Why I built Workbench
I set out to turn requirements into code. I ended up building the place where that work happens.
Workbench started life in a very different shape.
In late 2025, I began building a cloud platform under the Req2Code name for taking a product idea from requirements to working software. It had a web workspace, a requirements graph, a planning view and an assurance layer for checking delivered work against what was asked for.
The thinking was simple. As AI coding agents became better at writing software, more of the work would move to describing what to build and checking what had been delivered.
That need was real. But while I built the platform, I discovered that I needed a different place to do the work.
Where the work happened
Day after day, development happened on my MacBook Pro, in a terminal, with an agent running in the repository: at the coffee shop, at the kitchen table, at the airport gate.
I could give several agents work at once, but they still needed steering. Small decisions needed making. Bugs and bottlenecks needed discussing. That was easier where I could run the environment, test a change and try again.
My desktop told the story:
- A terminal window per project, each with a stack of tabs I struggled to tell apart.
- Codex in one pane, Claude in another, and no clear view of who had finished or needed an answer.
- Plans and decisions spread across transcripts, Markdown files and notes apps.
- Screenshots saved on the desktop and dragged into whichever conversation needed them.
The web platform helped describe the work. I also needed a workspace alongside my local shells, Git setup and toolchains.
The moment it clicked
I wanted my terminal sessions, Codex, Claude Code and Grok in the same app, using the accounts I already had and sharing the same project knowledge.
My MacBook Pro was where I worked. So I started again on the Mac and asked a narrower question: what would a workspace for developers working alongside AI agents look like?
A workspace organised around the project
That question became Workbench’s brief:
- Keep the project together. Terminals, conversations, documents and memory belong with the project folder. Switching projects should bring that context with me.
- Fit the existing environment. Use my login shell, installed tools and Git setup. Connect to AI tools through their local command-line tools and my own accounts, subject to each provider’s requirements.
- Keep knowledge across sessions. Record decisions and plans where another agent can find them, without relying on me to repeat every conversation.
- Make planning part of development. Keep design documents and milestone roadmaps close to the repository and the agents doing the work.
- Make attention visible. Help me see when an agent is working, finished or waiting for a decision.
Workbench is a native macOS app with a local workspace. That does not make an AI conversation entirely on-device: cloud AI tools send prompts and relevant context to their providers. The security page explains the boundaries.
Built in itself
Since mid-2026, I have developed Workbench inside Workbench, using a separate daily build alongside the one under development. Its terminals, conversations, project memory and roadmaps carry much of the project’s planning history.
Daily use shaped the details. Queued follow-up messages came from thinking of the next instruction while an agent was still working. Provider switching came from wanting another model to review a result without manually moving context between apps. A recent-work summary came from opening the laptop and needing to remember where I had stopped.
The original Req2Code ambition is still there: help people turn requirements into working software. Workbench puts that work in the place where I was already doing it.
Next in the series
The next two posts explore the frustrations behind the product:
- Too many agents, too many windows: keeping several agents and projects moving.
- Agents forget. Your project shouldn’t: keeping decisions, plans and progress available across sessions.
Explore Workbench or download Workbench.
Workbench requires macOS 27 or later on Apple Silicon. AI tools are installed and signed in separately. Provider requirements and feature support vary.