Your first agent

Before you begin

• A balance

  Deploying reserves the monthly runtime line, so top up under Billing first. Billing

• A model key

  The agent thinks by calling models through your key. Create one with the invoke scope, which is the default. API keys

• Channel credentials

  Only if you want a chat surface — a bot token you create yourself. Headless agents need none. Chat surfaces

The create wizard, step by step

1. 1

   Where it lives

   Pick a chat surface for people-facing work, or headless if you will reach it from your own code and the control panel only.

2. 2

   Credentials

   Paste the channel token you created. Credentials are stored encrypted and injected into the runtime at start; they never appear in a log line.

3. 3

   Persona and model

   Write what the agent is for and how it should sound, then choose the default model. Both can be edited later without a redeploy.

4. 4

   Region and deploy

   Choose where the runtime should sit, confirm the monthly line, and deploy. Progress shows in the dashboard while it provisions.

Chat surfaces

A chat channel needs a bot token that you create on that platform, not one we can issue for you. Paste it into the wizard and the runtime holds it encrypted. Rotating the token later is a settings change, not a redeploy.

Headless / API

A headless agent exposes an endpoint instead of joining a chat room. Call it the way you call any model — the request shape is identical, so the same client library works.

curl https://api.agentplatform.store/v1/chat/completions \
  -H "Authorization: Bearer ap_sk_..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "atlas-5.6-solo",
    "messages": [
      { "role": "user", "content": "Summarise this changelog." }
    ]
  }'

Parking and waking

Parking an idle agent releases its container and stops the compute line entirely. Persona, notes and channel configuration are archived on the platform side, so waking it brings back the same agent rather than a fresh one. Expect it warm in about fifteen seconds.

Scheduled work

A schedule is a prompt plus a cadence. The runtime wakes if it is parked, does the work, writes what it found, and goes back to sleep. Failures are retried once and then surfaced in the dashboard rather than swallowed.

Spend

Two lines make up the bill: the flat monthly runtime charge and the per-token cost of the calls the agent makes. The usage log itemises every request, so a surprise is always traceable to a specific call.

Set a ceiling per key if you want a hard stop rather than a warning. See the usage view.