A Kanban board where every card is a live AI coding agent.
Each task gets its own git worktree, its own terminal and its own agent — so a dozen of them
can run at the same time without ever touching each other's files.
Website · Quick start · Download · Docs · Issues
Every card is a real git worktree, a real terminal and a real branch.
Write a task → get a worktree → an agent works in it → you watch → you review → it merges
Six steps. You write the first one and approve the last; the middle four happen on the board.
You describe the work and choose the agent. Claude Code, Codex, Gemini, Cursor Agent, opencode — or several at once, each in its own pane of the same task.
A fresh git worktree off your base branch, a tmux session inside it, your per-project
setup script, and — if you asked for them — free ports reserved for that task's dev server.
Heavy directories like node_modules or .venv are copy-on-write cloned, so the sandbox costs
near-zero disk and appears instantly.
Nothing the agent does can collide with another task. That is the whole reason ten of them can run at once.
Hover a card and the live terminal comes to you. No clicking in, no losing your place on the board.
The card also carries the state you actually need at a glance: which agent and model, priority, memory use, a bell when the agent wants you, and a red badge when it asked a question.
Split panes in the same worktree — one writes, one reviews, one runs the tests. Or three independent variants of the same task on separate branches when you want to see which approach wins.
The built-in review panel shows the whole branch diff with syntax highlighting, per-file read state, and inline comments on any line range. When you are done, one click copies the review back into the agent's terminal as a prompt.
Not sure the diff is safe? Launch a pack of read-only bug hunters — several agents that comb the same branch diff in parallel, each seeded to look somewhere different, and report only what they can prove.
The agent opens the PR and enables auto-merge. The card then shows the PR number and live CI state, so a red build finds you instead of you going looking for it.
Every project, every running agent, one Activity view.
One click away: what the fleet actually got done. Tasks shipped, lines changed, velocity, completion rate, active-day streak and average task lifetime — each against last week.
Add your own columns — AI Review, PR Review, On Hold, whatever your pipeline actually is — and attach an agent to a column so that dropping a card there starts the work. Drag a task into AI Review and a reviewer agent spins up on its own.
dev3 remoteServes the full UI to any browser over a Cloudflare tunnel, your LAN, or an SSH forward. Scan the QR once and your phone stays connected for 8 hours. Same board, same terminals, same live previews — on a dev box you never sit in front of.
⌘K jumps between projects, ⇧⌘P runs any action, ⌥Tab cycles live tasks with a preview of each.
dev3 is not a side door — it is how agents talk back to the board. They set the task overview,
leave notes for whoever picks the task up next, ping you when blocked, show you an image or an
interactive HTML report, start a peer task, or peek at another task's terminal without
disturbing it.
dev3 attention "Need a decision on the schema" # red badge on the card
dev3 show-image before.png --caption "the bug" # in-app image viewer
dev3 peek --task seq:1383 # read-only look at a peer's terminal
dev3 message --task seq:1383 --in 30m "status?" # type into a live agent, now or later
dev3 doctor --worktrees --prune-orphans # reclaim disk from dead worktreesdev3 show-artifact report.html puts a self-contained interactive page into the task — charts,
filters, sortable tables, its own light/dark switch. It opens beside the terminal that produced
it, stays attached to the task, and downloads as a ZIP. No server, no upload, nothing external.
Dark and light themes · a hand-tuned 16-color ANSI palette so agents look right in both · English / Russian / Spanish · one-click dev server per task with auto-assigned ports · labels and full-text search · sparse checkout for monorepos · scheduled and recurring agent runs · drag-and-drop files and pasted screenshots straight into a prompt · multi-window, one per monitor.
AI writes the code now. It commits, opens PRs, reviews. Your job changed — from writing to commanding a fleet of agents across more tasks and projects than any one head can hold. The bottleneck moved: it's not your editor anymore, it's your focus. Everything above is built around that. Two things we optimize for, above all:
1. Your speed — as one person. dev-3.0 optimizes a single developer: you. The unit is always the individual, never the org. It works fine on a team — but it's not a tool for managing other people; it's a tool for each person to command their own fleet and hit their own top speed. Everyone focuses on themselves, and the whole moves faster.
2. Beautiful, and built around you. A cockpit you stare at all day should be fast, gorgeous, and keyboard-first — and it should bend to your way of working, not force one on you. Great tooling doesn't just make you productive; it makes the work fun again. We sweat the polish.
And what we refuse: dev-3.0 is not an IDE — and won't become one.
