How it works

From an empty folder to a signed-off result.

Castiri is a Mac app that runs a team of Claude Code agents in one window. This is the whole process, screen by screen: setting up a project, briefing the team, and running it until every agent has signed off.

Phase 1

Set up the project and the team

Five pages in the New Team assistant, in the style of Xcode's new-project assistant, with Previous and Next along the bottom. Each one takes a minute.

  1. Install Castiri and check Claude Code

    Castiri needs macOS 15 or later and Claude Code, installed and signed in. Download the app and open it. The Welcome window offers three things: Create New Team, Open Existing Project and Browse Role Library, with your recent projects on the right.

    Before any team launches, Castiri checks Claude Code and tells you exactly what to run if it isn't installed, is out of date, or is signed out.

  2. Choose a template

    Eight templates in three groups, Research, Build and Review, plus Blank. Each one fills in a mission, picks a team and gives the agents their first instruction. Pick the closest; everything it chose can be changed on the pages that follow.

    See the 8 templates and all 35 roles

  3. Choose where the team works and what it may do

    Project: a New Folder, named, with a Git repository if you want one; an Existing Folder; or a Gaitro Project, which gets a draft of its own. Then a permission mode, one for every agent, from Ask before acting to Skip all prompts. Each mode says in plain words what agents may do on this Mac.

    Name the team, or leave it blank and Castiri names it after its roles.

  4. Describe the mission

    Three boxes: Goals, Definition of done and Off-limits. Then the Lead's mandate: what the team Lead may decide on your behalf, and what it must bring to you. The template fills in a starting point. The mission is written into CLAUDE.md, so every agent loads it at start.

  5. Assemble the team

    The Role Library, grouped by category like System Settings: pick any of the 35 roles, or one you wrote. The chosen team sits along the bottom with each agent's model. One agent is the Lead, your delegate, on the strongest model. Set the model and effort for any role on this team, or apply one setting to everyone.

    Each agent is its own Claude Code session, so Castiri shows a cost note once the team is bigger than six.

  6. Give the instructions and inputs

    Instructions are the first message every agent gets: validate an idea, build a feature list, review a website. Inputs are optional and come in four kinds: a Product Brief, a Feature List, an Idea to Validate or Notes you write in, files and images you attach, and website links, which agents fetch and which show up in Preview. Drop them in.

    The footer checks Claude Code one more time. Then Launch Team.

Phase 2

Launch

  1. Agree once, then launch

    The first launch shows one screen, Before agents start on this Mac. It says what agents can do in the mode you chose; that they run as you, with the same access your Terminal has; where your inputs go, which is to Anthropic through Claude Code under your Claude account, because Castiri has no server; and whose terms apply. Agree and Launch. You see it again only if the text changes.

    Castiri then writes the team into the project and starts one Claude Code session per agent, each with a briefing: who is on the team, the mission, the inputs, the files to work in, and the sign-off rules. If Claude Code asks whether to trust the folder, you answer once for every agent.

Phase 3

Run the team to sign-off

From here the agents do the work. You watch, answer when a decision is yours, and decide what happens to the result.

  1. Watch the team work

    One window. The sidebar lists every open project and its agents, each with a status ring: blue working, green idle, orange needs you. The middle shows All Agents, where working agents keep a live terminal and idle ones shrink to tiles, or one agent's terminal, or Preview, where anything the team runs on localhost or writes as HTML appears, or Documents: the outcomes, the instructions and inputs, and the mission.

    The inspector on the right holds Team Chat, Mentions and Sign-off. The agents read the inputs, post their first take in the team chat, agree on a plan, split the work and review each other's. A meter on each agent shows the tokens it has used.

  2. Answer when it's your call

    Agents talk in the team chat and message each other directly. When a decision is yours, the agent asks you, and the question lands in Mentions with a badge: one question, a few options, the agent's recommendation and one reason. The Lead answers teammates' questions that fall inside its mandate and brings you the rest, so you get one question instead of six.

    Reply in the thread. The whole team sees your answer unless you send it to one agent only, so everyone works from the same decision.

  3. Sign off, then hand off or ship

    The team is done when every agent's row in the Sign-off table reads Agree on the same lead idea. Castiri reads the table from the outcomes file and shows who has agreed and, for everyone else, exactly why not: a Block with its objection, a vote still pending, a row that's missing. Once everyone has gone idle it checks in with the holdouts, and you can nudge them yourself.

    When the sidebar reads Signed Off, it's your decision: stop the team, hand off to a new team that works from these outcomes, or in a Gaitro project write a one-paragraph summary and Ship. Gaitro runs the project's checks and publishes once they pass. The outcomes file and the Claude Code transcripts stay in the project.

FAQ

About setting up and running a team.

What do I need before I start?
A Mac on macOS 15 or later, Claude Code installed and signed in, and a folder for the project. Castiri checks Claude Code before launch and tells you what to run if it is missing, out of date or signed out.
Can I use a project I already have?
Yes. Choose Open Existing Project on the Welcome window, or Existing Folder in the assistant. If the folder already has a brief at docs/PRD.md, Castiri adds it as the Product Brief input.
What does Castiri write into my project?
A marked block in CLAUDE.md with the mission, a docs folder with your brief, the instructions and inputs, and the outcomes file, and one agent file per role in .claude/agents. In a Gaitro project, Castiri's files go in a .castiri folder that git ignores and CLAUDE.md is left alone.
When is the team done?
When every agent's row in the Sign-off table reads Agree on the same lead idea. Castiri reads the table from the outcomes file, shows who has agreed and why anyone has not, and checks in with the holdouts once everyone has gone idle.

Ten steps. One window.

Open a project, brief the team, and watch the work happen until every agent has signed off.

Download for Mac, free

Free while in beta·macOS 15 or later·Uses your own Claude account through Claude Code