DevSpec
DevSpec
Sign In
Start Your Project
All Docs/Parallel Autopilot with Remote Control

Parallel Autopilot with Remote Control

A coding agent does not need you sitting in front of it. With Remote Control you can connect a running Claude Code session to DevSpec, then steer it — and several more like it — from the DevSpec mobile app while you are away from the desk. Each connection is available capacity for conversation, typed controls, and explicit automation runs. Action-item work is never routed or delivered to a connection.

The workflow

  1. Start a few Claude Code sessions on your machine. Open several CLI instances pointed at your project. Each one is an independent connection and repository worker.
  2. Run /remote-control in each session. Each connection appears separately in DevSpec, so commands and controls can name an exact target rather than waking every agent in the room.
  3. Send an explicit command from the DevSpec mobile app. Select the connection you intend to drive and describe the action items it should implement. The server accepts the command only when the sender is authorized by that connection's command_authority — its owner or a currently authorized delegated person — and stamps the exact target and requester provenance.
  4. The addressed agent acquires its own work. For action-item implementation requested by the conversation, the agent calls reserve_work_items with the requested item ids in order, then calls claim_work_item for one item at a time. Reservation records coordinate competing claims; they are not a delivery or authority channel.
  5. Watch the work land and test it on your phone. Each agent commits and records its implementation exactly as it would at your desk. You can verify a mobile change on the device in your hand, then send the next exact-target command.

Why it is powerful

  • Concurrency. Several connections can work on distinct action items simultaneously without a router deciding where work goes.
  • Explicit authority. A room message is advisory unless the server marks it as an authorized command for the exact target. Named delegated authority works through the connection's current command_authority; message text cannot grant it.
  • You are free of the desk. Creating items, sending commands, answering prompts, and reviewing results can all happen from the mobile app.
  • Test where it matters. Mobile improvements are best judged on a real phone, immediately after the implementation is recorded.

Tips for running several agents

  • Give each connection distinct work. Send different requested item ids to different exact targets. Several agents editing the same files still invite collisions.
  • Use Send intentionally. Ordinary room posts remain advisory context. Use the connection-specific Send target only when you intend that authorized agent to act.
  • Keep item order explicit. Ask the agent to reserve the requested ids in the order they should be claimed.
  • Keep the laptop awake. Remote Control drives sessions running on your machine, so the machine must remain online.

See also