Why a daemon

A terminal is not an agent host. A broker is the detached process that hosts one node’s agent engine and session; a viewer is only a terminal presentation attached to that broker. The daemon, crtrd, owns the durable canvas and starts, stops, and recovers brokers. This separation lets a node continue after the terminal that created or displayed it has gone away.

flowchart LR
  CLI[crtr CLI] --> API[/v1 API]
  SDK[SDK application] --> API
  Viewer[Terminal viewer] --> API
  API --> Daemon[crtrd]
  Daemon --> Canvas[Canvas state]
  Daemon --> Broker[Detached broker]
  Viewer -. live session .-> Broker

The broker has one job: host a node’s agent session. The daemon has the broader job: preserve the node graph, state transitions, waits, reports, and broker lifecycle. A viewer may stream a broker’s live session, but it does not become the authority that decides whether that node is active, dormant, or revived. Closing a viewer closes a view, not the node’s durable identity.

This is why a process crash is recoverable rather than an automatic loss of work. The daemon retains the node row, session information, artifacts, waits, and inbox state, then applies recovery to the broker execution. A later wake or explicit revival can continue the node’s saved conversation. The application should create nodes through client.nodes, not start an LLM process itself, because the daemon is the component that keeps the engine and durable canvas state coordinated.

The costs are real. A daemon must be running, and canvas state has one home instead of being scattered across terminals and application processes. Local clients use its owner-only Unix socket. Remote clients use its configured TCP listener and bearer token. A network interruption can therefore fail a mutation loudly rather than encouraging a client to replay it and risk doing the action twice.

Those costs buy a simpler model: one state owner, one broker launcher, and clients that can reconnect. A CLI command, an SDK application, and a terminal viewer all use the same API for canvas state. They can come and go without creating competing writers or losing the graph that explains what each agent is doing.

For the lower-level runtime model, run crtr memory read internal/nodes-and-canvas. For a complete remote application setup, start with the SDK getting started guide.