Create an agent with a persona and tool allowlist, open sessions from your backend, stream replies over SSE - with optional sandbox jobs that build deliverables headless.
Four hops from your backend to a running agent.
Your backend calls with a scoped key; the API isolates the session by app_id, loads the persona and allowlist, and streams. The same harness that runs the workspace runs your agent.
Persona & allowlist
Voice, instructions, and exactly the tools this agent may call - nothing more.
Assist mode (SSE)
Replies stream token by token into your UI - the same loop your users see in chat.
Sandbox deliverables
Optional headless jobs build documents and reports, filed back to storage.
Session isolation
Every session scoped by app_id. No context crosses between your apps or tenants.
Two curls to a streaming session.
Create the agent once, open sessions per user. Everything the agent does shows up in the console - sessions, runs, credits, and the last time it spoke.
Full API reference# 1 - create the agent curl -X POST /v1/managed-agents/agents \ -H "Authorization: Bearer sk_live_…" \ -d '{ "persona": "concierge-v3", "tools": ["storage", "web"] }' # 2 - open a session, stream messages curl -N -X POST /v1/managed-agents/sessions \ -d '{ "agent_id": "agt_91xk", "app_id": "acme-web" }' event: message.delta data: {"text": "Welcome back - let's finish setup…"}
Onboarding concierge
Managed agent
412
sessions this week
Run #442 finished
deliverable saved to storage
Your product's agent, running tonight.
Get API keys after sign-in, or talk to us about enterprise isolation and residency.
Or email contact@sundaypyjamas.com - a founder replies.