Vibe coding gave you a clear head. Then you opened three terminals, and the calm was gone. Vibecrew is the kanban board that brings it back: every feature is a card, every card gets an agent in its own git worktree, independent cards run in parallel, and one look tells you what’s done, what’s running, and what needs you.
I’ve been vibe coding for over a year. The best part was never the speed — it’s the clear head. You stop keeping every file in mind and start thinking about the product. Then I ran several agents across several projects, and the calm was gone: three terminals, constant checking, “wait, what was agent two doing?” I’d swapped remembering code for remembering process. Vibecrew is what I built to stop doing that.
Vibecrew carries the process for you — dispatching, answering, reviewing, merging — while leaving every part of it open. Any card’s agent is one keystroke from a live tmux session: read the transcript, type into it, take the wheel mid-task, hand it back. Assistance the whole way, with your hands never off the controls.
Where the task comes from doesn’t matter — your own head, a team backlog, someone else’s issue tracker. It becomes a card, and the card gets shipped.
A project can hold more than one repository — your library and the app built on top of it, on the same board. Every card spawns its agent in a fresh worktree of the repo it belongs to.
Assigned tickets come in as cards, keep their dependencies, and run as pipelines. Vibecrew becomes your personal planner on top of whatever tracker your team lives in.
Issues from every project you contribute to become cards. Each gets its own worktree and branch, so a good-first-issue and a deep refactor land as separate clean PRs.
Type or dictate one card, or drop in a whole plan and let a product agent break it into cards with the right order and dependencies. Then pick the pipeline each card runs — that's the one choice that matters.
A whole batch moves at once — every independent card gets its own agent, worktree, and branch. Each pipeline specs, plans, codes, reviews, and merges on its own. You read the reports.
Every feature is a card on a kanban board. Every card has a state you can see — speccing, planning, coding, in review, blocked, shipped. Cards relate to each other — dependencies and blockers are explicit — so independent cards run in parallel, each with its own agent, and nothing starts before what it needs. Across every project at once, because nobody works in a single repo any more.
The board dispatches ready cards, answers routine questions by policy, holds anything that needs judgment, and tells you what shipped. Queue cards in the evening; read the shipping reports with coffee. The number vibecrew exists to move: interruptions per shipped card.
Agents work in isolated worktrees — merging is your decision, or the pipeline’s when you dial it that way: auto-merge is opt-in per pipeline, behind the review gate. Diffs and comments in the app; real tmux terminals one keystroke away, running your subscription and your dotfiles. Everything local: one native app, a SQLite file, no account, no cloud.
One agent writing everything is how you get slop. Vibecrew gives each card a pipeline of specialists — spec, plan, code, review — with an independent harness reviewing the diff. That’s where the quality comes from, and why the same card ships well on a frontier model or an open-weight one. Pick the pipeline per card; if reality disagrees with the plan, the orchestrator says so instead of quietly carrying on.
Vibecrew runs tier-1 harnesses — Claude Code and Codex — and open-weight models — GLM, MiniMax, Kimi — under OpenCode, with Pi next. In practice the results come out much the same, because the quality comes from the pipeline: a proper spec, a plan, and an independent review agent on every card. Pick by what your subscription covers and which harness you like, not by how hard the card looks.
Every card, question, and merge goes through a single orchestrator agent — on your laptop, or through Telegram when you’re out. It dispatches the batch, answers what the agents ask, and merges what passes review. Take any of those decisions back whenever you want: total control, or top-notch assistance. Same agent either way.
Issues get handled in days, not quarters, and the roadmap is still soft enough to bend. Ask for the pipeline, the harness, or the board behaviour your work actually needs and there’s a real chance you’ll be running it next week. Download it, put a week of your own work through it, and tell us where it creaks.
Download for macOS — freeVibeCrew is free — now and after alpha. It’s not open source — it’s a native app you download and run. But there’s nothing to trust us with: no account, no cloud, no code leaving your machine. The app is closed; your data isn’t.
Yes — free, always. Not free-during-alpha, not free-tier: VibeCrew costs nothing, and local-first is permanent. You pay only your own model subscriptions or API keys.
Not today. We're open-source people — vibecrew runs open-weight models and open harnesses, and that isn't going away. But while the product is this young we want full control over how it develops, so the app itself stays closed for now. Either way there's nothing to trust us with: no account, no cloud, and your code, transcripts, and data never leave your machine.
A Mac (macOS 14+), git, and at least one coding agent CLI — Claude Code, Codex, or OpenCode — with your own subscription or API key. Vibecrew doesn't resell model access; your agents run under your accounts.
Tier-1 proprietary harnesses: Claude Code (Sonnet, Opus — headed or headless) and Codex. Open-weight models: GLM, MiniMax, Kimi under OpenCode, with the Pi harness landing during alpha. Mix them per card. Which one you pick is mostly about what your subscription covers — the pipeline, not the model, is what makes cards land.
Each card's agent works in its own git worktree on its own branch. By default nothing reaches your main branch without you — review the diff, choose merge or PR, per card. Pipelines can auto-merge past a passing review gate, but that's a dial you turn on, per pipeline.
As autonomous as you dial it. Run it fully hands-off: the orchestrator dispatches a whole batch in parallel, answers the agents' routine questions from your policies, and lets pipelines merge once the review gate passes. Turn any of that back to manual per pipeline — question answering, approvals, the merge itself. Anything outside policy escalates to the approvals inbox, a native notification, or your phone via Telegram.
It's the escalation path that works from anywhere without us running a relay: your own bot, your own chat, your machine on the other end. (More channels later.)
Not yet. macOS first — alpha feedback decides what's next.