kdlbs/kandev

What it solves

Kandev addresses the difficulty of running and managing multiple AI coding agents at scale. When you launch several agents on a large codebase, you quickly lose track of what each one is doing, and reviewing their output in a terminal doesn't scale. Kandev gives you a single control plane to define tasks, run agents from many providers in parallel, and review every change before it ships.

How it works

Kandev runs as a self-hosted server (a native Go binary with an embedded web frontend) that connects to agent CLIs through the Agent Client Protocol (ACP) or plain terminal passthrough. You create tasks on a kanban board, optionally define multi-step workflows that mix different agents per step, and assign each task to a runtime. The backend orchestrates agent sessions, isolates concurrent work using Git worktrees (one worktree per repository so agents don't conflict), and exposes an integrated web workspace with a file editor, terminal, git changes panel, and browser preview. Executors can be local processes, Docker containers, SSH servers, Kubernetes Pods, or cloud environments like sprites.dev. The server also supports automations (schedules/webhooks), sub-tasks, session resumption, and a plugin marketplace.

Who it’s for

Kandev is built for power users and teams who want deeper control over how AI agents work in their development process. It suits developers who need to review and iterate on agent-generated changes, teams that want consistent workflows and review gates across multiple providers, and organizations that want to self-host without telemetry or vendor lock-in. It can be run locally, on your own infrastructure, or accessed remotely from a phone via Tailscale or a VPN.

Highlights

• Run agents from 20+ providers side by side (Claude Code, Codex, Copilot, Gemini CLI, Devin, etc.)
• Parallel task execution with kanban and pipeline views, plus drag-and-drop automation
• Integrated workspace: terminal, code editor with LSP, git changes panel, and embedded VS Code in one IDE-like view
• Agentic workflows that mix-and-match agents per step, e.g. plan with Claude Code Opus, implement with Copilot Sonnet, review with Codex
• Workspace isolation via Git worktrees to prevent concurrent agent conflicts
• Flexible runtimes: local, Docker, SSH, Kubernetes, or cloud executors
• Open source (AGPL-3.0), no telemetry, no cloud dependency
• Experimental "Office mode" adds agent roles, permissions, approvals, budgets, and task delegation for persistent agent teams (disabled by default)