Self-hosted · open source

Mission control for your AI agent team.

Agent Teams turns Claude Code or OpenAI Codex into a persistent, governed team of specialists. File the work on a board, step away, and trust it is handled — with receipts.

Works today with Claude Code and OpenAI Codex. Runs in Docker on your machine.

The Agent Teams Kanban board with task lanes, a burndown chart and a velocity chart.
The problem

One CLI is a brilliant brain with no memory.

The session drifts.

What started sharp ends up re-explaining itself, holding your whole plan hostage in one sprawling chat.

Nothing persists.

No project structure, no task state across sessions. Close the window and the plan is gone.

No safety rails.

Did it finish, or just say it finished? Without a contract, "done" is whatever the model claims.

What it is

A governance layer on top of your coding CLI — not a replacement for it.

Keep the CLI you already trust. Agent Teams wraps it with the operating layer that makes raw agent power actually land.

  • A Postgres-backed Kanban holds the plan so your agents never have to.
  • A Lead orchestrator resolves the project, reads the team playbook, and spawns the right specialist per task.
  • Five context zones keep standards, team, project and role state scoped with no bleed-through between tasks.
  • A defense-in-depth safety layer adds acceptance criteria, human-in-the-loop gates and hard budget caps.
What's genuinely special

Built like an engineering org, not a chat window.

01

One board. Every agent, every project.

Board, list, calendar and Gantt in one switcher. Queue tasks, watch them move across lanes, and see the whole operation at a glance. No scrolling through a chat to find where things stand.

The Agent Teams Kanban board with New, In Progress, Review, Blocked and Done lanes.
02

Done means proven done.

Every task is a contract. Each acceptance criterion is checked against evidence before the task can close — so "done" is the workflow's verdict, not the model's claim.

A full history trail records who verified what, and when.

A task drawer showing acceptance criteria, one of three checked.
03

Run your AI team like a real engineering org.

Burndown and velocity come built in. Track throughput across 90 days, spot stalls early, and plan the next sprint from real numbers instead of vibes.

A weekly velocity chart in green bars over the last ninety days.
04

Beyond code: many domains, many specialists.

Eight team playbooks and around 39 specialist roles ship today: development, content, SEO, data and more. Add a new team by dropping in a markdown file, no database migration required.

A grid of project cards tagged DEV, NOVEL and GENERAL.
05

Goes easy on your subscription.

Lean, scoped context per task and the right model tier per specialist stretch your plan further. Usage metering and hard budget caps keep spend visible and bounded.

A usage dashboard showing input and output token counts and session runs.
06

Self-hosted. A layer on your CLI.

Everything runs in Docker on your machine and wraps the CLI you already use. Choose Anthropic, OpenAI, Google Gemini, or Ollama for fully local inference. With Ollama, nothing leaves your network.

How it works

File it, step away, get it back verified.

  1. 1

    File tasks on the board

    Write each task as a contract with acceptance criteria. Queue as many as you like.

  2. 2

    The Lead resolves your project

    It binds the active project, loads the matching team playbook, and plans the work.

  3. 3

    A fresh specialist runs each task

    Scoped context in, results out. No sprawling conversation, no stale state carried between tasks.

  4. 4

    Done is verified against the contract

    Each criterion is checked before the task closes, and the audit trail proves it.

A closer look

See the whole operation.

Who it's for

Delegation with receipts, for people who ship.

Solo builders

Get the leverage of a whole team without the burnout of being one. Queue the work and let it run.

Small teams

A shared board, scoped tasks, and an audit trail everyone can read. The plan lives in one place.

Anyone running coding agents

If you drive Claude Code or OpenAI Codex and want structure around it, this is the layer you were going to build anyway.

Built in the open

Agent Teams builds Agent Teams.

I built this because I was tired of brilliant sessions falling apart halfway through. So I gave my coding agents a board, a memory, and a contract for "done."

The system you're reading about is the system that builds itself. Its own commit history and live Kanban are the proof: every feature here shipped through the same board it describes.

Follow the build on LinkedIn.

Get started

One command and a board.

Requires Docker Desktop running. Services keep running after you close the terminal.

Prefer cloning? git clone https://github.com/bankung/agent-teams and run the installer in bin/.

Then open http://localhost:5431. The command seeds a demo-tour project to explore.

  • Mode A - on your subscription (today): open the agent-teams folder in Claude Code or OpenAI Codex, tell the Lead which project to run, and it orchestrates the team interactively with per-action approval. No API key needed.
  • Mode B - headless via API (actively in development): tasks run with no terminal open, metered usage and hard budget caps.

Works with both Claude Code and OpenAI Codex, in the terminal or inside VS Code through their extensions.