeek, Groq, Together AI, Cerebras, and Perplexity. Existing ChatGPT/OpenCode/Grok OAuth, Mistral, z.ai, local runtimes, and custom compatible endpoints retain their existing integrations.

The provider picker exposes More cloud providers. Select a provider, enter its API key, and optionally restrict the account to comma-separated upstream model IDs. An empty list uses the bundled models.dev text-generation catalog. Model names are prefixed with the provider ID, for example anthropic/claude-sonnet-4-6, google/gemini-2.5-flash, or openrouter/anthropic/claude-sonnet-4.6. Aliases can refer to these names as usual. Manual model IDs allow newly released models before the next catalog update.

Request path

Rust remains the public API edge and owns authentication, account selection, retry/failover, tracing, and client protocol conversion. For accounts whose provider is ai-sdk, it calls the internal SDK adapter hosted by Node. The adapter reads the selected account and instantiates an AI SDK language model:

  • @ai-sdk/anthropic for Anthropic's native Messages API.
  • @ai-sdk/google for Google's native generation API.
  • @ai-sdk/openai-compatible for the registered compatible providers.

The adapter uses the SDK's standardized LanguageModelV4.doGenerate and doStream contracts. It does not implement provider-specific HTTP codecs, use Vercel AI Gateway, execute tools, or perform a second retry loop. The same adapter is available to the standalone Node proxy. Existing providers incur no new hop.

The internal adapter requires V1_EDGE_INTERNAL_JOB_TOKEN in the native profile. Standalone Node generates a process-local token. Provider keys never authorize adapter calls; the adapter loads them from the account store and only sends them to the provider's registered endpoint. Rust's public fallback refuses adapter paths. Remote catalog data cannot select executable packages or change endpoints.

Supported generation features are text, supported image inputs, function calls and tool results, JSON output formats, reasoning effort, common generation controls, token usage, and incremental streaming. The existing edge converts Chat Completions, Responses, and Messages requests to/from the adapter's chat surface. Unsupported input structures return a clear error. Provider-specific options can be supplied as provider_options on chat requests; provider support still governs the available features. This adapter does not provide embeddings, audio generation/transcription, image generation, realtime, provider-executed tools, or arbitrary provider credentials such as AWS signing or Azure tenancy.

Upstream HTTP errors are returned before committing streaming headers. Later stream errors emit an error frame without a fabricated successful finish. Disconnects cancel generation, and response writes respect backpressure.

Model metadata

npm run catalog:sdk:refresh downloads https://models.dev/api.json and updates src/ai-sdk/catalog.generated.ts. The catalog is bundled: startup and inference do not depend on models.dev availability. Review and commit refreshes. Only non-deprecated text-generation entries for the installed registry are included.

GET /admin/provider-catalog serves provider choices and model metadata behind admin authentication. Model discovery preserves context/output limits, tool and reasoning capabilities, input modalities, source timestamp, and pricing metadata. Pricing is models.dev's advertised USD per million tokens, not measured billing or an account-specific quote. Catalog inclusion does not prove entitlement or current availability. Upstream failures remain authoritative.

Input modalities describe the upstream catalog; the adapter currently accepts text and supported image inputs only.

Neither AI SDK nor models.dev provides account subscription quotas. These new accounts report quota windows as unsupported; request token usage remains in normal traces and statistics. Provider-specific quota integrations can be added separately when a documented authenticated API exists.

Extending the registry

Add a reviewed provider in src/ai-sdk/providers.ts. Compatible providers need an ID, label and fixed endpoint. Native protocols require an installed SDK adapter factory in models.ts plus its pinned package dependency. Add the ID to the catalog refresh script and refresh metadata. Add request/response fixtures and account-isolation tests before exposing the new provider in setup. A catalog entry alone never implies inference support.

Validation

Worktree checks passed: git diff --check and syntax checking the catalog refresh script. Dependencies were installed and validation run on local main: web/API builds, 41 targeted TypeScript tests (including OpenCode and quota regressions), and the Rust ai_sdk integration test using Rust 1.88. The native test covers discovery, Chat Completions, Responses, streaming, Messages, and internal-route isolation. The protocol tests passed again after account-ID validation was added. Browser checks verified provider selection, review, saved vendor labels, and unavailable quota explanations. SDK codec tests use mock upstream responses; live paid-provider inference and production deployment were not performed.

Sources checked during implementation:

Migrated from docs/ai-sdk-providers.md on 2026-09-07.

Documentation home · Source code · Releases · Apache 2.0 license

MultiVibe documentation

Clone this wiki locally