Parallel agents on the board
Dispatch on as many cards as you want; each agent runs in its own git worktree on a kanbots/issue-N branch. The board updates live as runs progress, decisions surface, costs accrue.
v1.0 · MIT licensed·Open source on GitHub
A kanban that runs parallel agents on every card.
Drop a folder. Get a board. Dispatch Claude Code or Codex agents on as many cards as you want — each in its own worktree. Or hit autopilot and let personas split the work, run in parallel, and check their own output while you sleep.
Download desktopmacOS · Linux · WindowsSee cloud for teams
Free forever · Pay-what-you-can donations · MIT licensed
2
CLIs supported
Claude Code · Codex
0
bytes leave your machine
MIT
license · free forever
Two products on purpose
Solo on a laptop, or a team that ships together.
Same kanban metaphor. Same agent runtime. Different scale of collaboration. Both are legitimate paths.
Every Cloud-only feature requires either another person or another devicein the picture. None of them help a solo user on one machine — so we don't gate them. OSS is what one person does on their machine; Cloud is what a team does together.
Plays nice
Plays nice with the tools you already use.
Capabilities
Built for the way agents actually ship.
Not a chat tab. A board. Eight features that turn AI agents from a curiosity into a system of record.
Autopilot — self-evolving feature dev
Plug in personas — product, engineer, reviewer, tester — and a parallelism count up to 4. The orchestrator round-robins through personas, splits parent issues into subtasks, and evolves the backlog as agents discover work. Personas spawn personas.
Decision prompts you actually answer
Agents pause and ask. You click an option; the run continues. Numbered shortcuts, edit-and-resubmit, slash commands like /spec, /review, /split. Reviewable decisions, not silent tree mutations.
Bring your own CLI
Claude Code or Codex. Same board, same worktrees, same decision UI — kanbots speaks both stream formats behind a single AgentCliAdapter. Use your existing claude /login or OPENAI_API_KEY.
Local-first, zero servers
Everything lives in .kanbots/ next to your repo: SQLite database, configs, worktrees. No cloud account, no telemetry, no HTTP server. Your code never leaves the machine.
Cost analytics, live
Per-run, per-card, per-project rollups. Watch the cost meter accrue as agents work. Set per-run and per-session caps; runs stop on cost-budget. No surprise bills.
GitHub mode + draft PRs
Drive real GitHub issues with your personal PAT. Promote a worktree to a commit, or open a draft PR with one click. A pre-push hook means agents never publish on their own.
MCP server included
kanbots-mcp-server exposes the board over the Model Context Protocol so Cursor, Claude Desktop, or anything MCP-aware can drive it. The board becomes a first-class tool for your other agents.
Inside the app
Every surface, designed for the loop.
Not just a wrapper around a CLI. A full UI for dispatching, reviewing, splitting, and shipping agent work.
AUTOPILOT
Pick personas. Set parallelism. Walk away.
Up to 4 parallel slots round-robin through your roster. Each slot atomically claims the next persona; agents split parents into subtasks as they go. Stops on completion or session budget.
DECISIONS
Decisions, not silent mutations.
Live thread streams every tool_use and tool_result. When the agent needs a call, the run pauses with numbered options. Reply box accepts slash commands: /spec, /review, /split.
PERSONAS
Personas as first-class lenses.
A persona is a named system prompt snippet. Built-ins ship with the app; New persona lets you write your own — save, reuse forever. Custom personas stay on your machine.
PROVIDERS
Bring your own CLI.
Claude Code or Codex behind one AgentCliAdapter. Reuse your existing claude /login or codex login — no extra account, no extra key management. Switch per dispatch.
TASKS
Spec first. Or dispatch immediately.
Bug fix · Feature · Refactor · Review · Spike templates. Three start modes: spec-first (run /spec, await approval), create-and-dispatch, or queue-for-later. Title becomes branch + PR title.
CHAT
Talk to the workspace.
Ask anything — "which routes don't have rate limiting?", "where do we use the old auth helper?". A general-purpose agent that knows your repo, your tests, and your git state.
Autopilot
Personas spawn personas.
The backlog evolves itself.
Hand kanbots an issue and a budget. The orchestrator round-robins through a roster of personas, runs up to four slots in parallel, splits parent issues into subtasks, and keeps cycling until the work converges or the cost cap hits.
- 01
Pick a roster
Built-in personas ship with kanbots; or write your own — define the system prompt, save, reuse forever. Custom personas never leave your machine.
- 02
Set parallelism (1–4)
Each slot atomically claims the next persona via a round-robin counter. Four agents, four lenses, four worktrees — at the same time.
- 03
Personas split work
As agents discover work, they file new cards on the board. Later cycles pick those up. The backlog grows and shrinks under the orchestrator.
- 04
Stop on budget or completion
Per-session cost budget caps total spend. Stop button kills the parent and all children. In-flight runs finish their iteration cleanly.
QA mode
Run typecheck / tests / lint / build / e2e in the worktree. Optionally start your dev server and watch it. For every failing check, dispatch a fix run on a derived child issue. Repeat until green.
tsctestslintbuilde2e