How do I get an API key?

Sign up for Cruxy and create a key with the device-login flow described in the quickstart: request a device code, approve it in the browser and poll for the key. API keys come from the same account as the app, and the API is billed per token, separately from your plan.

Which model should I pass?

Pass auto unless you have a reason not to. Mira is the fast size for short, high-volume calls, Vaani is the balanced default, and Kavi is the deep size for the hardest tasks. The model behind a size can change; the size id is the contract.

Any OpenAI client works for chat completions and models, in whatever language you already use. For Cruxy's own routes, the docs show the request in curl and TypeScript with fetch.

Is my data used for training?

Your prompts and outputs are not used to train AI models without your explicit, separate consent. Request logs keep the model, token counts, cost and timing, and none of the content. The privacy policy has the full account.

Developers

Build what comes next.

One base URL, one key and three model sizes behind it. An OpenAI-compatible API for chat, Cruxy's own routes when you want memory, vaults, files and evals, and a CLI that works in your repo.

Get started free Read the quickstart

import OpenAI from 'openai';

const cruxy = new OpenAI({
  baseURL: 'https://api.cruxy.ai/v1',
  apiKey: process.env.CRUXY_API_KEY,
});

const res = await cruxy.chat.completions.create({
  model: 'auto',
  messages: [{ role: 'user', content: 'Summarize this ticket' }],
});

import os
from openai import OpenAI

cruxy = OpenAI(
    base_url="https://api.cruxy.ai/v1",
    api_key=os.environ["CRUXY_API_KEY"],
)

res = cruxy.chat.completions.create(
    model="auto",
    messages=[{"role": "user", "content": "Summarize this ticket"}],
)

curl https://api.cruxy.ai/v1/chat/completions \
  -H "Authorization: Bearer $CRUXY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "auto",
       "messages": [{"role": "user",
                     "content": "Summarize this ticket"}]}'

Response · chat.completion

{
  "id": "chatcmpl-8f3a1c2d",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Checkout fails on UPI when the cart
          holds a gift card. Steps to reproduce and
          a fix are below."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 212,
    "completion_tokens": 96,
    "total_tokens": 308
  }
}

The request is real. The reply is illustrative and trimmed to its shape.

Base URL: api.cruxy.ai/v1
Models: kavi · vaani · mira · auto
Streaming: OpenAI's SSE, ends with [DONE]
Errors: A stable code on every one

The platform

Everything you can build on. All of it live, all of it documented.

Chat APILiveOpenAI-compatible chat completions on three model sizes. Streaming, tool calling and a reasoning_effort you set per request.Chat guide Cruxy CodeLiveA coding agent for your terminal. Reads your codebase, edits files with a diff before each approval, runs your tests and opens pull requests.Explore Cruxy Code Evals APILiveSuites of cases you define, runs you start, results you read back. Point an evaluation harness at the OpenAI-compatible routes, or use the evals routes directly.API reference Embeddings, rerank and moderationLivecruxy-embed turns text into vectors, cruxy-rerank orders candidates against a query, and cruxy-guard classifies text against content-safety categories.Cookbook DocsLiveA quickstart, a guide for every route, a cookbook with retrieval, tool loops and streaming in the browser, and the full API reference.Read the docs StatusLiveLive probes and incidents, on their own domain, so you can check us before you check your code.See status

Chat

Two routes, one set of models. Use the one your client already speaks.

Chat guide

POST /v1/chat/completions

OpenAI-compatible

Point any OpenAI client or evaluation harness at https://api.cruxy.ai/v1 with your Cruxy key. Standard chat.completion objects, OpenAI's SSE and OpenAI's error envelope, with a stable Cruxy code inside it.

temperature, top_p, max_tokens and reasoning_effort are honoured
Fields the gateway does not use are ignored, not refused, so a harness that sends them still runs
A reproducible run of your messages and the model. No memory, no vaults, no files

POST /api/v1/chat

Cruxy chat

The same models with everything the Cruxy app has: memory recall and capture, vaults and standing rules, files by id, and usage.cost on every reply.

memory: { recall, capture } per request, rules on or off, a vault_id
usage.cost with the tier, the amount and the currency on every response
cruxy.request_id to look the call up in your logs and pair it with feedback

const stream = await cruxy.chat.completions.create({
  model: 'auto',
  stream: true,
  messages: [{ role: 'user', content: 'Draft the release note' }],
});

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '');
}

stream = cruxy.chat.completions.create(
    model="auto",
    stream=True,
    messages=[{"role": "user", "content": "Draft the release note"}],
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

curl -N https://api.cruxy.ai/v1/chat/completions \
  -H "Authorization: Bearer $CRUXY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "auto", "stream": true,
       "messages": [{"role": "user",
                     "content": "Draft the release note"}]}'

STREAMING

Set stream to true and the reply arrives as OpenAI's server-sent events: chat.completion.chunk frames with content or tool-call deltas, then [DONE].

