Your AI interns finally learned to talk to each other

What it does

Agent Teams is an Electron desktop app that turns LLM runtimes into autonomous dev teams. You create squads with assigned roles, drop tasks on a kanban board, and the agents message each other, split work, write code, and review diffs without your intervention. It wraps Claude Code, Codex, and OpenCode (200+ models, 75+ providers) in a management layer that handles the coordination glue.

The interesting bit

The project treats agent coordination as a first-class problem, not an afterthought. Agents have real mailboxes, cross-team messaging via @team-name, and a built-in review workflow where they inspect each other’s hunks. When a runtime compacts context, the app restores team-management instructions so the kanban doesn’t drift. There’s even auto-resume after Claude rate limits.

Key highlights

• Runs fully offline with free models; no signup or API key required to start
• Hunk-level code review with accept/reject, plus a built-in editor with Git support
• Per-task log isolation and token-usage breakdowns by context category
• MCP server integration for external tools and extensible plugins
• Solo mode: one agent that self-assigns tasks, expandable to full teams later

Caveats

• Windows requires Administrator launch, especially for OpenCode runtimes, and may trigger SmartScreen
• The “200+ models, 75+ providers” claim is stated but not enumerated in the README
• Comparison table positions the project against Cursor and Claude Code CLI, but some cells are truncated in the source

Verdict

Worth a look if you want structured multi-agent workflows rather than single-threaded chat sessions. Skip it if you’re happy with Cursor’s IDE integration or need fully proven enterprise orchestration.