Introducing ADHDev
The control plane for your coding agents.
Claude Code, Codex, Gemini — running on your Mac, your Windows box, your home server. ADHDev coordinates them all: dispatch tasks in parallel, approve from your phone, and let the Refinery land every branch safely on main.
📸 [screenshot: the ADHDev dashboard showing several agent sessions running across machines at once]
You already have the agents. ADHDev is the layer above them — one place to see every session, send work to whichever machine has capacity, and get finished branches merged without babysitting a single terminal.
The four pillars
1. The Refinery — landing the work
Spawning ten parallel agents is easy. Merging ten branches is the hard part. The Refinery validates each worktree against your repo's own gates (typecheck, tests, lint), verifies patch equivalence, publishes submodule commits, fast-forwards main, and cleans up the worktree — automatically.
Why it matters: this is the part vendor tools leave to you. They create the worktree and stop. ADHDev takes it all the way to main — and it never force-pushes. Anything ambiguous is kicked back for human review instead of being forced through.
2. Any agent, mixed freely
Vendor-neutral by design. Run Claude Code on the Mac, Codex on the Windows box, Gemini for a review pass, Hermes alongside them — in the same mesh, under the same policies.
Why it matters: every vendor ships only its own agent. ADHDev lets you mix them and switch between them without switching platforms.
3. Your machines
P2P-first. Your Mac, your Windows box, your Linux server become a single control plane. Chat, commands, and screenshots flow directly between your browser and your daemon over WebRTC; the cloud only handles auth and signaling metadata.
Why it matters: the work runs on hardware you already own and trust, coordinated as one — not scattered across terminals you have to alt-tab between.
4. Remote control + HITL
Approve tool calls from your phone. Drive a session over a P2P connection. Keep long-living sessions alive across daemon restarts through the session-host stack.
Why it matters: the human decision is often the bottleneck. ADHDev routes approvals and status to wherever you are, so a waiting agent doesn't stall for 20 minutes because you weren't at the desk.
Who it's for
Developers who run several coding agents across several machines and are tired of alt-tabbing between terminals to ask "is it done? is it stuck? did it need approval an hour ago?"
Two ways to run it:
- Standalone — AGPL, no account required, runs entirely on
localhost:3847. A fully local dashboard for the machine in front of you (and the rest of your LAN). - Cloud — multi-machine management, push notifications, webhooks, and a hosted API, at adhf.dev. Reach any of your machines from anywhere.
What makes it different
Parallel agents are easy. Landing the work is hard. ADHDev does the hard part.
Most tools help you start more agents. ADHDev is built around what happens after they finish — getting real, verified work back onto main safely, across every machine and every vendor you use.
To see it in practice: a recent 7-task protocol migration on this very repo was developed across a Mac and a Windows worker and every branch landed on main through the Refinery — validated, patch-checked, fast-forwarded, cleaned up.