Stop Juggling Agent TUIs in Twenty Terminals

What it does

Kandev is a web-based development environment that treats AI agents as parallel workers on a Kanban board. You assign tasks to agents from virtually every major provider—Claude, Codex, Copilot, Gemini, Cursor, Devin, and many more—then review their diffs, terminal output, and chat history in a single IDE-like interface before merging. It uses git worktrees to keep concurrent agents from colliding, and can run them locally, in Docker, over SSH, or in remote cloud sandboxes.

The interesting bit

The project is essentially opinionated glue: it speaks the Agent Client Protocol (ACP) to normalize a zoo of incompatible agent CLIs, then wraps them in a review-first workflow that insists a human understands the code before it ships. That boring orchestration layer—worktree isolation, session resumption, multi-repo branching—is where it tries to earn its keep.

Key highlights

• Supports 20+ agents via ACP adapters or raw CLI passthrough, including Claude Code, GitHub Copilot, Codex, Cursor, Devin, and Grok.
• Runs agents in parallel across isolated git worktrees, with support for multi-repository tasks and multiple PRs from a single job.
• Integrated workspace bundles a code editor with LSP, terminal, git changes panel, browser preview, and chat in one view.
• Runs locally or self-hosted over Tailscale or VPN, with no telemetry and no cloud lock-in.
• Workflows are portable YAML you can export and import across workspaces.

Caveats

• Several agent integrations require external CLI installs, paid subscriptions (e.g., Cursor Pro), or beta ACP adapters; not all are turnkey.
• Office mode, an autonomy layer for persistent agent teams with memory and budgets, is currently in progress and not yet a supported feature.

Verdict

Teams already experimenting with multiple AI agents and craving a single review surface will find this a useful upgrade from terminal juggling. If you are happy with one agent in one TTY, it is probably overkill.