One-Click Handoff to Your Coding Agent
You don't have to copy a prompt, switch windows, open a terminal, and cd into the right repo just to hand a piece of work to your coding agent. DevSpec does all of that in one click.
Anywhere you can ask an AI agent for help — on an action item, on the testing page, on a failed deployment — there's a small rocket icon (🚀) sitting next to the "copy prompt" button. Click it, pick your agent, and DevSpec launches your local desktop app already pointed at the right project, with the work pre-loaded. The agent does the work on your machine, and the result pushes straight back into DevSpec. You go from reading an action item in your browser to a coding agent actively working on it, in a single click.
Want the bigger picture? Read From a Browser Tab to a Working Agent in One Click — why this small feature changes how a workday feels.
What the rocket actually does
When you click the rocket and choose an agent, DevSpec opens that app's desktop deep link and hands it three things automatically:
- The right repository. The app opens in your local clone of the project's repo, not your home directory — so the agent is in the correct working directory from the first keystroke.
- The prompt, pre-filled. Whatever you were about to copy — the "Implement", "Plan", "Investigate", "Fix this", or "Fix build" prompt for that item — is dropped straight into the agent's composer, already loaded with the action item's ID and instructions to pull its full context (description, history, linked commits) over DevSpec's MCP connection.
- The intent you picked. The selected action in the copy-prompt popover drives the launch too. Choose "Resume session" and the rocket resumes the exact coding session that last worked the item; choose "Fix build" on a broken deployment and the agent opens with the failing branch, commit, and deployment context baked in.
So the agent doesn't start cold. It starts knowing which item it's working on, where the code lives, and what you want done.
And the work flows back
This is the half that makes it a loop rather than a one-way door. The prompt DevSpec hands the agent tells it to finish the job the DevSpec way:
- It implements the change in your local repo.
- It generates a properly tagged commit message (with the action item's
[devspec:…]reference) and links the commit back to the item. - It records the implementation against the item over MCP.
The moment that lands, DevSpec sees it: the action item moves to implemented, the commit shows up in its activity timeline, and it appears on the testing page ready for verification. You never had to come back and tell DevSpec what happened — the agent reported it.
That's the whole trick: read it in the browser → one click → agent works locally → result is back in DevSpec, all without manually shuttling prompts, paths, or status updates between two windows.
Before it'll work
The rocket only launches agents you've turned on, and it needs the desktop app actually installed:
- Enable your agents. Go to Settings → Connections and switch on the coding agents you use (Claude Code, Cursor, and others). The rocket's menu lists exactly the agents you've enabled there — Claude Code even offers both its CLI and Desktop targets.
- Install the app. Launching works through the app's custom URL scheme (for example, Claude Code's), so the desktop app has to be installed on the machine you're clicking from. If nothing opens, the app isn't installed or its handler isn't registered — the "copy prompt" button right next to the rocket is always there as the manual fallback.
- Clone the repo locally. For the agent to open in the right place, you need a local clone of the project's repository on that machine.
- Use a desktop. Launching a desktop coding tool is meaningless on a phone, so the rocket is hidden on mobile and appears on larger screens.
Where to find it
Look for the rocket beside the terminal / "copy prompt" icon on:
- Action item cards — hand any open item straight to an agent to implement.
- The testing page — kick off a fix or a re-run on something a tester sent back.
- The deployments view — launch an agent into a failed build with the deployment context already loaded.
If you don't see it, check that you've enabled at least one agent under Settings → Connections.