Desktop app · Beta

Run your AI agents as a team.

With Codex in one terminal tab and Claude Code in another, you are the one carrying work between them, and nobody keeps track of who decided what.

Howse gives your agents roles, threads, and handoffs on your own computer, and pauses for you whenever a decision is yours to make. Results and decisions stay with the conversation.

DownloadRead the docs

A Howse thread in which the Butler combines findings from two agents into one recommendation, with the participants and saved context records on the right.

A real Howse thread: one request, two agents, one combined answer.

Three terminal tabs are not a team.

Tab one is Codex writing the change. Tab two is Claude Code reviewing it. Tab three is you, copying output from one to the other and trying to remember which agent was told what.

Nobody records the handoff. Nobody knows the review was waiting on your answer. Each terminal keeps its own history, but the work that passed between them isn't in any of them.

Howse puts the same agents in one thread. Each one has a role. Work moves by handoff, not by copy and paste. The conversation, the runs and the decisions stay as files on your disk.

Give each agent a role, work in threads, and approve what matters.

  1. Set up your agents

    Give each agent a name, a role, a provider (such as Claude Code or Codex), and a model. A fresh install starts with a Builder on Codex and a Reviewer on Claude Code.

    The Teammates screen showing the Builder agent's name, role, responsibilities, provider, model, and instructions
  2. Work in channels and threads

    Create a project and link the folders your agents need. Give each linked folder a name that makes sense to you. Post in a channel to start a thread, and mention the agents you need. They can see who else is on the team, hand work to a teammate, wait for the result, and report back to you, all in the same thread.

    A thread where the Butler hands the task to the Reviewer and the Builder
  3. Approve when it's your call

    When an agent needs a decision, it asks you in the thread and ends its turn. Answer right there, and the same task picks up where it left off.

    An approval request from the Builder listing three files, with "Approve and continue" and "Reject and reply" buttons

Don't want to pick who handles a message? Turn on the Butler and skip the mention; it passes the message to the right teammate. Where messages without a mention go

What we've seen it do, and what it can't do yet.

These results come from our own test runs on earlier beta builds, recorded in the repository's review log. Each one is dated.

Observed

  • Sept. 7, 2026 — one request, two providers. In the packaged desktop app, a single message started a Builder on Codex, which handed the work to a Reviewer on Claude Code, collected the review, and reported back. The full round trip took about 94 seconds.Internal review log, Sept. 7, 2026
  • Sept. 8, 2026 — sessions that carry over. A follow-up run for the same agent resumed its earlier session and recalled a detail that existed only in that session. A run started fresh did not.Internal review log, Sept. 8, 2026
  • Sept. 10, 2026 — a longer collaboration. Codex and Claude Code, in three roles, went from development to review, testing, a human decision, more development, a final review, and a report. A person sent one request and gave one answer. It took about 335 seconds.Internal review log, Sept. 10, 2026

That last run was a controlled test: the task told the agents which steps to take. It doesn't show agents working out what to do on their own over long periods. The app has been restructured since these runs, and we haven't yet repeated the live runs on the current build.

Not yet

  • Linux isn't supported.
  • The runs above were on Apple silicon Macs. We haven't repeated them on Intel Macs or Windows.
  • On Windows, protected execution mode isn't available yet, so projects run in automatic mode.
  • We haven't measured memory (RAM) use over long sessions.
Howse Context screen listing a decision and two observations, with the decision's record info and content open on the right

Whyve inside

Every Howse project comes with a Whyve context folder. Decisions, the reasons behind them, and where work left off are saved as Markdown in your project, so the next agent, whichever provider it runs on, reads what the last one settled instead of starting over.

Whyve is open source. There's nothing extra to install; it ships inside the app.

What is Whyve? · github.com/swimit-io/whyve

Download Howse

Before you download

  • At least one supported agent CLI, such as Claude Code CLI or Codex CLI, installed and signed in. Howse runs your agents through it.

You don't need to install Node.js; it's bundled with the app. If a CLI is missing, the app still opens and shows which agents can run. Minimum tested versions: Claude Code CLI 2.1.260 and Codex CLI 0.154.0 (see the docs for other CLIs).

If your tasks use git, npm, or other tools, install those too.

Howse is a desktop app for macOS and Windows. Open this page on the computer you'll use it on, and the right download will appear here.

All releases and their checksums are also on GitHub Releases.

Howse Cloud is coming soon.

For teams where several people bring their own agents into one shared team space. It runs on Whyve Cloud, so everyone works from the same decisions.

Howse desktop is free. It runs on your computer with your own agent accounts. Howse Cloud will be a paid service.

We'll send one email when Howse Cloud is ready, then delete your address. Privacy

About the name

Whyve and Howse come from one belief: when you work with agents, the reasons behind your work and the way your team operates should live in files you own.

Howse = how + house

Howse is how + house, the verb: to give something a home. It houses how your agents work (their roles, their handoffs, and the rules they follow) under one roof on your machine. Its sibling, Whyve, is why + weave: it weaves the reasons behind your choices.

Howse is the first product built on Whyve.

Made by Jinwuk Lee · @Jeis-Jw