The problem with most AI coding assistants is they queue up like customers at a deli counter—one task, then the next, then the next. KanBots, a new open-source desktop tool, takes a different approach: it lets developers dispatch multiple AI agents simultaneously across a visual project board, each working in its own isolated git worktree.
The launch drew immediate attention on Hacker News—206 points and 118 comments, with developers debating whether parallel agent execution is brilliant or a recipe for merge conflicts. At its core, the tool marries something familiar (Kanban task boards) with something still experimental (letting AI agents loose on your codebase in parallel).
What makes the technical architecture noteworthy is the isolation mechanism. When you dispatch an agent to work on a card, KanBots spins up a separate git worktree and branch for that specific run. Multiple agents can therefore modify code simultaneously without colliding. A pre-push hook keeps agents from sending anything to remote repositories. Nothing lands in production—or even in a commit—without explicit human approval.
Co-founders Alex Kim and Jordan Lee released everything under an MIT license: free to use, no telemetry, no external server requirements for the open-source version. Which raises the obvious question: how do they plan to make money?
The Local-First Bet
Drop a project folder into KanBots and you get a Kanban board. The interface shows live progress as agents work, cost tracking per card, and prompts whenever an agent needs a human decision. The tool supports Claude Code and Codex agents, running them through what amounts to isolated sandboxes—each agent operates in its own branch, labeled kanbots/issue-N, while others work elsewhere.
The approach feels almost deliberately restrained. In an industry where AI tools often promise to "automate everything," KanBots seems designed around human oversight. You can set spending caps per run or per session. You control when—or if—agent work gets promoted to a pull request.
There's an Autopilot mode for the adventurous. Give it an issue description and a budget, and it orchestrates work through two paths. Feature-dev mode runs up to four agents in parallel, cycling through different personas: product manager, engineer, reviewer, tester. QA mode runs a continuous loop of typechecking and testing, fixing failures until the build goes green. Whether that actually saves time or just burns through API credits faster remains an open question.
Coordination Without Centralization
The architecture choice is deliberate. Local mode stores everything in SQLite—no external services, no data leaving your machine. GitHub mode is optional: it mirrors column movements to status labels and can create draft pull requests, but authentication pulls from local credentials.
Perhaps more interesting is the MCP server bundled with the tool. MCP—Model Context Protocol—lets other AI coding tools like Claude Desktop or Cursor interact with the KanBots board programmatically. Localhost-only, with per-session bearer tokens. A developer could, in theory, orchestrate multiple agents from inside their existing editor while the Kanban board tracks everything in another window.
Kim and Lee say the open-source edition will remain free indefinitely, sustained through GitHub Sponsors and similar platforms. The macOS builds are currently unsigned—a friction point that signals either scrappiness or a team still figuring out distribution.
The Cloud Upsell
Alongside the free desktop tool, the founders offer KanBots Cloud. Starting at $19 per seat monthly, it adds real-time presence, cross-device sync, Slack integration, and organization-wide cost rollups. Enterprise plans layer on SAML authentication and immutable audit logs.
The cloud tier coordinates only metadata—cards, status updates, costs—while keeping agent execution local. It's an interesting compromise: you pay for collaboration infrastructure but keep the actual AI work on your own machine. Whether enough teams find that model compelling to sustain a business is an open question, particularly in a landscape where companies like Cursor and GitHub Copilot already have distribution and brand recognition.
A Crowded Moment
KanBots isn't alone in experimenting with visual interfaces for parallel agent work. Cline Kanban runs CLI agents in a local web app. Emdash supports multiple agent CLIs including Claude Code. Dispatch offers a macOS Kanban interface with, oddly, Telegram integration. Nous Research added Kanban-style multi-agent flows to their Hermes Agent tooling.
The common thread: a shift away from chat interfaces toward task boards that make agent work more legible. Rather than conversing with a single assistant, developers coordinate multiple agents across a project's full surface area.
It's a bet that the future of AI-assisted development looks less like pair programming and more like managing a small team of junior developers who work fast, make mistakes, and need supervision.
Maintenance Questions
The codebase itself is an Electron shell—React, Vite, SQLite via better-sqlite3, an agent runtime that spawns CLI processes and parses streamed JSON. Standard modern web tooling, organized into packages for core logic, storage, dispatching, LLM interfaces, and the desktop shell.
Whether the architecture holds up under real-world usage depends partly on contributor interest. The GitHub repository had accumulated over a hundred stars shortly after launch, and scattered Reddit posts suggest developers are intrigued. But open-source maintainability is less about initial enthusiasm and more about whether a small core team can keep up with issues, pull requests, and the inevitable feature requests without burning out.
Kim and Lee are betting that enough developers find the parallel-agent workflow compelling—and controllable enough—to build on top of. The tool's documentation emphasizes containment and cost tracking almost as much as capability, which suggests they've thought hard about what goes wrong when you let AI agents run wild.
The Hacker News thread, predictably, debated whether this solves a real problem or creates new ones. One commenter noted that visual task management doesn't eliminate the fundamental challenge: knowing when to trust an agent's work and when to intervene. Another pointed out that parallel execution only helps if your tasks are actually parallelizable—something that's true for testing and linting, less obvious for feature development where architectural decisions cascade.
Still, the project taps into something developers are clearly curious about. Not chat. Not full automation. Something in between—visible, controlled, parallel. Whether that curiosity translates into adoption, contributions, and eventually paying Cloud customers will likely become clear over the next several months.
For now, it's an interesting architectural choice in an industry still figuring out what AI-assisted development should actually look like.
