Providers

Relay-E uses the Vercel AI SDK underneath, so adding a provider is wiring an adapter into the registry. Six providers are configured as of v0.1.0.

What's wired

Set at least one. The engine throws a clear error at startup if none are configured.

OpenRouter — the recommended cloud path

OpenRouter aggregates 100+ models behind a single OpenAI-compatible API. With one key, you can route between Claude, GPT, Llama, Mistral, and more — useful when you want flexibility without managing a key per provider.

OPENROUTER_API_KEY=sk-or-v1-...

OpenRouter model keys use the openrouter:<provider>/<model> convention:

modelKey: "openrouter:anthropic/claude-sonnet-4"
modelKey: "openrouter:openai/gpt-4o"
modelKey: "openrouter:meta-llama/llama-3.3-70b-instruct"

OpenRouter's pricing is per-model. The engine doesn't pre-bake every model's price — set yours in packages/providers/src/pricing.ts for the ones you use.

Anthropic — for Claude direct

ANTHROPIC_API_KEY=sk-ant-...

Claude direct gets you:

• Prompt caching (huge cost win on stable system prompts — see Token Optimisation)
• Lower latency (no aggregator hop)
• Specific features like extended thinking

Defaults bundled:

OpenAI — for GPT direct

OPENAI_API_KEY=sk-...

OpenAI is also the embedding provider. text-embedding-3-small (1536 dims) is a sane default; switch to voyage-3 for better retrieval quality (separate Voyage API key).

Ollama — fully offline

For air-gapped dev or compliance-heavy environments:

docker compose --profile local-llm up -d
docker exec relay-e-ollama ollama pull llama3.2
docker exec relay-e-ollama ollama pull nomic-embed-text

OLLAMA_BASE_URL=http://localhost:11434

Local models are slower and lower-quality than frontier ones, but useful for:

• Local development without burning API credits
• Air-gapped self-hosted deployments
• Cheap classification / routing models

Tier-based routing

You don't pick a model per request — you pick a tier in the skill:

defineSkill({
  name: "...",
  preferredTier: "balanced",  // "fast" | "balanced" | "premium"
  ...
});

The router resolves the tier against the configured providers in priority order:

Anthropic ▶ OpenRouter ▶ OpenAI ▶ Ollama

If Anthropic is configured, "balanced" → claude-sonnet-4-6. If only OpenRouter is configured, "balanced" → openrouter:anthropic/claude-sonnet-4. Routing is in packages/providers/src/registry.ts.

You can also override per request:

curl ... -d '{"prompt":"...","modelKey":"claude-haiku-4-5-20251001"}'

MiniMax

MiniMax is an OpenAI-compatible Chinese LLM provider.

MINIMAX_API_KEY=your-key

Moonshot / Kimi

Moonshot AI (Kimi) is an OpenAI-compatible Chinese LLM provider with large context windows.

MOONSHOT_API_KEY=your-key

Adding a new provider

The registry is small (~100 lines). To add a provider:

1. Add the package: npm i -w @relay-e/providers your-ai-sdk-provider
2. In packages/providers/src/registry.ts:
   • Create the client in the ProviderRegistry constructor.
   • Add models to CHAT_MODELS with provider: "yourname".
   • Add a case to providerFor(id).
3. In pricing.ts, add per-model pricing.
4. Done — routeBy() and chat() will use it automatically.

Embeddings

Separate from chat. Configured via EMBED_MODELS:

The schema's embedding columns are 1536-dimensional by default — match this when picking your embedding model, or migrate the column when you change.

Cost-saving tips

See Concepts → Agent Loop for cost accounting, and Reference → Token Optimisation for the full playbook.