multi-agent coding, in your terminal
A swarm of coding agents you can talk to.
Plan it like a senior engineer. Decompose it into slices. A swarm of agents builds them in parallel — one git worktree each, dependency-ordered, merged back — while you steer any one of them, live.
One agent is a conversation.
A swarm is a team — and you're still in the room.
From a sentence to a merged feature
No dashboards. No hand-offs. A single flow you watch happen.
-
topic
You say what to build
One line. dogukan treats it like a brief, not a prompt.
-
plan
It plans like a senior
Problem, data model, API, sequence, slices, tests, risks — reviewable before a line is written.
-
swarm
A swarm builds it
The plan is sliced into tasks; one agent per slice spawns in its own git worktree — parallel, no clobbering.
-
talk
You steer any agent
Tab between agents. Type to redirect the focused one. Broadcast to all. Interrupt mid-turn.
-
merge
It merges back
Each slice integrates into the base in dependency order, so dependents fork from work that already landed.
Watch the swarm work
The real TUI: a sidebar of agents, live status, and a conversation pane you can type into.
Built for the terminal
- One worktree per agent
- Agents edit in isolation; no two ever touch the same checkout.
- Dependency-ordered integration
- Each finished slice merges before its dependents start — they fork from a base that already has it.
- Conflict-aware slicing
- Slices that touch the same files are chained, not run in parallel — so merges don't collide.
- Headless mode
dogukan run "…" --json— embed the whole pipeline in another agent or CI.- Honest reporting
- A conflict or a failed start is reported as failed — never a no-op disguised as success.
- Talk, don't wait
- Steer, interrupt, and broadcast to agents while they build.
Start a swarm
Requires Node 20+. Drives the full pipeline non-interactively — or run dogukan with no args for the interactive TUI.