ai-workbench: Terminal cockpit for AI development

Four windows, one directory

You navigate to a project in the file browser – and the AI assistant, LazyGit and your shell are in the same directory. No cd in three windows, no searching for which terminal is where.

That is the core of ai-workbench: a multiplexer for the terminal interface, written in Rust. On the left the file tree, on the top right the preview with syntax highlighting, on the bottom right three switchable areas for AI assistant, LazyGit and shell.

To the project on GitHub →

Runs wherever a shell runs

This is where a terminal tool differs from an IDE: You connect via SSH to a server – and start the complete cockpit there. File browser, AI assistant, Git and shell run on the machine where the code is located.

No local client, no extension that installs a counterpart, no file synchronization between two computers. A single binary on the target system is sufficient – even on a machine without any graphical interface.

So that it doesn't fail in everyday life due to trivialities, a lot of work goes into the margins:

• Clipboard over SSH – five methods in sequence, up to the OSC-52 sequence, which copies through the terminal connection itself.
• Preview and export from the server – an exported PDF lands via the existing connection in your local download folder, instead of starting a console browser that messes up your interface.
• Insert images over SSH – the setup wizard detects an SSH session and explains how to build the bridge to your local clipboard.
• XRDP and Kitty are explicitly tested together, not just the local terminal.

Practically, that means: the same workspace on the laptop, on the development server, and on the client machine – via the same SSH connection that you already have open.

Why not just an IDE?

Because modern development environments get heavier with each version. Visual Studio Code starts slowly and brings features that only get in the way in daily workflows. I tried Zed and Google IDX – both well made, both lack the shell-first mindset.

What I wanted: an environment that responds as quickly as the thought before it. Over the turn of the year 2025/2026, ai-workbench was born from that.

No overhead. A single binary. Starts in under a second.

Six AI backends, one interface

The AI area is not tied to one provider. You choose at startup – or switch at runtime with F8:

The last two run locally via Ollama – without code leaving your own network. The tool remembers your choice until the next start.

What you get from it in everyday life

• File browser with Git status – Colors show you immediately what has been changed, new, or is in conflict. Also summarized for entire directories.
• Preview with syntax highlighting for over 500 languages, plus rendered Markdown and a built-in editor with search and replace.
• Mouse support that feels right – click to focus, wheel to scroll. If an application with its own mouse control runs in an area, it receives the events passed through.
• PDF and HTML export from Markdown files, directly from the preview.

Installation

On Mac and Linux with Homebrew:

brew install eqms/ai-workbench/ai-workbench

Alternatively, you download the appropriate archive directly from the Release page – there are builds for Linux and macOS as ARM64 and x86_64, plus Windows packages. The archives are signed.

You can update from within the tool: The help page checks on request whether a newer version is available, and replaces the binary in place.

What you should bring

git is mandatory. lazygit you need for the Git area, and the CLI of your chosen AI backend must be in the path. On first start, a setup wizard checks what is missing and tells you how to install it.

For the clipboard under X11, xclip or xsel, under Wayland wl-clipboard. Without these programs, the OSC-52 sequence is used – it works over SSH as well, but cannot read back.

Honestly

ai-workbench does not replace an IDE. If you need graphical debuggers, refactoring across the entire project, or a database interface, you are better off in a proper development environment.

What it does well: keep you in the terminal when you are working there anyway – with AI assistant, Git, and Shell side by side instead of in four windows. And on any machine you can connect to.

The source code is openly available under the MIT license. Check it out on GitHub, try it out, and write to me what is missing.