desk.md is a local-first workspace for projects, tasks, docs, and meetings, stored as plain Markdown files you own, and structured so an AI agent can find its way around them. It runs as an offline desktop app, or you can self-host it and connect Claude, ChatGPT, or Claude Code to your work over MCP.

I built it for my own work and use it every day to keep track of everything from client projects to personal tasks. The point isn't to generate the notes, it's to have a nice note-taking and project-management environment that happens to keep real context somewhere an agent can reach it in a structured way.

• Projects and tasks. Workspaces, projects, tasks, docs, and meetings as real things, with statuses, priorities, due dates, Kanban or list views, a week planner, and a quick-capture inbox. Built in, not assembled out of plugins.
• Plain Markdown. YAML frontmatter in an ordinary folder. No database, no lock-in, nothing to export, since the files already are the format. Drag in Word, PDF, Excel, CSV, or HTML and they convert to Markdown docs.
• Agents can navigate it. The Smart Index summarizes your files into a catalog agents can search, and desk.md writes CLAUDE.md, AGENTS.md, and GEMINI.md so local coding agents understand your projects without you re-explaining them every session.
• MCP from anywhere. Self-host desk.md and your workspace becomes an OAuth-protected MCP endpoint, so you can browse, search, and read your projects from claude.ai, ChatGPT, or your phone. Not a localhost port that needs the app awake on one machine.
• Local or self-hosted. Run fully offline on your desktop, or host it on your own server for browser/PWA access and MCP.

Screenshots

Dashboard	Projects
Docs	Meetings

Why not just Obsidian?

Obsidian gives you notes with properties. Everything above (projects, task statuses, boards, meetings) you assemble yourself out of community plugins, and the stack is fragile: the Projects plugin, which came closest to this, was discontinued in 2025. Bases covers tables and cards now, but Kanban still isn't there, and every schema is one you define and maintain.

desk.md has the model built in, and the agent side is a real difference: desk.md self-hosts an OAuth-protected MCP endpoint, so your work is reachable from claude.ai or your phone.

The files stay plain Markdown either way, so nothing stops you opening them in Obsidian, an editor, or grep. You just won't need to.

AI & agents

desk.md isn't an agent platform. It keeps your context organized and reachable, then lets AI tools read it when you want them to. There's no chat box in the app; the sparring happens in the tool you already use, with desk.md underneath it as the source of truth.

Records and context. Your files split by lifecycle. Docs, tasks and meetings are records: they are dated, they accumulate, and nothing rewrites them. A March meeting note stays a March meeting note, so a record cannot go stale. Everything an agent needs to actually understand the work lives in context, a small folder kept current: what this is, which systems it touches, what was decided.

The brief and the current state. Each project's context holds two small files with a clean line between them. The brief is yours: what this project is and what it runs on, the part no AI can work out for itself. The current state is the AI's: a short snapshot of where the project stands, rewritten in the background as your tasks, docs, and meetings change, and marked as AI-written right in the file. This runs wherever your data lives: on your desktop with your local key, or on your own server when self-hosted. desk.md tracks how many records the snapshot hasn't seen yet, so you can tell a current map from a rotting one at a glance.

Your words are protected structurally, not by asking the model nicely: the AI only ever writes its own file, so nothing you wrote is ever in an AI write path.

Smart Index. Optional, and off until you add your own Anthropic or OpenAI key (Settings → AI). It summarizes your files into a catalog so agents can find the right note fast instead of crawling the whole folder.

Local agent files. Because your data is just Markdown, local tools like Claude Code, Codex, and Gemini CLI can open the folder directly. desk.md writes CLAUDE.md, AGENTS.md, GEMINI.md, and WORKSPACE_CONTEXT.md so those agents know what they're looking at.

Hosted MCP. If you self-host desk.md, the server exposes an MCP endpoint over OAuth. Claude.ai, ChatGPT, Claude Code, and other MCP clients can browse, search, and read your workspace over the network. Useful for questions like "what did we decide about this project?", "find the latest client notes", or "draft a reply using the related meeting notes and tasks". The tools are read-only for now; letting an agent file something back when you ask it to is on the roadmap.