If the upstream fails mid-stream, the stream ends with an error frame instead of [DONE], so your client can tell a cut-off from a finish. Tool calling works on both routes: send the functions the model may call, run the ones it returns and feed the results back.

Tool calling guide

The API

Every route, one line each. Paths as the reference lists them.

Full API reference

GROUPROUTESWHAT IT DOESDOCS

Chat

POST /v1/chat/completions
POST /api/v1/chat
POST /api/v1/tokens/count

OpenAI-compatible chat completions, Cruxy's own chat with memory, vaults, rules and files, and a token counter.Chat guide

Models

GET /v1/models
GET /api/v1/models

The tier catalog, as an OpenAI list or with Cruxy's own metadata: context window, pricing, tool support and reasoning defaults.Model reference

Tool calling

"tools": [ ... ] on every chat route

Send the functions the model may call, run the calls it returns and feed the results back.Tool calling guide

Embeddings and rerank

POST /api/v1/embeddings
POST /api/v1/rerank

Turn text into vectors with cruxy-embed and order candidate documents against a query with cruxy-rerank.Embeddings guide

Moderation

POST /api/v1/moderations

Classify text against content-safety categories with cruxy-guard and act on the flagged boolean.Moderation guide

Files

POST /api/v1/files
GET /api/v1/files/{id}/content

Upload, list, download and delete files, then attach them to Cruxy chat requests by id.API reference

Evals

POST /api/v1/evals/suites
POST /api/v1/evals/suites/{id}/runs

Suites of cases you define, runs you start, and results you read back.API reference

Feedback

POST /api/v1/feedback

Report the outcome of a prior call by its request id: accepts, rejects, thumbs and test results.Feedback guide

Memory, vaults and rules

GET /api/v1/memories
POST /api/v1/vaults
PUT /api/v1/rules

Read and delete memories, manage vaults and their members, and set standing rules for an account or a vault.API reference

Limits and logs

GET /api/v1/limits
GET /api/v1/logs/{request_id}

Your per-key and per-organisation request ceilings, and the log of your requests by request id.Errors reference

Keys and organisation

POST /api/v1/device/code
GET /api/v1/org/members
POST /api/v1/workspaces

Device login for a key, your organisation's members and invitations, and workspaces.Authentication guide

Operate

Limits, logs and errors you can reason about.

Errors reference

Limits you can read

GET /api/v1/limits returns your per-key and per-organisation chat request ceilings for each 60-second window, and the gateway sends X-RateLimit headers. There is no tokens-per-minute limit.

Two kinds of 429

rate_limited means requests are arriving too fast; it clears once you slow down. budget_exhausted is a spend ceiling that a retry alone will not clear.

Logs by request id

Every request is logged with the model, token counts, cost, timing and status, and none of the prompt or the reply. Read one back at GET /api/v1/logs/{request_id}.

Keys shown once

Get a key with the device-login flow. It is shown exactly once and stored only as a hash. Secrets at rest are encrypted field by field with AES-256-GCM.

Live

Cruxy Code

The same models, in your terminal.

Plans multi-step changes and shows every diff before you approve it
Asks before every edit and every command, by default
Runs your test command, reads the failures and keeps fixing
cruxy rollback restores the tree to before any turn
cruxy pr opens the pull request after one approval

npm i -g @cruxy/cliExplore Cruxy Code

~/web · Cruxy Code

$cruxy run "fix the failing tests in the auth module"
cruxy › fix the failing tests in the auth module
! cruxy wants to run tests: pnpm test (destructive)
  $ pnpm test
  in /home/dev/web
  [y] once · [a] always allow test runs of exactly `pnpm test` this session · [t] reject & instruct · [n] no: a
✗ tests failed · 2 of 42 failed · 3.1s · pnpm test
  ✗ signs a session token src/auth/session.test.ts:18
      expected  "ok"
      received  undefined
  ✗ rejects an expired token src/auth/session.test.ts:41
  why (inferred)  TOKEN_SECRET is not set in the test environment
✓ read_file      8ms  src/test/setup.ts
? cruxy wants to edit: src/test/setup.ts (mutate)

A real session, cut short. The CLI output is real; the model's replies are illustrative.

API pricing

Per token, in USD. No setup fee and no minimum.

API pricing in full

Fast· Mira

Short, high-volume calls where latency and cost matter most.

INPUT: $0.20 / 1M
OUTPUT: $0.40 / 1M

Context: 64K
Max output: 2K

Balanced· Vaani

Everyday calls: strong quality at lower latency and cost.

