AI providers

Splice can call language models for agents and skills — with explicit routing, usage reporting and no hidden provider substitution.

Use it

splice ai models --search gpt-4o
splice ai generate "Explain EIP-1967 in one sentence" --max-tokens 80
splice ai generate "Reply with one word" --provider gemini
splice ai generate "Capital of France as JSON" --schema schema.json

From TypeScript:

const r = await splice.ai.generate({ prompt: "…", maxTokens: 200 });
if (isLive(r)) console.log(r.data.text, r.data.usage, r.data.routing);

Over MCP (with splice mcp --data): ai_generate and ai_models. Skills use the ai.generate host capability (Host capabilities); their output is capped at 2,048 tokens per call.

Routing

• A request uses the requested provider (or AI_PROVIDER, default openrouter) and the requested model (or the provider's default model). No model is hardcoded in Splice.
• Nothing else is tried unless AI_FALLBACK_PROVIDER and AI_FALLBACK_MODEL are set and the request does not pass fallback: false.
• Every answer carries routing: requested and actual provider and model, whether a fallback was used and why. OpenRouter's upstream provider (e.g. "OpenAI", "Azure") is reported too.

Limits and behaviour

• Up to 100 messages and 200,000 characters per request; maxTokens up to 16,384; up to 32 tools; images as https or data:image/* URLs (Gemini: data: URLs only).
• Timeout 60 seconds by default (maximum 120); one retry for network errors, 5xx and short rate-limit waits.
• Invalid requests (unknown model, bad schema) are PROVIDER_ERROR without fallback; missing credits or provider outages may use the configured fallback.
• Completions are never cached. Health checks use free metadata calls (OpenRouter /key, Gemini model list), never a completion.
• Streaming is not offered: results are complete objects with provenance.

Keys stay in the host process and are redacted from every result and error.