DoReply
Learn the system

Automation, channels, and MCP

Start governed work from schedules, state changes, communications, or a card-scoped external AI client.

Not every useful task begins with a person opening a card. Delegate can start or route work through automation, communication channels, and MCP. Each route has an explicit principal and visible record, but they do not all inherit the same card, project, or budget context.

Scheduled tasks run by time

A scheduled task is appropriate when the trigger is a known time or cadence: prepare a Monday brief, check a queue each afternoon, or assemble a monthly report. It runs a recurring prompt in team or agent chat under the authority of the person who created it and produces a scheduled-task conversation and work stream. It affects a project or card only when the prompt explicitly asks the agent to use one.

Scheduling does not bypass team capacity, funding, or permission controls. A project budget applies only when the scheduled work actually operates in that project. When a boundary prevents a run, the pause remains visible.

Open team schedules for shared recurring work, or the selected agent's schedules for work owned by one agent.

Monitors run when a condition changes

A monitor watches a supported condition on a card, board column, or agent workspace. Use it when the meaningful trigger is state rather than the clock: a card enters review, an item remains in a column too long, or a workspace condition needs attention.

Avoid creating a monitor when an explicit dependency or ordinary human assignment already expresses the workflow. Automation should remove repeated coordination, not hide responsibility.

Channels bring conversations to the work

Team chat provides general coordination, while card activity is the right place to start and steer a specific outcome. Configured email and Slack channels can bring requests and replies into the team. SMS is available only when the deployment has it configured.

Channel identity and routing matter. A message should resolve to the right team, verified sender, and conversation or handling rule before an agent acts. When the conversation creates a trackable outcome, move its durable decisions and outputs onto a relevant card instead of leaving them only in a private thread.

MCP lets an external client work a card

From a card, Send to Codex / Claude creates a short-lived, single-use launch reference. You choose Act as to select the Delegate agent whose persona, tools, and ceiling the external worker will inherit, then choose Codex, Claude Code, or another compatible client.

After authorization and claim, the external worker is bound to that card and delegator. It appears in the card's work stream, can be paused or revoked, and can return a result like an in-product run. The external client pays for its own model execution.

MCP is the supported external-agent interface today. Delegate does not currently offer a general public API. A launch reference is not a reusable credential or an open route into the team.

One operating model

Whether time, a state change, a message, or an external client starts the work, ask the same questions: Where does this outcome and its record belong? Which agent and human authority apply? Should it use a card or remain a governed conversation? Where will the result and evidence appear? What limits should stop it?

Next steps