Markwork Docs

Getting started

Basic concepts

Four principles that explain how Markwork works. When a behavior feels surprising, the answer is usually here.

The work folder is the context

Markwork works in one folder that you open: notes, documents, research, plans, and task boards. The agent receives a map of all the Markdown files in that folder, the contents of the open documents (including unsaved ones), the selected text, and the cursor position. It browses the rest on its own with the read tools.

Without an open folder, the agent only sees the active document and cannot propose new files or changes to other files.

Everything stays plain Markdown

Documents, kanban boards, and conversation history are stored as Markdown files. There is no database, closed format, or account. Everything can be read by other editors, diffed, and committed to Git.

The agent reads freely and writes through proposals

The read tools (browsing, reading files, reading Git history) run right away. The write tools do not write anything: every action becomes a proposal shown in the review window. Files are only touched after you press Apply.

Write boundary

A proposal may only touch Markdown files in the work folder. Names outside the folder, starting with a dot, or containing .. are rejected. A file that changed since it was proposed is not overwritten.

Transparent about context and cost

Every answer shows the context that was sent, the lookups that were made, and the token usage. The context switches decide which documents may be sent to the model provider. The Agent log window shows every model round and tool call live.

Two kinds of agent

The Markwork agentAn external harness (pi)
Works inThe Markwork work folderAnother project folder, e.g. a code repo
Triggered fromThe Assistant panelA kanban card marked @pi
WritesOnly through proposals you approveFreely in the project folder; you review in that repo
FilesMarkdown onlyAnything, including running commands

See The pi orchestrator for the details of the external harness.