kanbots

A local-first desktop kanban board where each card is an AI coding agent running in its own git worktree. Drop a folder on it and you get a board; drag a card to "in progress" and it dispatches an agent on a fresh branch (kanbots/issue-N) in an isolated worktree, and the board updates live as the agent works. The pitch is that a kanban board is already the right shape for supervising parallel agents — columns are states, cards are tasks, and the worktree-per-card model keeps the agents from stepping on each other.

How it works

One AgentCliAdapter sits behind eleven agent CLIs: Claude Code, Codex, Gemini, Cursor, Copilot, Amp, OpenCode, Droid, CCR, Qwen, plus anything that speaks ACP. It reuses whatever auth those CLIs already have, and you can switch which agent runs on a per-dispatch basis — so one card can run Claude Code and the next Codex, against the same board.

Autopilot is the multi-agent mode. You pick personas — Product Manager, Senior Engineer, UX Designer, Growth Lead, Reliability Engineer, or a custom one — set parallelism from 1 to 4, and the agents split parent cards into subtasks and file new cards themselves. It stops on completion or when a per-session cost budget is hit. The tagline is "personas spawn personas," which is also the risk: it's a tree of agents mutating a board.

The answer to that risk is decision prompts. Instead of silently editing the tree, an agent pauses and asks numbered questions; you answer in a reply box that also takes slash commands — /spec, /review, /split. The framing is "reviewable decisions, not silent tree mutations," which puts a human-in-the-loop gate at the branch points.

Everything lives in a .kanbots/ directory: SQLite for state, configs, and the worktrees. No cloud account, no telemetry, no HTTP server — "0 bytes leave your machine." There's live cost analytics with per-run and per-session caps, and a pre-push hook containment so agents can't publish anything on their own. GitHub Issues mode (via a PAT) and draft PRs are there when you do want to reach out. A bundled kanbots-mcp-server exposes the board over MCP, so other tools can read and drive it.

Isolation model

The worktree-per-card design is the same isolation instinct as sandboxing-ai-agents, done with git rather than containers — each agent gets a clean branch and directory instead of a microvm or matryoshka-isolation sandbox. It's lighter and repo-native, but the boundary is filesystem-and-branch, not a real security sandbox, so it protects agents from each other's edits, not the host from a hostile agent.

Usage

There's no CLI to paste — it's a desktop app. Install it, point it at a repository folder, and the folder's issues (or GitHub Issues, in that mode) become cards. Drag to dispatch, answer decision prompts as they come, watch cost accrue against the caps.

Business model

Open-core. KanBots OSS is MIT, free, has every feature, and runs on macOS, Linux, and Windows. KanBots Cloud ($19/seat/mo or $190/yr) gates only the multi-user and multi-device pieces: presence, sync, SSO/SCIM, audit log, Slack integration, and org-wide cost rollup. The single-developer experience is entirely in the free tier.

Limitations and watchlist

No public source repo — the site is a product landing page, so the OSS license (MIT) is stated but the code and implementation language aren't published for inspection yet. It's a new product with unproven maturity and an open-core model whose free/paid line could move. On watchlist for that reason — re-check in ~90 days to see whether the OSS build ships a real repo and whether the paid gating stays where it is.

claude-code-workflow-creator — the other new agent-orchestration entry; deterministic scripted fan-out versus kanbots' interactive board
agentsview — session cost/analytics ledger that could sit alongside kanbots' built-in cost caps
mcp-vs-skills — the MCP surface kanbots exposes its board through
sandboxing-ai-agents / microvm / matryoshka-isolation — heavier isolation models than kanbots' worktree-per-card
human-in-the-loop — the decision-prompt gate

Repo

kanbots.dev — desktop app (macOS/Linux/Windows), OSS build MIT-licensed. No source repository or star count published as of July 2026.