Agent Teams AI
A desktop control plane for autonomous, reviewable coding teams across multiple AI runtimes.
Per-dimension scores and reasoning
Evidence shows: README mentions configurable autonomy levels (fully autonomous or step-by-step approval), indicating user confirmation mechanisms; token analytics and budget controls suggest data flow transparency; dependency audit in CI (pnpm audit) provides dependency security; atomic file write tests show attention to data integrity. Deductions: no specific permission model details, least_privilege only partially satisfied; user confirmation not detailed; data flow transparency only implied by features; sensitive data handling not explicit; external effects (e.g., network requests) not specified; rollback mechanism not specified; source attribution only via author info.
Evidence shows: extensive tests (e.g., atomicFile.test.js, cascadeGuard.test.js, controller.test.js) indicate good self-consistency; dependency locking (pnpm-lock.yaml) and CI dependency audit ensure dependency availability; failure messages are reflected in tests (e.g., error throws). Deductions: self-consistency not fully verified as static review cannot run tests; dependency availability relies on external services, not fully guaranteed; failure messages only partially covered.
Evidence shows: README describes multiple use cases (coding, project management), supports multiple AI providers, indicating broad audience and scenarios; environment fit via installers for macOS, Windows, Linux; trigger precision via built-in triggers (e.g., notifications). Deductions: capability boundaries not explicit (e.g., which actions require approval); trigger precision not detailed; environment fit only via installation instructions, not deep.
Evidence shows: README has clear structure, installation notes, FAQ, comparison table; license is AGPL-3.0; version number (package.json 2.1.2) and CI badges. Deductions: naming stability not explicit (e.g., API changes); known limitations not listed; version changelog not provided; maintenance responsibility only via author info.
Evidence shows: output usability via screenshots and demo videos; marginal value via comparison table (vs Gas Town etc.); cost-benefit via free model and budget controls. Deductions: output usability not actually tested; marginal value based on subjective comparison; cost-benefit not quantified.
Evidence shows: README comparison table has sources and methodology, indicating claim traceability; test files support feature claims. Deductions: cross-source verification insufficient (only internal sources); facts and inferences not clearly separated.
- Static review cannot verify actual runtime behavior; all feature claims should be treated with caution.
- Dependency security relies only on CI audit; no vulnerability remediation strategy provided.
- User confirmation mechanism not detailed; actual approval flow may be incomplete.
- Data flow transparency only implied by features; no specific data flow diagram provided.
What does this agent do, and when should you use it?
Agent Teams AI is an Electron desktop orchestration layer for Claude Code, Codex, OpenCode, Cursor, SuperGrok, GitHub Copilot, Z.AI, MiniMax, and Kiro. A user defines a project, team roles, and a provisioning prompt; teammates can then work in parallel while task state, dependencies, messages, and logs appear on a kanban board. It includes task-level code-diff review with accept, reject, and comment actions, plus optional Git worktree isolation per teammate. The app reads local project, app, and runtime/session data, including supported Claude Code data under ~/.claude/. The desktop application is the primary product, while the web version remains under active development.
The app auto-discovers installed Claude Code, Codex, and OpenCode runtimes, and its UI can connect the other listed providers. After a team is launched, teammates can create and manage tasks, exchange direct messages, add comments, and review work while the board updates task state. Each task can retain attachments, runtime logs, tool calls, messages, and workflow history; detected code changes are presented as hunk-level diffs for review. Its TerminalWorkspacePanel.tsx terminal workspace can run team-runtime and local-shell commands, and the built-in mcp-server supports MCP integration. It also aggregates input, output, cache, and reasoning usage, supports monthly token or estimated-cost budgets with 80% and 100% alerts, and documents hard caps for scheduled runs in ScheduledTaskExecutor.ts.
- An engineering lead using Codex, Claude Code, and OpenCode needs to split a cross-module feature among specialized teammates and follow progress on one board.
- A developer who must inspect AI-authored edits before adopting them can accept, reject, or comment on individual diff hunks within a task.
- A team running parallel changes across several repositories can enable Git worktree isolation for selected teammates to reduce workspace conflicts.
- An engineer investigating an automated task can inspect its runtime logs, tool calls, messages, timeline, and detected changes in one task view.
- A solo developer evaluating the product without account setup can begin with the documented free model that requires no signup, API key, or card.
What are this agent's strengths and limitations?
- Combines multiple coding runtimes in one team and automatically discovers installed Claude Code, Codex, and OpenCode runtimes.
- Provides a task-centered interface for real-time kanban progress, isolated execution records, teammate messaging, and code-diff review rather than only individual chat sessions.
- Supports optional Git worktrees, an integrated visual terminal, task attachments, and cross-team communication for parallel engineering work.
- Tracks token and estimated-cost usage across teams, members, tasks, models, runtimes, and sessions, with budget alerts.
- The desktop app is the current primary product; the web version is still in active development, and 24/7 cloud teams remain on the roadmap.
- Use of advanced models and runtimes depends on external provider subscriptions, API keys, access, and terms.
- The standalone HTTP dashboard has no built-in authentication for public deployment; documentation limits it to local or trusted networks unless placed behind an authenticated reverse proxy.
- Virtual GPUs in some Linux RDP sessions can cause Electron rendering failures; the documented workaround is launching with AGENT_TEAMS_DISABLE_GPU=1.
How do you install or deploy this agent?
Download a release for macOS Apple Silicon or Intel, Windows, or Linux as AppImage, .deb, .rpm, or .pacman. To run from source, install Node.js 24.16.0 LTS and pnpm 10+, then run:
git clone https://github.com/777genius/agent-teams-ai.git
cd agent-teams-ai
pnpm install
pnpm devOn first launch, use the no-auth free model or allow the setup wizard to discover local Claude Code, Codex, and OpenCode runtimes. The other listed providers are connected in the UI and generally use an existing subscription or API key.
How do you use this agent?
Launch the desktop app, choose a project, create a team, define roles, and write a provisioning prompt. Select the free model or a connected runtime, then launch the team; teammates create tasks, collaborate, and update the board. Open a task to inspect attachments, messages, logs, and diffs, then accept, reject, or comment on code changes as needed. For development-time visual verification, run pnpm dev:mcp; it starts the Electron app with hot reload and exposes local CDP debugging on 127.0.0.1:9222.
How does this agent compare with similar options?
The repository’s comparison table positions Agent Teams alongside Gas Town, Paperclip, Cursor, and the Claude Code CLI, covering live collaboration, approvals, visual progress, kanban boards, code review, workspace isolation, and budget controls. These are the project’s own product-positioning claims; teams should validate them against their runtime accounts and security requirements.