Start here

A reading path through the Cruxy API in nine pages, from your first call to tool calling, retrieval, moderation, and running in production.

Each step names one page and what you take from it. The pages build on each other, so read them in order the first time; after that, each one stands on its own as a reference.

1. 1

   Make your first call

   Quickstart takes you from an empty terminal to your first model response: you get an API key with device login, then send a chat request to the kavi tier.

2. 2

   Choose a tier

   Choosing a tier compares the tiers you pass as model (kavi, vaani, and mira) by profile, output cap, and how each one reasons by default, and shows how to read the live catalog from GET /api/v1/models instead of hardcoding it.

3. 3

   Hold a conversation

   Chat documents POST /api/v1/chat, the OpenAI-compatible chat completion: the request options, a single JSON completion or a Server-Sent Events stream, the cruxy block, and usage.

4. 4

   Let the model call your functions

   Tutorial: build a tool-using agent builds on the same endpoint. You describe two functions, run the calls the model returns, send the results back, and loop on finish_reason until the model answers.

5. 5

   Answer from your own documents

   Retrieval-augmented generation is a cookbook recipe that chains three endpoints: embeddings for a first-pass search, rerank to sharpen the shortlist, and chat to write the answer.

6. 6

   Screen content

   Moderations checks text against the content-safety categories with the cruxy-guard model and returns an overall flagged result.

7. 7

   Handle failures

   Errors and retries defines the error envelope, lists every error code, and says which failures to retry and how long to wait, so your client branches on the stable code, never on the wording of the message.

8. 8

   Run it reliably

   Reliability puts the production rules in one place: the per-tier deadlines, which failures to retry, Idempotency-Key, how to tell a finished stream from a failed one, and where to check platform status.

9. 9

   Know the limits

   Known limitations lists what the API does not do today, such as returning kavi's thinking or running a non-streaming call past 90 seconds, and the workaround for each.

Next steps

Cookbook has more recipes that combine endpoints to finish a real job, and API Reference lists every endpoint with its exact request and response shape.

Moving code that already calls OpenAI? Migrate from OpenAI changes the base URL, the key, and model, and OpenAI SDK compatibility says what carries over.