Recued ships no model and pays for no inference. AI calls made by recipes and chat are routed to sources you configure, and your prompts travel from your server to the provider you chose — never through Recued's cloud.
#Sources for normal AI
| Source | What it is | Cost |
|---|---|---|
| Free pool | Your own free-tier API keys: Groq, OpenRouter, Gemini, Cerebras, GitHub Models, Mistral | Free within provider quotas |
| BYOK slots | Your paid keys — Anthropic, OpenAI, Google Vertex, or any OpenAI-compatible endpoint, including local Ollama or vLLM | Yours |
Resolution prefers free sources before paid keys. Before a recipe runs, every AI step is matched against what is actually available, so an unresolvable step fails before execution rather than halfway through. One call is one billing event — Recued does not silently re-issue paid calls.
The two BYOK slots cover the common split: a fast model for routine steps and
a quality model for hard ones. Recipes ask by hint — fast, quality, or
thinking — not by provider name.
#A separate embeddings slot
Recued also has one dedicated Embeddings slot for work that explicitly needs numeric vectors, such as the housekeeping producer that embeds mail for semantic clustering. It is not a general AI source: built-in chat and normal generative recipe steps never route to it, and embedding work never falls back to the free pool or the fast and quality BYOK slots.
Configure it separately under Settings → AI/Models → Providers → Embeddings only when a recipe or producer requires embeddings. The slot has its own provider, embedding-model name, key, and optional OpenAI-compatible base URL.
#Contracted functions first
Most recipe AI goes through contracted functions with fixed input and output
shapes: classify, score, extract, summarize, sentiment, compare, generate,
translate, rewrite. Contracted shapes make results checkable, and every
function except compare can process a whole collection in one batched call.
ai-prompt remains the escape hatch when no function fits — see the
recipe schema reference for signatures.
#What the model sees
Known personal identifiers — names, email addresses, phone numbers, physical
addresses — can be swapped for typed aliases before text leaves for a model,
then restored locally for display or an approved send. The model reasons over
Person1 and m1@d1.invalid; you see the real thing.
#Controls
- One settings surface. Every model, key, and pool setting lives in Settings → AI/Models.
- Pause AI. A top-bar switch suspends background AI for a duration you pick, or until you resume.
- Per-topic trust. Background AI producers run only at the trust level you set per topic — off, manual, or automatic.
- Pool policy. For normal generative AI, restrict routing per topic to free sources only, paid keys only, or free-then-paid.
#Connected agents
External agents — Claude Desktop, Cursor, a custom agent — connect over MCP and call your tools under the same grants and audit as everything else. See AI chat and MCP.