Run a team of coding agents on your machine.
Agent Teams turns the Claude Code, Codex or goose you already have into a team: a planner, a builder and a reviewer that work in your repository. A finished task opens a pull request, and you merge it. No sign-in.
Not released yet. Requires Node 24 or later and Claude Code, Codex or goose installed.
$ lucitra
Open the shell in your repository.
$ lucitra plan
Break a goal into tasks. Nothing runs yet.
$ lucitra candidates
List work the repository already says needs doing.
$ lucitra start-build
Start a task in its own git worktree.
$ lucitra tasks
See what is running and what is waiting on you.
$ lucitra ask
Search the project memory.
A planner, a builder and a reviewer
The planner breaks your goal into tasks. The builder writes each one. The reviewer checks the result and recommends. You decide what ships.
Every task in its own git worktree
The builder works in a separate worktree, so the branch you are on stays as you left it. If you decline a push your policy held, the worktree is kept for you to inspect.
You merge what ships
A finished task pushes its branch and opens a pull request. Read the diff and the reviewer's findings, then merge. To hold every push for you first, a team policy gates it.
Project memory in your repository
What the team learns is written as markdown in your repository, where you can read it, diff it and correct it. lucitra ask searches it first.
See what is running and what is waiting on you.
lucitra tasks lists every task, the agent on it and the worktree it runs in.
Read the diff, then approve.
A push your policy holds stops here. The reviewer posts its findings; you approve or decline.
Every repository in one view.
lucitra dashboard shows each repository and what the week cost.
Ask the project memory.
lucitra ask searches what the team has written down, file by file.
Meet the team.
Four agents, each with one job and one thing it cannot do.
- LLucitraaccept-edits
Your point of contact. It takes requests and hands work to the right teammate.
Cannot: merge a pull request - PPlanning Leadplan
It breaks a goal into tasks small enough for one agent.
Cannot: write any code itself - BBuilderaccept-edits
It builds one task in its own worktree and opens a pull request.
Cannot: merge it - RReviewerplan
It reviews the pull request and posts findings with a recommendation.
Cannot: approve it
Commands.
lucitraOpen the shelllucitra planBreak a goal into tasks without running themlucitra candidatesList work the repository already says needs doinglucitra start-buildStart a task in its own git worktreelucitra tasksSee what has run and what is runninglucitra approveLet a waiting task pushlucitra dashboardEvery repository in one view, with what the week costlucitra askSearch the project memorylucitra mcpExpose these commands to any agent over MCPRuns on your machine.
Terminal
Your tasks and approvals live in the terminal you already use.
Daemon
It runs the team, the task store and the MCP server on your machine.
Open source. It runs on your machine.
On your machine
Open source- Every command on this page
- Claude Code, Codex or goose
- Project memory in your repository
- Token budgets with hard stops
For a team
- Roles, permissions and approvals
- Controls mapped to your compliance framework
- Work from Linear and GitHub Issues
Questions.
Is it released?
Not yet. It needs Node 24 or later, and Claude Code, Codex or goose installed.
Is my code private?
Your code goes where it already goes: to the model provider behind Claude Code, Codex or goose. Agent Teams starts each of them with its telemetry turned off, collects none of its own, and keeps usage counts on your machine. Anything else leaves only through what you set up: the push and pull request at the end of a task, through your own git and gh, and any integration or notification channel you add.
Which coding agents does it work with?
Claude Code, Codex and goose. lucitra runtimes shows which are installed and switches between them.
Where does work come from?
From you, or from your repository: lucitra candidates lists open issues, drafts and old TODOs it already names.
Can an agent push without me?
Yes, when a task finishes: it pushes its branch and opens a pull request, and a person merges it. To hold every push for your review first, list push under gated in a team policy.
Where does the project memory live?
As markdown in a repository you own, so you can read it, diff it and correct it.
Do I need an account?
No. It runs on your machine with no sign-in.
Is it open source?
Yes. Agent Teams is an open source tool that runs on your machine.
A team that plans, builds and reviews. You merge what ships.
Not released yet. Requires Node 24 or later and Claude Code, Codex or goose installed.