v0.6 · open source · by Lakeba

Your AI dev team.
On any model, any machine.

agent-team plans a roadmap, builds features as pull requests, reviews them, fixes what reviewers ask for, merges and releases, through GitHub, like a small team would. The rules that keep it safe are enforced in code, not left to the models.

Get started View on GitHub

$pipx install agent-team

ClaudeOllamaOpenAI-compatible Your other machinesmacOS & Linux

agent-team build "add CSV export" --branch team/export --pr

        Atlas · lead (claude-opus-5-5)
          Wrote PLAN.md: export command, CSV writer, 4 tests.
        Probe · qa  wrote tests/test_export.py first
        Forge · builder (qwen on your Mac)
          <edit path="todo.py"> 2 changes · export.py created
        ✓ tests: 31 passed (in a sandbox, no network)
        Atlas read export.py, ran the tests
          VERDICT: APPROVED after 1 round
        Sentinel · security  SIGN-OFF: YES
        ▸ secret check passed · opened PR #12
        Harbor · devops CI green, lead approved 4d74eaa
        ▸ merged PR #12 · $0.31 · 6 min
      

From idea to market

Conceive. Create. Commercialise.

Take a product from an idea to the people who use it, with a team that never sleeps and never skips a review.

Conceive

Agree what to build

The members discuss and agree a roadmap, weigh what developers ask for, and turn it into GitHub issues.

discussissues syncAtlas

Create

Build it properly

Plans, tests first, code, reviews with tools, security checks and fixes, a pull request at a time.

buildfixautopilotForge

Commercialise

Ship it and tell the world

Merges and releases with your approval, and a positioning brief, website and pitch deck checked against the code.

shipmarketHarborSpark

The team

Members with jobs, not one chatbot.

Each member has a role, a model and the files it owns. Name them, add more, put each on the model that suits the job: Claude where judgement matters, a local model where volume does.

Atlas

Lead

Plans the work, reviews every change against the plan and the project's guarantees.

claude-opus-5-5

Forge

Builder

Writes the code with small search/replace edits, round after round, until the lead approves.

qwen on your Mac

Harbor

DevOps

Runs shifts on GitHub: triage, sign-offs, fixes, merges, releases, a feedback digest.

claude-sonnet-5-5

Spark

Marketing

Writes the brief, the website and the pitch deck, checked against the code and reviewed with screenshots.

claude-sonnet-5-5

Sentinel

Security · optional

Reviews every approved build and signs off every pull request.

Probe

QA · optional

Writes the tests for a plan before anyone builds it.

Quill

Docs · optional

Brings the README, docs and changelog up to date after every build.

Yours

Any role

Any role with its own prompt and skills, on any model, on any machine.

How it works

From roadmap to release, a pull request at a time.

Run each step yourself, or let autopilot work through a milestone: it builds the next issue, and each devops shift reviews, fixes and merges, until the lead says the milestone is done.

PlanThe team agrees a roadmap; items become issues

BuildThe lead plans, builders write the code

TestEvery round, in a sandbox if you like

ReviewThe lead reads files and runs tests itself

FixRequested changes go back to the builder

MergeOnly with green CI and every approval

ReleaseAutomatic, or when you approve it

The team learns: reviewers' lessons go into every later prompt. ·A scoreboard shows what each member costs and how often it's approved first time.

Guardrails

Rules in code, not in prompts.

Models are asked nicely; the tools they act through check. A shift can't merge what the rules don't allow, whatever the model says.

✓

Approvals tied to a commit

Green CI, the lead's approval and every owner's sign-off, all on the exact head commit. A new commit needs them again.

◆

You decide what matters

Paths like auth/ or dependency files wait for your approval, and releases can too.

▣

Sandboxed tests

The code the team writes runs in a throwaway container, with no network and none of your keys.

⚿

No secrets pushed

Every push is scanned for keys, tokens and private keys first. Keys stay in your environment.

Spend limits

Per run, per day and per milestone, checked before every model call, with an estimate before you start.

⌧

Outside text is data

Issues, PR text and comments reach the models marked untrusted, never as instructions.

Any model, any machine

Mix Claude, local models and your own hardware.

A team runs entirely on Ollama with no API key, entirely on Claude, or anything in between. A weaker model hands its work to a stronger one after a rejected round.

anthropic

Claude

Opus to lead and review, Sonnet to build and ship, Haiku for routine checks.

ollama

Local models

Ollama on this machine or another: Qwen, Llama and friends, at no per-token cost.

openai

Any compatible server

vLLM, LM Studio, llama.cpp or a hosted gateway.

agent-http

Your other machines

A GPU box runs agent-team serve; the team finds it on the network.

The dashboard

See everything. Change anything. Approve with a click.

agent-team ui runs on your machine: the team and its health, every run with each member's verdicts, costs and the scoreboard, live activity, easy settings and templates, and every project you have.

127.0.0.1:8790/#overview

The dashboard's overview of the todo-py team

Overview: the team's health, recent runs, spend and GitHub at a glance.

Claude Code

Drive the team from Claude Code.

The plugin adds slash commands and a side panel with the team, its health, recent runs, spend and anything waiting for your approval. Claude runs the workflow where you can watch, checks the result, and reports.

/team-initset up a team

/team-planagree a roadmap

/team-buildbuild a task as a PR

/team-shipone devops shift

/team-fixapply a review to a PR

/team-autopilotwork through a milestone

/team-issuesthe to-do list on GitHub

/team-scheduleshifts on a timer

/team-panelthe side panel

Get started

Three commands to a working team.

macOS or Linux with Python 3.11+ and git. GitHub's CLI for the pull-request workflows. An Anthropic key only if members use Claude.

Install

The CLI, and the Claude Code plugin if you use it.

pipx install agent-team

or the one-line installer from the README

Set up a project

A checklist of what's missing, then a team file you can edit.

cd my-project agent-team ready agent-team init # asks: Claude, Ollama or both

Put it to work

Plan, build, or open the dashboard.

agent-team discuss "next features" agent-team build "add export" \ --branch team/export --pr agent-team ui

Shared by Lakeba

CONCEIVE · CREATE · COMMERCIALISE

agent-team is an open-source project that Lakeba shares with its partners, its clients and the whole community of software developers. It's free under the MIT licence: use it, change it, and tell us what you'd like it to do next.

Visit lakeba.com Suggest a feature

Give your repository a team.

Open source, runs on your machine, works with the models you already have.

Get started Read the guides Star on GitHub

Your AI dev team.On any model, any machine.