INPUT: $0.65 / 1M
OUTPUT: $1.30 / 1M

Context: 128K
Max output: 4K

Deep· Kavi

The hardest tasks: deep reasoning and the largest output budget.

INPUT: $0.76 / 1M
OUTPUT: $1.52 / 1M

Context: 256K
Max output: 8K

USD per 1M tokens, metered per request. API prices exclude GST. 18% GST is added for India-billed accounts (input credit available). With auto, you pay the rate of the model that served the call.

Questions

Good to know.

Moderation on any keyScreen input with cruxy-guard before it reaches chat. The moderation API takes any key on any plan.

Yes, for chat and the model list. https://api.cruxy.ai/v1 serves /v1/chat/completions and /v1/models in OpenAI's shapes, so existing clients and evaluation harnesses work once you change the base URL and the key. The rest of the API, from files to evals, is Cruxy's own and lives under /api/v1.

Change the base URL. Keep your code.

Start free, pay per token, and read every limit before you hit it.

Get started free API reference

import OpenAI from 'openai'; const cruxy = new OpenAI({ baseURL: 'https://api.cruxy.ai/v1', apiKey: process.env.CRUXY_API_KEY, }); const res = await cruxy.chat.completions.create({ model: 'auto', messages: [{ role: 'user', content: 'Summarize this ticket' }], });

import os from openai import OpenAI cruxy = OpenAI( base_url="https://api.cruxy.ai/v1", api_key=os.environ["CRUXY_API_KEY"], ) res = cruxy.chat.completions.create( model="auto", messages=[{"role": "user", "content": "Summarize this ticket"}], )

curl https://api.cruxy.ai/v1/chat/completions \ -H "Authorization: Bearer $CRUXY_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model": "auto", "messages": [{"role": "user", "content": "Summarize this ticket"}]}'

{ "id": "chatcmpl-8f3a1c2d", "object": "chat.completion", "choices": [ { "index": 0, "message": { "role": "assistant", "content": "Checkout fails on UPI when the cart holds a gift card. Steps to reproduce and a fix are below." }, "finish_reason": "stop" } ], "usage": { "prompt_tokens": 212, "completion_tokens": 96, "total_tokens": 308 } }

const stream = await cruxy.chat.completions.create({ model: 'auto', stream: true, messages: [{ role: 'user', content: 'Draft the release note' }], }); for await (const chunk of stream) { process.stdout.write(chunk.choices[0]?.delta?.content ?? ''); }

stream = cruxy.chat.completions.create( model="auto", stream=True, messages=[{"role": "user", "content": "Draft the release note"}], ) for chunk in stream: print(chunk.choices[0].delta.content or "", end="")

curl -N https://api.cruxy.ai/v1/chat/completions \ -H "Authorization: Bearer $CRUXY_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model": "auto", "stream": true, "messages": [{"role": "user", "content": "Draft the release note"}]}'

$cruxy run "fix the failing tests in the auth module" cruxy › fix the failing tests in the auth module ! cruxy wants to run tests: pnpm test (destructive) $ pnpm test in /home/dev/web [y] once · [a] always allow test runs of exactly `pnpm test` this session · [t] reject & instruct · [n] no: a ✗ tests failed · 2 of 42 failed · 3.1s · pnpm test ✗ signs a session token src/auth/session.test.ts:18 expected "ok" received undefined ✗ rejects an expired token src/auth/session.test.ts:41 why (inferred) TOKEN_SECRET is not set in the test environment ✓ read_file 8ms src/test/setup.ts ? cruxy wants to edit: src/test/setup.ts (mutate)

Build what comes next.

Everything you can build on. All of it live, all of it documented.

Two routes, one set of models. Use the one your client already speaks.

OpenAI-compatible

Cruxy chat

Every route, one line each. Paths as the reference lists them.

Limits, logs and errors you can reason about.

Limits you can read

Two kinds of 429

Logs by request id

Keys shown once

The same models, in your terminal.

Per token, in USD. No setup fee and no minimum.

Good to know.

Is the API OpenAI-compatible?

How do I get an API key?

Which model should I pass?

Do you have an SDK?

Is my data used for training?

Change the base URL. Keep your code.

Build what comes next.

Everything you can build on. All of it live, all of it documented.

Two routes, one set of models. Use the one your client already speaks.

OpenAI-compatible

Cruxy chat

Every route, one line each. Paths as the reference lists them.

Limits, logs and errors you can reason about.

Limits you can read

Two kinds of 429

Logs by request id

Keys shown once

The same models, in your terminal.

Per token, in USD. No setup fee and no minimum.

Good to know.

Is the API OpenAI-compatible?

How do I get an API key?

Which model should I pass?

Do you have an SDK?

Is my data used for training?

Change the base URL. Keep your code.