An AI agent that works in your folder of notes, plans, and task boards. It browses, plans, and proposes real changes, then waits for you to approve the diff before it writes anything.
Agent features
Ask for results, not only answers. The agent reads freely, writes only through proposals, and checks its own results.
The open documents (including unsaved ones), the selection, the cursor, and a map of the whole folder are assembled for you. Attach a whole file with @filename, and limit what is sent with the switches.
list_files, search_documents, search_text, and read_file. Answers quote the file and the line number.
“When was the release date changed and by whom?” Answered from git_log, show_commit, and file_at_commit, without ever changing the repository.
The sparkle button next to the commit box writes a one-line message from the diff of the checked files, in the style of your recent commits. You check it; nothing is committed until you press Commit.
Commit messages →Create a file, replace text (once or every occurrence), insert text, move to the Trash, or move and rename a file.
Available actions →Add, move, mark, edit, or delete cards, and add and rename lists. Moving a card only marks that card in the diff.
The board and the agent →Every proposal appears as a line diff with numbers and context. Nothing is written before you press Apply.
The review window →Up to 20 actions are reviewed at once. Uncheck a file to apply only part of it; the agent knows which were accepted.
Change batches →Add a note to your decision so the agent fixes its proposal. A change that has been applied can be undone safely.
Undo →A request with several steps is recorded with set_work: a goal and steps whose status is visible in the panel.
verify_work rereads the real result: text that must be present or gone, the position of cards, and the Markdown structure. Not passed, not finished.
Proposals, decisions, and snapshots are recorded together with the conversation. Interrupted work can be resumed after the app is opened again.
Recovery →Streamed answers, the context and token details of every answer, a choice of DeepSeek models, and the Deep thinking mode.
The Assistant panel →Conversations are saved as Markdown in .markwork/chats. Open them again and continue at any time.
A live timeline of every model round, the reasoning, the tool calls with their arguments and results, the tokens, and the duration.
The agent log →A card marked @pi is worked on by the pi harness in another project repo. Answer its questions and permission requests straight from the card.
How it works
Everything is a Markdown file in one work folder, so each step feeds the next.
1. Home
Open Markwork and Home already shows where you left off.
2. Inbox
Ideas, links, and quick notes go into the inbox in one line.
#tag
3. Journal
You jot short notes; Markwork adds what already happened.
4. Kanban
A kanban board is also just a Markdown file.
[due:: date] show up on Home
5. Agent
Ask in plain language; the agent reads your folder and proposes the work.
@pi goes to another agent in a project repo
Orchestrator
Write @pi on a kanban card, right-click, then Work on it with pi. The pi harness works in the project repo that you choose, and Markwork handles the rest.
Free and open. Its foundation is a Markdown editor that formats live, with a kanban board, Mermaid diagrams, and Git history, all in a light GTK 4 window.
For developers: learn to build an agent from Markwork, learn performance from its editor, learn GTK from how it grew, and learn clean code from its refactors.