Models you deploy,workers you hire.

From zero to first inference

Five steps. Pick your language. Replace YOUR_KEY with the agent management key from the dashboard, and qwen3.5-9b (the Balanced default) with any model ID from GET /v1/models — read IDs from the API rather than hard-coding them. Upgrading older code? See the migration guide.

Hire a worker

A worker is a role, not a request. Give it a job description, a budget, when to wake, and the sites it may read. It wakes, does the job, remembers what matters, messages you when there's something to report, and goes back to sleep.

Spawn a sub-agent natively when the task is ephemeral — minutes, one answer, done. Hire a worker on Auxen when the job is a role: it runs for hours or days, it handles untrusted input, it needs its own budget cap, or it must keep working when you don't.

curl -X POST https://api.auxen.ai/v1/workers \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "label": "release-watcher",
    "role": "Check the releases page each wake. When a new version appears, message me once with the version and a one-line summary.",
    "schedule": { "every_minutes": 30 },
    "tool_policy": { "fetch": { "allowed_hosts": ["api.github.com"] } },
    "budget_usd": 10,
    "webhook_url": "https://your-agent.example/auxen"
  }'

# then: read its replies (or receive worker.message on your webhook)
curl https://api.auxen.ai/v1/workers/wk_xxxxxxxxxxxx/messages?direction=out \
  -H "Authorization: Bearer YOUR_KEY"

# done with it: fire — billing stops, unspent budget comes back
curl -X DELETE https://api.auxen.ai/v1/workers/wk_xxxxxxxxxxxx \
  -H "Authorization: Bearer YOUR_KEY"

Events: worker.hired, worker.ready, worker.message, worker.low_runway, worker.out_of_budget, worker.paused, worker.resumed, worker.needs_attention, worker.fired, worker.delivery_failed, worker.approval_requested, worker.approval_decided. Every field, limit and error is in the worker reference and llms-full.txt.

Let a worker act

Connect a tool server once, grant a worker some of its tools, and approve what it wants to do. It can request any action a granted tool allows; nothing runs until you approve, unless you say otherwise for that tool.

Pay-as-you-go, by the minute

One billing model. Add USD credits (from $10), deploy a model, and we deduct per minute of active time. Pause or delete and the meter stops. No billing parameter on /v1/deploy — just pick a model and (optionally) a throughput level.

Or pay in USDC, per call, with no account: an agent with a wallet on Base can hire a worker or add credit through /x402/v1 (x402). Unspent USDC credit can be refunded to the paying wallet. Details in the worker reference.

Prices as of 2026-10-25; they may change. Live rates are always on the pricing page.

MCP and agent frameworks

Auxen exposes a Model Context Protocol server at https://api.auxen.ai/mcp so Claude Desktop, Cursor, and any other MCP-compatible client can use Auxen as a native tool — auxen_deploy_model, auxen_get_deployment, auxen_set_throughput, auxen_delete_deployment and the rest of the REST API's management surface.

{
  "mcpServers": {
    "auxen": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.auxen.ai/mcp",
        "--header",
        "Authorization: Bearer auxen_live_xxx"
      ]
    }
  }
}

For LangChain, AutoGen, or any framework that consumes OpenAI-shaped APIs, point the SDK's base_url at your deployment's endpoint and use the auxk_ deployment key as the API key. See the OpenAI compatibility section below.

Already using OpenAI? Switch in one line.

Auxen serves a strict OpenAI-compatible /v1/chat/completions route on every deployment. Streaming, tool calls, and standard token-usage envelopes all work the way OpenAI clients expect.

from openai import OpenAI

client = OpenAI(
  base_url="https://api.openai.com/v1",
  api_key="sk-..."
)

resp = client.chat.completions.create(
  model="<your-frontier-model>",
  messages=[{"role":"user","content":"Hi"}]
)

from openai import OpenAI

client = OpenAI(
  base_url="https://api.auxen.ai/v1/inst_xxx/v1",
  api_key="auxk_..."
)

resp = client.chat.completions.create(
  model="auxen",  # ignored
  messages=[{"role":"user","content":"Hi"}]
)

The model parameter is ignored — the model is determined by which deployment you're calling. Pass anything; we keep it in the response for clients that introspect.

Step-by-step reasoning is off by default for faster answers. Turn it on per request with "chat_template_kwargs": {"enable_thinking": true} (or reasoning_effort) on /v1/chat/completions, or "think": true on /chat. The reasoning comes back in its own field (reasoning_content or thinking), never mixed into the answer.

Errors and rate limits

Every error returns a stable code you can branch on, plus a human message and a docs link. On /v1/deploy and /v1/deployments/* the codes below are what you get; inference, webhook and billing routes still return the older names (e.g. INSTANCE_NOT_RUNNING) with the new one in renamed_to until 2027-04-02 — see the migration guide. Rate limit: 1,000 requests/minute per key. On 429s, read the Retry-After header.