Adapters
Five LLM adapters ship with v0.1, covering 14+ providers in total. A sixth private package (adapter-contract-tests) is not published; it's the shared conformance suite every adapter must pass.
alpha.30 adds two subprocess-driven agent adapters — @llm-ports/adapter-codex and @llm-ports/adapter-aider — plus a companion observability adapter, @llm-ports/telemetry-otel, that turns any ObservabilitySink slot into an OpenTelemetry gen_ai bridge.
| Adapter | Covers | Implements |
|---|---|---|
@llm-ports/adapter-anthropic | Anthropic Claude (Opus, Sonnet, Haiku) | LLMPort |
@llm-ports/adapter-openai | OpenAI + 10+ compat providers via baseURL (Azure OpenAI, Groq, Together AI, Fireworks AI, DeepInfra, Perplexity, Cerebras, LiteLLM proxy, ...) | LLMPort, EmbeddingsPort |
@llm-ports/adapter-google | Google Gemini (Pro, Flash) | LLMPort |
@llm-ports/adapter-ollama | Local LLMs via the Ollama daemon | LLMPort, EmbeddingsPort, model management |
@llm-ports/adapter-vercel | Migration from @ai-sdk/* | LLMPort, EmbeddingsPort (limited multimodal in v0.1) |
@llm-ports/adapter-codex | OpenAI Codex CLI (subprocess) | LLMPort (runAgent only) |
@llm-ports/adapter-aider | Aider CLI (subprocess) | LLMPort (runAgent only) |
@llm-ports/telemetry-otel | OpenTelemetry gen_ai semconv bridge (companion) | ObservabilitySink factory |
Feature matrix
✓ = supported, ✗ = not supported, ✓* = model-dependent, n/a = doesn't apply.
On the two chat methods. generateChat and streamChat hand tool calls back to the caller instead of executing them, for a caller that owns the loop. They are optional port methods and only adapter-openai implements them today, which covers the OpenAI-compatible surfaces that mostly want this. The registry knows which aliases implement them and filters a fallback chain accordingly, so a mixed chain works and a chain with no support throws a named error rather than failing mid-call. Every other adapter offers the same capability through runAgent, which runs the loop for you. See Tool use.
| Feature | Anthropic | OpenAI | Ollama | Vercel |
|---|---|---|---|---|
| Text generation | ✓ | ✓ | ✓ | ✓ |
| Structured output (Zod schema) | ✓ | ✓ | ✓* | ✓ |
| Streaming text | ✓ | ✓ | ✓ | ✓ |
| Streaming structured (partial JSON) | ✓ | ✓ | ✓ | ✓ |
| Tool use | ✓ | ✓ | ✓* | ✓ (multi-turn via Vercel's own loop) |
generateChat / streamChat (tool calls returned, not executed) | ✗ | ✓ | ✗ | ✗ |
generateStructured from a JSON Schema (jsonSchema) | ✓ | ✓ | ✓ | ✓ |
| Tool parameter schemas advertised to model | stub** | stub** | stub** | via Vercel SDK |
| Vision input (base64) | ✓ | ✓ (data URI) | ✓* | ✓ (data URI) |
| Vision input (URL) | ✓ | ✓ | ✗ (Ollama doesn't fetch URLs) | ✓ |
| Audio input | ✗ (Anthropic chat) | ✓ (wav, mp3) | ✗ | ✓ (base64 only) |
| Audio output | ✗ | ✓ | ✗ | ✗ |
| Prompt caching | ✓ native | partial (cached_tokens reported) | n/a | via Vercel |
| Embeddings | ✗ | ✓ | ✓ | ✓ (via Vercel) |
| Batch embeddings | n/a | ✓ | ✓ | ✓ |
| Model management (list/pull/delete) | ✗ | ✗ | ✓ | ✗ |
| Cost tracking (pricing tables) | ✓ | ✓ | ✓ (zero-cost default) | ✓ (user-supplied pricing) |
baseURL override (compat providers) | ✓ | ✓ (primary feature) | n/a (always local) | n/a |
** Tool parameter schemas advertised to model — "stub" means: in v0.1 the Anthropic, OpenAI, and Ollama adapters convert your Zod inputSchema to a generic { type: "object", properties: {} } shape before sending the tool definition to the model. The Zod schema still validates execute's input at runtime; only the model-facing tool advertisement loses structural information. Until #1 lands, name parameters explicitly in the tool's description string. The Vercel adapter uses the Vercel SDK's own schema handling, which preserves the schema shape.
Gaps to address in v0.2
The full inventory — including workarounds and tracking — is on the v0.1 status page.
- All adapters: full Zod-to-JSON-Schema conversion for tool parameters (#1).
- All adapters:
onRetryobservability hook for capability-rejection / transient-401 / reasoning-starved retries (#3). - Vercel: reasoning-model headroom multiplier (#4) and typed
EmptyResponseError(#5). - Ollama: pre-emptive
streamStructuredhandling once Ollama's API exposes partial JSON natively. - OpenAI: prompt caching first-class once OpenAI ships its caching API surface.
- Anthropic: embeddings, if and when Anthropic ships an embedding model.
- Vercel: full multimodal content-block round-trip (currently limited to text via stringify on the way in).
v0.3 adds
@llm-ports/adapter-googlefor Gemini (different API shape)@llm-ports/adapter-bedrockfor AWS users (Claude on Bedrock, Titan embeddings)
Writing your own
If you need a provider not listed here, see Custom adapters →.
Every adapter must pass @llm-ports/adapter-contract-tests. The shared suite asserts adapter-agnostic invariants (port shape, content-block round-trip, cost field populated, error mapping). New invariants added there automatically apply to every adapter on the next CI run.