Agent Doorbell

Start here

How Doorbell works

Send matching events to your agents and send their results to other workflows.

When to use it

Connect one source, choose the data Doorbell retains, and send matching messages to named agents you own and run. Several agents can use the same source without duplicating its canonical messages.

Example

An order needs review. Doorbell sends its selected details to order-reviewer. The agent reviews the order and sends its result to a review-results webhook output.

Set it up

  1. Name your agent and choose how it receives events. With MCP, your agent checks for messages. With a webhook, Doorbell sends events to your workflow's HTTPS URL.
  2. Choose event sources, rules, and fields to share. Save a draft and preview a sample. Saving does not change anything live.
  3. Approve a connection test. A webhook must accept it. An MCP agent must receive and acknowledge it without doing real work.
  4. Review the saved settings and activate. Send a new matching event to check delivery.
  5. Agents can send messages to each other, reply to messages, or publish results to named webhook outputs that you approve.

Important limits

  • Doorbell delivers events. You run the agents, give them task instructions, and check their results. Delivery does not mean a task is finished.
  • Keep MCP agents running and checking for messages. Doorbell cannot wake a sleeping agent.
  • Your agents share your Doorbell permissions. Names, sender labels, and rules do not restrict access between agents.
  • Doorbell stores only permitted source fields. It does not store email bodies or attachment contents. Receiving or acknowledging a notification does not delete the stored message.
  • Agents choose each message's recipient. Reply and conversation IDs link messages but do not choose where a reply goes.
  • Doorbell does not run tasks, wait for several agents' results, or take email actions on your behalf.

Troubleshooting

  • If nothing arrives, check that your settings are active, the source is connected, and a sample matches. Then check the receiver's connection test and message polling or webhook history.
  • If work repeats, compare message IDs and notification IDs before retrying. Your receiver should use message IDs to avoid repeating work.
Technical reference

Configure with save_agent_address, test_agent_receiver, activate_agent_address, pause_agent_address, and delete_agent_address. Runtime tools are receive_agent_message, send_agent_message, reply_agent_message, and publish_agent_result. Use list_agent_addresses to find saved agents and outputs.

Retry keys apply across send, reply, and result calls for seven days for all your agents. An identical accepted retry returns its original receipt without sending again, even after settings change or content is erased. It does not restore erased content.

Named addresses own retained inboxes and notification settings. Named messages use the service-managed doorbell-addresses source. Retained reading, notification leases, and acknowledgments keep separate progress. Configuration inspection is read-only; change inputs and notifications through the address.

Webhook workflows can publish through Doorbell's authenticated MCP HTTP endpoint. Results go to saved outputs, not arbitrary URLs supplied at runtime.

Open Agent Doorbell