You run the deck.
Agents run the work.
A desktop app where AI agent teams claim their own tasks, message each other, and review each other's code. You watch one kanban board and decide what ships.
Bump deps to latest minor
Add rate limit to /api/sync
Extract auth into a module
Write tests for parser
Fix null deref in session
Migrate config loader
- Providers
- 75+Providers
- Models
- 200+Models
- Runtimes
- 9Runtimes
- To start
- $0To start
Not one chatbot. A small crew with roles.
Each agent has a job — lead, builder, reviewer, tester. They claim cards, hand work off, and check each other while you watch one board.
75+ providers · 200+ models · mix them inside one team
One agent in a chat window makes you the bottleneck.
You copy context between sessions, remember who was doing what, and read scrollback to find out whether anything actually worked.
Work happens in a queue
One agent means one thing at a time, no matter how many terminals you open.
State lives in scrollback
Ten minutes in, nothing tells you what is done, stuck, or quietly burning tokens.
You are the message bus
Decisions get relayed by hand because sessions cannot talk to each other.
Review happens elsewhere
Diffs land in a separate Git client, detached from the reasoning behind them.
Everything the orchestration layer actually needs.
Eight parts that make multi-agent work supervisable instead of hopeful.
- 01
Agent teams
Build a team with real roles. Agents work in parallel, talk to each other, and hand off across teams.
- 02
Kanban board
Five columns that move on their own. Tasks change status the moment an agent picks them up.
- 03
Code review
A diff per task with accept, reject and comment — plus a built-in editor with Git support.
- 04
Token analytics and budgets
See exactly where tokens go by team, agent, task and model. Set caps and get warned at 80%.
- 05
Organizations and global overview
Nest teams into departments and squads, then watch every one of them on a single live map.
- 06
Terminal workspace
A real terminal in the app, with team shells, local shells, history and autocomplete.
- 07
Cross-team communication
Agents message each other directly, inside a team or across teams, and you can read all of it.
- 08
Solo mode
Start with one self-managing agent. Turn it into a full team the moment the work outgrows it.
Four steps between installing and reviewing your first diff.
- 01
Install and open
The app ships with a free model already wired up, so you can take the first run the moment it opens.
- 02
Point it at a project
Open any local repo. AgentTeams reads the structure and sets up a workspace per teammate.
- 03
Create three roles
A lead, a builder and a reviewer is the right starting team. Add more once review is the bottleneck.
- 04
Give it an objective
The lead breaks it into cards and the board starts moving. You review what reaches the review column.
Start from the job, not the feature list.
Each one comes with a suggested team, the workflow, and the guardrails to set first.
- For teams drowning in small stuff
Clear a backlog of small tasks overnight
Point a team at thirty low-risk tickets and wake up to thirty diffs waiting for review.
LeadBuilder ×3Reviewer - For anyone who inherited the code
Refactor a legacy codebase without breaking it
Split the refactor across agents, block risky work behind tests, and review every diff before it lands.
LeadTest writerBuilder ×2 - For the person who has to fix the number
Raise test coverage on a project nobody tested
One agent per module, tests written in parallel, every suite run in the built-in terminal before review.
LeadTest writer ×3Reviewer - For solo founders and indie builders
Ship a product as one person with a team of agents
Start in solo mode, promote to a team when the work outgrows one agent, and keep the spend capped.
Solo agentBuilderReviewer - For regulated and privacy-sensitive teams
Use agents without sending your repo to a vendor
Coordination stays on your machine, and you choose which model — including fully local ones — sees any code.
LeadBuilderReviewer - For people building an AI-first operation
Run several teams like departments
Nest teams into an org, watch every board on one map, and route work between departments.
Org leadTeam leadsSpecialists
Where AgentTeams wins, and where it doesn't.
Every competitor here is good at something. The table says which one, honestly.
| Feature | AgentTeams | Cursor | Claude Code CLI | Paperclip |
|---|---|---|---|---|
| Live team collaboration | ||||
| Kanban board | ||||
| Code review in app | ||||
| Visual team progress | ||||
| Agent to agent messaging | n/a | |||
| Budget controls | ||||
| Built-in terminal | ||||
| Built-in code editor | ||||
| Separate agent workspaces | ||||
| Linked and blocking tasks | ||||
| Mixed runtimes in one team | ||||
| Setup effort | ||||
| Runs locally by default | ||||
| Price to start |
Free to install. Free model included.
There is no paid tier. Provider usage is optional, and you set the ceiling.
$0
Open source. No seats, no trial, no card.
- Free model with no authentication
- Unlimited agents and teams
- Kanban board with real-time updates
- Code review with diff view and Git
- Token analytics, budgets and hard caps
- Optional paid providers when you want them
Paid provider usage is billed by that provider, never by us.
Questions people ask before installing.
What is AgentTeams?
A desktop app for running teams of AI coding agents. It brings its own local coordination layer, so agents can hold roles, claim tasks, message each other and review each other's diffs. You supervise all of it from one kanban board instead of several terminal windows.
Is it really free?
Yes. The app is free and open source, and it ships with a free model that needs no authentication — no account, no API key, no card. Paid providers are entirely optional and only apply if you connect one.
Which platforms are supported?
macOS on both Apple Silicon and Intel, Windows on x64 and ARM64, and Linux via AppImage. The download page detects your platform and puts the right build first.
How does it actually work?
Install the app, open a project, and create a team with a few roles. Give the team an objective. The lead breaks it into tasks and assigns them, builders work in parallel, and the reviewer checks the diffs. You watch the board and step in when a card reaches review.
Is my code private?
Coordination state, task history and app data stay on your machine. Code only leaves your machine if you point an agent at a hosted model — and you choose that per team, so you can keep sensitive repos on a local runtime.
Which models and providers can I use?
Claude Code, Codex and OpenCode run as first-class runtimes. Cursor, GitHub Copilot, SuperGrok, Z.AI, MiniMax, Kimi and Kiro connect through supported provider routes, and OpenRouter opens up 200+ models across 75+ providers. Local models work through Ollama.
Stop being the message bus.
Install it, open a repo, and give three agents an objective. Free, no account, no API key.