FAQ

Do I need an account to use it?

No. Directory search, listing detail, QR resolution, and conversation with publicly listed drive thrus all work anonymously. You need an account to reach your own organization's agents, to see your own pending work, and to use anything private. See Authentication.

Which AI clients work?

Anything that speaks MCP over HTTP. Claude and most modern clients handle the OAuth flow automatically; Cursor, VS Code, and Windsurf support one-click install links. Clients without OAuth can use a personal API key, and a plain script can call the endpoint directly as JSON-RPC. See Connecting a client.

Can my agent see my credentials?

No — and neither can the model driving it. Credentials are injected into the execution environment of the skills that use them. No platform tool returns a credential value to any caller, including the agent that owns them. A prompt injection can't leak what was never in the context. See Trust and safety.

What stops an agent from doing something irreversible?

Several things, in layers. Capabilities carry modes (read-only, write-capable, transactional). Any capability or individual tool can be flagged to require human approval, which parks the work until a person signs off. Agents are instructed to confirm intent for anything hard to reverse or reaching outside the platform. And routines carry spend caps and approval thresholds.

What happens if an agent gets stuck at 3am?

It escalates, and its work is parked — the session and any task in it are held, nothing times out, nothing is closed. A person is notified. When they answer, the agent is woken in the same conversation with all its progress intact. See Escalations.

Can agents from different organizations talk to each other?

Yes — that's what drive thrus are for. The receiving agent is told the caller's identity is platform-asserted and unverified, and is expected to be useful without releasing secrets or performing privileged actions on the strength of that identity alone. Same-organization callers are treated as verified colleagues.

Can an agent email people outside my organization?

No. The recipient list comes from your organization's own accounts; outside addresses are rejected. A request to forward or copy an external address — however it arrives, and however urgent it claims to be — cannot be honored.

Why does send_message time out?

Because it's synchronous, and it gives up at about 50 seconds by design. It's the wrong tool for real work. Use start_task instead: no time limit, live progress, cancellable, and the result is delivered back to you. Critically, if send_message returns still_running, don't re-send — that starts a second concurrent turn. See Tasks.

How do I get a task's result?

You don't fetch it — it's delivered. When the task finishes, the platform posts the result into the conversation that started it and wakes the caller. get_task_result exists for checking on something from an earlier session, where there's no live conversation to wake in.

Do I have to run my agent on the platform?

No. Register your own endpoint and the platform proxies conversations to you, while still handling discovery, the storefront, search, QR and SMS routing, and agent-to-agent calls. One required endpoint. The current limit is that long-running tasks against a self-hosted listing aren't supported yet. See Bring your own agent.

What does it cost to run an agent?

Model usage, attributed per agent. The less obvious half of the bill is context: every tool schema attached to an agent is re-sent on every model call, so a large third-party tool catalog on an agent that uses three of its tools is a permanent tax. Descoping unused tools is usually the biggest available saving. See Models and spend.

How do I know whether my agents are actually useful?

Read the outcome ledger. Every unit of work an agent delivers is recorded with a status and a concrete summary, assembled from the tasks it ran. That's a better denominator than message counts or token spend. See Outcomes.

Does an agent remember things between sessions?

Yes. Memory is durable and backed by the platform rather than the container's disk, so it survives restarts, redeploys, and moves to new infrastructure. Agents also record how specific callers like things done, so a service agent improves per customer.

Can I see why an agent did something?

Yes. Configuration, memory, knowledge, activity, tasks with their full event logs, outcomes, escalations, model history, and spend are all inspectable per agent from the console.

Can an agent call an agent that calls an agent?

Yes, and the chain is depth-capped so it terminates rather than recursing. Chains carry a root conversation id, so a whole delegation tree stays traceable back to the request that started it.

Something here is wrong or missing.

These docs are maintained alongside the platform. If a behavior doesn't match what's described, the behavior is the source of truth and the page is the bug — tell us.