- The code is the agent's job. No embedded editor; one click to your real VS Code or Cursor when you truly need it — and the goal is to need it less.
- Git is the agent's job too. No manual staging, no hand-written commits.
- Integrate through your agent. Claude Code, Codex & co. already speak MCP to Linear, Jira, and the rest. dev-3.0 is the cockpit; your agent is the adapter.
Other tools in this space are great — if you want to live in an editor. dev-3.0 makes a different bet. Pick by your goal, not a feature checklist:
| If you want to… | Reach for… |
|---|---|
| Stay in an editor, hands on the code and git | an agent IDE |
| Buy a platform for a whole team (SSO, seats, audit) | a team orchestrator |
| Run a fleet of agents solo, at speed, without drowning | dev-3.0 |
| Situation | What dev-3.0 does about it |
|---|---|
| Five agents editing the same repo | Five worktrees. They physically cannot conflict |
| You have no idea which of three approaches is best | Run all three as variants of one task, compare the diffs, keep one |
| An agent stops and waits for you — and you don't notice | The card turns red and the bell rings; Has Questions is a column |
| Reviewing agent output in a terminal is miserable | Full diff viewer with inline comments, and the review goes back as a prompt |
| You left the laptop and the build is still running | dev3 remote, scan the QR, keep going from your phone |
| A long task finished hours ago and you forgot it existed | The card carries the agent-written overview and notes, not just a title |
| Three hundred old worktrees ate your SSD | dev3 doctor --worktrees shows exactly what is reclaimable, and only deletes on your say-so |
🤖 The fastest way — paste this into Claude Code, Codex, Gemini CLI, whatever you already run:
Install dev-3.0 by following the guide at https://dev3.h0x91b.com/ai-install.txt
The agent reads the guide, detects your OS, and does the whole install itself.
macOS, by hand:
brew tap h0x91b/dev3
brew trust h0x91b/dev3 # newer Homebrew refuses untrusted third-party taps (skip on older brew)
brew install --cask dev3Linux (headless box, full UI in your browser):
brew tap h0x91b/dev3 && brew trust h0x91b/dev3 && brew install h0x91b/dev3/dev3
dev3 remoteWindows — download
canary-win-x64-dev-3.0-canary.zip
(~121 MB), extract it, run bin\launcher.exe. Windows support is brand new and may still be
rough: the build ships in the canary channel only — the stable download page has no Windows
file yet — and it is unsigned, so the first launch needs two clicks past a SmartScreen warning.
Details: Install guide.
Then: add a project (point it at a git repo), press ⌘N, describe a task, pick an agent, hit Run. That is the whole onboarding.
Prefer a direct
.dmg, a CLI tarball without Homebrew, a systemd service, or a build from source? → Install guide
| Runs today | |
|---|---|
| Agents | Claude Code · Codex · Gemini CLI · Cursor Agent · opencode · any CLI tool you configure |
| Desktop | macOS — Apple Silicon and Intel. Windows x64 — brand new and may still be rough; published in the canary channel only, unsigned |
| Headless | Linux x64 and arm64 (dev3 CLI + browser UI) |
Claude Code and Codex additionally report their status back automatically through hooks; the others do it through the installed dev3 skill. Session resume, system-prompt injection, rate-limit tracking and skill directories differ per agent — the full grid is in agent-support-matrix.md.
| Install guide | Every install path, tmux versions on Linux, cloud-VM caveats, build from source |
| Remote access | dev3 remote in depth — tunnels, systemd, sessions, exposed ports |
| Troubleshooting | dev3 doctor, disk reclamation, Full Disk Access, terminal colors and agent themes |
| Keyboard shortcuts | The complete list, mirroring the in-app ⌘/ panel |
| Agent support matrix | What each agent supports, feature by feature |
| CLI exit codes | The dev3 exit-code contract, for scripting |
| Diagnostic logs | What is logged locally, for how long, and what is redacted |
| AGENTS.md | Architecture and contributor guide |
bun install
bun run dev # Build + launch the app locally (no HMR)
bun run build # Staging build
bun run build:prod # Production build
bun run lint # TypeScript type-check
bun run test # Tests (fast subset; `bun run test:full` for CI parity)Built on Electrobun (native webview, no Chromium) and Bun, with React 19 + Tailwind in front, ghostty-web for GPU-accelerated terminals, and tmux underneath.
Contributions welcome — read AGENTS.md first; it is the architecture doc and the house rules in one file.
Apache 2.0 — © 2026 Arseny Pavlenko