Cruxy Platform
One gateway for every model tier
An OpenAI-compatible LLM gateway for the Cruxy model tiers: chat, embeddings, and moderation behind one API key.
curl https://api.cruxy.ai/api/v1/chat \
-H "Authorization: Bearer $CRUXY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "kavi",
"stream": true,
"reasoning_effort": "none",
"max_tokens": 256,
"messages": [
{ "role": "user", "content": "Say hello, then describe what the Cruxy API does in one sentence." }
]
}'Platform
Choose how you build
Two ways into the same gateway: call the endpoints directly, or start from a pattern and copy what fits.
The gateway surface itself: five endpoints behind one key. Each row links the guide that covers it.
- ChatPOST /api/v1/chat
- EmbeddingsPOST /api/v1/embeddings
- RerankPOST /api/v1/rerank
- ModerationsPOST /api/v1/moderations
- FeedbackPOST /api/v1/feedback
- API ReferenceOpenAPI 3.1
Patterns to copy, terms defined, and the whole API condensed for an agent context window.
- Cookbookrecipes
- API guide for agentsagent context
- Glossarydefinitions
- Changelogversions
Developer journey
From key to production
Every integration walks the same path; each stage links into the guide that covers it. To read it as one ordered path, follow Start here.
Model tiers
Pick a model
You pass a tier id as model (or auto to let the router choose), and the
gateway routes the call. The tier is the contract; the upstream model behind it
is never exposed. Choosing a tier compares
them, and Models documents every id and alias.
Deep
Deepest reasoning and the largest output budget; reach for it on the hardest tasks.
- Context window
- 256K tokens
- Max output
- 8,192 tokens
- Latency target
- 4,000 ms
kaviBalanced
Strong quality at lower latency and cost, a sensible default for everyday calls.
- Max output
- 4,096 tokens
vaaniFast
Lowest latency and cost for short, high-volume calls; the budget-downgrade target.
- Max output
- 2,048 tokens
miraCookbook
Start from a recipe
Four working patterns that combine the endpoints to do a real job; copy the one shaped like yours. Cookbook indexes them all.
- Retrieval-augmented generation
Embed a corpus, rerank the shortlist, and ground the answer in what wins.
- Tutorial: build a tool-using agent
Send tools, run the calls the model returns, and loop until it stops.
- Streaming chat in the browser
Consume the SSE stream and render deltas without jank.
- Moderation gate
Screen input with cruxy-guard before it ever reaches chat.