Agent Teams AI

You will get the most from this if you already run several agents in parallel and lose track of which one is stuck. The kanban board is the trait that decides it day to day: every task carries its own log and diff, so answering what the agents did stops being archaeology. Teammates also message each other and file tasks, which reads as useful or as noise depending on how closely you supervise.

Pick it if orchestration is your real bottleneck. Pick Vibe Kanban for the same board shape with less desktop machinery, or Claude Squad if a terminal list of sessions was all you needed.

The structural risk is delegation without control. Claude Code, Codex, OpenCode, Cursor, Copilot and Kiro supply the actual behaviour, so a breaking change in any of them lands here first and gets fixed last. The row records per-runtime MCP curation as an unchecked roadmap item, meaning this cannot yet constrain what a teammate reaches for. Six vendors, one integration layer, and no version contract between them.

What it does right: each teammate can take its own git worktree under a configurable branch strategy, so a bad run gets deleted instead of untangled.

Three observations. 1. Verification is human and visual: changes are accepted or rejected at hunk granularity, which is a checkable property rather than an asserted one. 2. Inter-agent review is asserted, not specified; nothing states what a reviewing teammate reads or on what grounds it declines. 3. Nothing is claimed on any benchmark, which spares the reader a methodology audit.

Documentation is thin for a project first released in March 2026. The design is legible exactly where it is mechanical and undescribed where it is agentic, which is the usual ordering and rarely the reverse.

Ownership sits with 777genius, a single publisher, and the pricing row says free with no paid tier anywhere. The licence is the tell. AGPL is what you choose when you intend to sell exceptions later, so monetisation here is deferred rather than absent, and a buyer should expect a commercial edition before a rug pull.

There is no moat in orchestration itself; the runtimes it launches are the valuable assets and none of them belong to this vendor. The likely exit is a hosted team tier or quiet abandonment, not an acquisition. Position: use it, do not build around it, keep the worktrees.

The board demo is legible, which is more than most give me. Licensing for sixty is zero, so the spend arrives as tokens, and per-task token and cost analytics are the one feature my finance partner will like on sight. The identity story is missing entirely: nothing on the row records SSO, SCIM or an audit log. There is no headless mode, so this never runs in a pipeline, and distribution means a .dmg, .exe, .AppImage, .deb or .rpm per machine.

Onboarding is an afternoon. Not yet. Revisit when I can see what sixty desktops did without asking them one at a time.

AGPL-3.0 is the strongest copyleft anything in this class ships with, so a fork outlives the publisher and anyone who hosts it owes the source back. It ships its own MCP server package, which makes the app addressable from outside instead of only clickable. The credentials stay mine: Z.AI, MiniMax, Kimi or OpenRouter accounts plug in behind the runtime I choose.

The gap is hardware. Local models are not supported, so nothing here runs air-gapped, and the alternative to an installer is a pnpm clone and a dev build. Copyleft plus a clone command is still ownership.