A workbench for coding agents

Pi, Claude Code, Codex, OpenCode and Gemini in one window. Threads on the left, the conversation in the middle, terminal, files and preview on the right. Finished work waits in Reviews like a local merge request.

macOS, Windows and Linux · Free and open source (MIT) · Alpha

Every agent, through its official route

Pick the runtime for each thread. Tau starts the tool its maker ships and uses the sign-in you already have there. It never borrows a subscription for a tool it wasn't sold for.

Every thread in view

The list shows each thread with its project, title, branch, runtime and age, newest first. Press / to search. Threads that run on your other machines sit in the same list.

A thread can work in a Git worktree of its own. Three agents can change one project at once without touching each other's files.

Terminal, files and preview beside the chat

The stage opens next to the conversation. A terminal in the thread's folder, the files with the thread's diff, a browser preview of what the agent builds. Split it, maximize it, or open a file in your editor.

Check finished work like a merge request

When a thread is done, its branch lands in Reviews with what changed, which model did it and what it cost. Merge it, or send it back to the thread with a note. A branch that conflicts gets a button that asks its thread to rebase.

Know what's left of your plans

Small bars at the foot of the sidebar show your Codex and Claude Code limits, the way Juicebars does in the Mac menu bar. The Usage page adds up what you spent and what your plans would have cost at API prices, including the work you did in the CLIs outside Tau.

Put another computer to work

Pair a second Mac or a Linux box once, with a link or over SSH. Its threads show up in your list, and a new thread can start there. Automatic picks the machine with the most room when you send.

Work handed to another machine comes back as a branch. Nothing is pushed to your Git remote on the way. On a server, Tau runs as a service without a window.

Answer your agents from your phone

The app for iPhone and Android pairs with your computer through a QR code and shows the same threads. Read along, answer an agent's question, check a review, start the next task.

There is no account and no Tau server in between. The phone talks to your computer directly, on your network or over Tailscale, and checks its key on every connection.

Rebuild it yourself

Terminal, files, reviews, usage, even the connection to Claude Code are kits: packages that use the same interface that is open to you. If something bothers you, you don't have to wait for someone else. Write a kit, load it into your Tau, and it's there. Save the file and it reloads.

Download Tau

Release notes · Getting started

Tau is in alpha. Settings, the interface and the extension API can still change between releases. Feedback is welcome.

Updates itself

You install Tau once. After that it keeps itself current without a terminal: it looks for a new release every six hours, checks the download against the release's SHA-512 checksum, and installs it when no agent is working.

A machine you paired updates the same way, and Settings shows where each one stands.

On a Mac, Tau is signed with an Apple Developer ID and notarized by Apple, so it opens like any other app. The Windows installer isn't signed yet, so SmartScreen may warn the first time: choose More info, then Run anyway.

Thanks

Tau wouldn't exist without two projects. Pi gave me the idea that an agent is yours and that you can rebuild it. T3 Code showed how a workbench like this can feel, and I followed it in a lot of places. I have a lot of respect for both teams.