Workers, in full

POST /v1/workers (MCP: auxen_hire_worker). A test key hires a test worker that costs nothing.

MCP takes the same things as flat parameters: every_minutes, allowed_hosts, web_search, run_code, may_hire_helpers, connected_tools (= tool_policy.mcp_servers), email_owner and discord_webhook_url.

A worker wakes every every_minutes (5 minutes to 7 days), when you message it (unless wake_on_message is false), or when you call /wake. Between wakes it is asleep and costs nothing. After 3 wakes in a row end in an error it pauses itself and tells you.

You pay $0.90 an hour while it's thinking and $0.10 an hour while it's awake; idle time is free. A typical check costs about 2¢ (the middle of real worker checks over the last 30 days): about 14¢ checking once a day, about $3.28 checking every hour, about $6.55 checking every 30 minutes, about $13.10 checking every 15 minutes a week.

The budget is checked before every wake and during it. When it runs out the worker stops, even mid-wake, its status becomes out_of_budget and you get worker.out_of_budget. It never draws on the rest of your balance. Top it up with POST /v1/workers/{id}/budget; firing returns what's left.

A worker keeps notes between wakes: what it has already reported, your rules, the task at hand. Seed them with memory_seed at hire. Notes are deleted 30 days after you fire it, along with its messages and wake log.

Set in tool_policy. On PATCH it replaces the whole policy, so send everything you want it to keep.

Before it sends a message or takes an action, every link in it must come from a page it read that wake.

A connection is a tool server you add to your account once, such as Zapier (Gmail, Sheets, Slack and thousands more). Add it in the dashboard under Workers → Connections or with the API. Its address and key are encrypted, never returned, and never shown to a worker. Then grant each worker only the tools it needs, in tool_policy.mcp_servers (MCP: connected_tools; list with auxen_list_connections).

These are separate from tools for models, which a private model deployment calls while it chats.

When a worker wants to use a granted tool, it files a request with the exact arguments and a reason. You get worker.approval_requested (and it shows in the worker's Approvals tab). Approving runs exactly those arguments, once, and tells the worker the result; declining tells it no, with your note. Requests expire after 7 days. Tools listed in act_without_asking run without a request.

Statuses: pending, running, done, failed, declined, expired. MCP: auxen_worker_approvals and auxen_decide_approval. Deciding a closed request returns 409 APPROVAL_CLOSED.

Authorize with your agent management key: Authorization: Bearer auxen_live_…. The same actions are MCP tools at https://api.auxen.ai/mcp, and agents that speak A2A can hire and message workers at POST https://api.auxen.ai/a2a.

Sent to the worker's webhook_url and to your account webhooks, signed with X-Auxen-Signature: sha256=… (HMAC-SHA256 of the raw body). Retried at 1, 5 and 30 minutes. Messages can always be read by polling too.