VibeSwarm · Orchestration
One prompt in.
A whole team out.
VibeSwarm turns separate AI coding CLIs into one coordinated crew — a Coordinator that plans, a Scout that reads, a Builder that ships and a Reviewer that refuses to take anyone's word for it. Each role can run a different model.
Split goal into 3 parallel-safe tasks → delegated
Mapped auth flow · findings posted to board
Editing LoginController.php in worktree wt/login
Re-running tests before approving “done”…
Shared task board
- Add login page doing
- Map existing auth code done
- Wire session tests review
- Dark-mode toggle todo
.ovibe/team/blackboard.jsonl — live
The roster
Specialists, not one overworked generalist
A single agent doing everything drifts, forgets and marks its own homework. VibeSwarm splits the work the way a real team does — and each seat can be filled by a different CLI and model, so the strongest tool gets each job.
Plans & delegates
Breaks your goal into parallel-safe tasks and hands each one off decision-complete: goal, files in scope, constraints, definition of done. It's read-only by design — it plans, it never touches source.
Reads before anyone writes
Maps the relevant files and where a change should go, grounding every claim in code it actually opened — cited as path and line, never speculation. Findings land on the board for the whole team.
Ships the change
A senior-engineer persona that implements the task end-to-end and leaves the build green — reporting the files it touched and the verify command it ran.
Trusts nothing
Independently re-runs the build and tests, and rejects any “done” claim it can't reproduce. Flags scope creep — refactors and renames nobody asked for — with findings cited as path:line.
Need more hands? The pool also includes a Debugger that reproduces bugs before fixing the root cause, and a Tester that writes edge-case tests and never weakens one to make it pass.
Reliable hand-offs
The app is the message broker
Agents don't shout into each other's terminals and hope. To delegate, a teammate appends one JSON line to a shared outbox file — and The Vibe Father watches it and delivers the task into the right agent's terminal. Deterministic, replay-proof, and robust where scraping a rich terminal UI is not.
- ▸Addressed by role or name. Send to “builder”, “reviewer” — or a teammate you named yourself.
- ▸Duplicate-proof. Fingerprinting stops a repainted terminal line from dispatching the same task twice.
- ▸A live activity feed. Every routed hand-off shows up in Team Activity, so you always know who asked whom for what.
coordinator → {"to":"scout","task":"Map the auth flow — where do sessions get created?"}
coordinator → {"to":"builder","task":"Add /login page. Scope: routes, LoginController, blade view. DoD: tests green."}
builder → {"to":"reviewer","task":"Login page done — please verify. Touched 3 files."}
✓ delivered to Reviewer's terminal
Builder
Reviewer
Scout
Builder
The shared task board
Watch the work move, card by card
Every agent posts its own work to a shared blackboard the human actually watches: a card enters at todo, moves to doing, lands in review at hand-off and finishes at done. No digging through four terminals to figure out where things stand — the board is the status meeting.
The Coordinator also summarizes each worker's result on the board, so teammates read a digest instead of re-reading each other's raw logs.
Parallel without collisions
Git worktrees keep agents out of each other's diffs
Two agents editing the same checkout is how AI teams eat their own work. VibeSwarm gives parallel tasks separate git worktrees — real branches in real folders — and the Coordinator only dispatches tasks in parallel when their file scopes don't overlap.
The Changes panel shows every worktree live: its branch, dirty-file count, and how far it's ahead or behind your main checkout — so merging back is a decision you make with the facts in front of you.
Reviews with teeth
“Done” is a claim. The Reviewer treats it like one.
Most agent stacks let the same model write the code and declare it finished. In VibeSwarm the Reviewer independently re-runs your build and tests, rejects anything it can't reproduce, and answers with either APPROVED or a numbered list of concrete required fixes — specific and honest, not reassuring. Pair it with VibeSafe's AutoVibe gate and nothing reaches you without a genuine green.
Pairs with
Stronger together
The AutoVibe gate runs your real build and tests before work ships — plus one-click checkpoints and rewind.
Explore VibeSafe VibeBrainOne brain, any model — Anthropic, OpenAI, xAI, Kimi, MiniMax, OpenRouter or your own endpoint, with persistent memory.
Explore VibeBrain VibeBenchStop guessing which model is best. Live cross-model benchmarks, leaderboards and head-to-head compares.
Explore VibeBench
Hire the whole team at once
VibeSwarm ships in The Vibe Father — with the brain, the safety gate and everything else. One flat price.