How it works

Chats and workflows

A chat starts empty. You attach one or more workflows to it by typing a command. Each attachment is a workflow, optionally on a project folder, and the chat keeps them as a numbered list. The same folder can have several workflows.

A workflow that writes files needs a folder. If you type its name alone, Workpen opens a dialog to choose the working directory.

Stages

A workflow moves through stages. The one included with Workpen interviews you, writes a plan, and builds the files. At the end of a stage the model asks for your approval, and an Approve button appears in the chat. Typing a message instead cancels the approval and keeps you in the stage.

Once the last stage is done the workflow stays in an assist stage for small requests, such as changing a heading or adding a page. If you attach a workflow to a folder that already has files, it starts by reading them instead of interviewing you.

Project folders

The model works in your folder with a few tools: list files, read a file, edit a file, write a file, delete a file. Before it changes a file it must have read it, and if the file changed since, it has to read it again. That keeps your hand edits safe. At the start of each turn the chat shows what changed in the folder since the last one.

History

Workpen records your project folder at the end of every turn, including your own edits, in storage of its own. Your folder gets no extra files, and an existing Git repository in it is left alone. Open History from an attachment's menu to see each turn, the files it changed and the differences, and to restore a turn. Workpen needs Git installed to do this, and shows nothing about Git itself.

Models

Workpen talks to services that follow the OpenAI chat format, which includes OpenAI, Gemini through its compatible endpoint, and local servers such as Ollama. The model must be able to call tools. You add a service in Settings with its address, the model names, and your API key.