Running it: local or self-hosted

desk.md runs the same app two ways. Installed locally, it's an offline desktop app that reads and writes Markdown straight from your disk. Self-hosted, the same app runs on a server you control, so you can reach your desk from a browser or phone and let AI tools connect over the network. Either way your data stays plain Markdown you own; the difference is where it lives and who can reach it.

	Local (desktop app)	Self-hosted (your server)
Install	Native app on Mac, Windows, Linux	docker compose on a box you control
Reach it from	The one machine it runs on	Any browser, on any device, all sharing one copy. Installable as a PWA on phone and tablet
Where data lives	A folder you pick at setup (default ~/DeskMD/)	A folder on the server you bind-mount into the container
Multi-device sync	Your own tool (iCloud, Dropbox, git, Syncthing)	Built in, the server holds the shared copy
Login	None, fully offline	Account and session (first user wins), needs the network
Smart Index	Yes, with your own API key	Built by the server with its own key (set it in the deploy .env), so browser and phone get it too
External AI agents	Point Claude Code, Codex, or Gemini at the folder	Connect Claude.ai, ChatGPT, Claude Code, or any MCP client to your workspace
MCP	No server needed; agents read local files	OAuth-protected MCP endpoint, read-only for now

Both modes write the same folder: your Markdown under workspaces/, app metadata under .desk/. So a folder made by the desktop app can be handed to a server, and back, with nothing to migrate. Hosting adds one file, .desk/auth.sqlite, for accounts and sessions. Everything else is Markdown you can open, back up, or git like any other directory.

The native desktop app can also point at a self-hosted server instead of local disk, giving you the desktop UI on top of server-side data.

Self-hosting is one container (docker compose up) that serves the web/PWA app, the API, the OAuth server, and the MCP endpoint. See deploy/README.md.

Install

Download the build for your OS from the latest release. desk.md then keeps itself up to date automatically. To build from source, see Run from source.

Windows and Linux builds are beta. desk.md is developed and tested on macOS. Please report anything broken.

• macOS (Desk_*.dmg, one universal build for Apple Silicon and Intel). Drag Desk into Applications. The app isn't notarized by Apple yet, so macOS calls it "damaged" on first launch. Clear the quarantine flag once: xattr -dr com.apple.quarantine /Applications/Desk.app
• Windows (.exe installer). Not code-signed yet, so SmartScreen warns: More info → Run anyway.
• Linux (.AppImage, .deb, or .rpm). For the AppImage, chmod +x Desk_*.AppImage and run it. Storing AI API keys needs a desktop secret service (GNOME Keyring or KWallet), present on most desktop installs.

To self-host instead, see deploy/README.md.

Roadmap

Ideas I'm exploring:

• More project-management depth, still lightweight: timelines, milestones, task dependencies
• MCP write tools, so an agent can file a note or a task back into the workspace when you ask it to
• Ollama support, so the Smart Index can run on a fully local model
• Server-side AI, so self-hosted installs get the Smart Index without a desktop
• Mobile polish: make the hosted PWA better for quick capture and read-only review
• Time tracking: log time on tasks with simple per-project reports

Run from source

desk.md is organized as an npm-workspaces monorepo under packages/: @desk/core for the domain layer, @desk/app for the Tauri and React UI, and @desk/server for the Node/Hono self-host server with the domain API, web/PWA, OAuth, and MCP. All commands run from the repo root:

npm install
npm run dev          # Browser with mock data (fast UI loop, no Rust needed)
npm run tauri:dev    # Desktop app with the real file system

Use Node 22. Newer versions currently break Rollup's native dependency. See CONTRIBUTING.md for the full setup.

Contributing

Bug reports, feature ideas, and pull requests are welcome. See CONTRIBUTING.md for setup and guidelines.

To regenerate the screenshots above after a UI change: npm run screenshots.

License

GPL-3.0-or-later. You're free to use, modify, and share desk.md. Because the GPL is copyleft, any distributed fork or derivative must also stay open-source under the GPL.