Plonk

您的代理可驅動的 Mac 視窗管理程式。貼齊區域、已儲存工作區、保持喚醒、螢幕截圖與裝置端 OCR — 19 個 MCP 工具,橋接至您自己機器上的回環 API。無帳號、無雲端、無遙測。

文件

Plonk

A Mac window manager with zones you draw yourself. Put a window in one by dragging it, by shortcut, by voice, or by asking an agent over MCP.
To plonk is to set a thing down exactly where it belongs.

Version macOS 13+ Swift 6 No dependencies MIT MCP

ostapondo.github.io/Plonk

A pile of overlapping windows is cleared four ways in turn: one window is dragged into a zone as the zones light up, the next is sent to zone two with ⌃⌥2, the third by holding ⌃⌥V and saying “zone three”, and the last two are placed together from a sentence typed to an agent

Why another one

You most likely already have Rectangle, Magnet, Loop or Raycast. If a fixed grid of halves and thirds is all you want, they do it well. Plonk is for three things they do not do.

Zones you draw. Not a preset grid. Any number of zones, any size, a different set per monitor, overlapping if that suits you. Click a zone to split it, -click to split the other way: a narrow rail for chat, a wide middle split in two, a strip for the terminal.

Workspaces that remember which monitor. Save the desk: the apps, every window's frame, and the display each one belongs on. Launch it onto an empty desktop and it rebuilds itself, on the right screens.

An agent can drive all of it. Plonk ships an MCP server that covers its whole surface: layouts, workspaces, zones, keep-awake, screenshots, on-device OCR, measuring. Generic macOS automation servers can nudge a window around. This is the window manager itself, so "browser on the left 60%, terminal top right, save that as a workspace called review" is one sentence rather than a script.

The other seven tools, OCR and a ruler and keep-awake among them, are there because each was otherwise its own icon in the menu bar.

It is early. Version 0.2.x, one author. Shortcuts, zone files and workspaces are settled. The MCP tool names and the HTTP API are not, and can still change between minor versions. CHANGELOG.md says what moved.

Three ways to move a window

Drag it. Zones light up as you move a window, and it drops into one. Hold too and it takes two of them at once. Turn on grab-and-move to pull a window from anywhere inside it instead of aiming for the title bar.

Press a key. ⌃⌥← for the left half. ⌃⌥1 to ⌃⌥9 for the numbered zones on that screen. ⌃⌥0 puts a window back where it was before Plonk touched it.

Say it. Hold ⌃⌥V and name the place: "snap this left", "zone three". That runs in the app, offline, on-device.

Five zone sets ship with it, and past those you draw your own. Swap the set on a screen and any window already in a numbered zone moves to wherever that number is now.

Five clicks in the zone editor cut one screen into six zones, six windows fill them, and ⌃⌥⇧2 then swaps the whole set for another while the windows already in it follow their numbers across

Plonk's Zones page: zone sets across the top with Thirds selected and its three numbered zones drawn below, controls for drag-to-snap and the modifier that shows the zones, and the list of halves, quarters, maximize and centre with the shortcut bound to each

Install

macOS 13 or newer, Apple silicon.

brew install --cask ostapondo/plonk/plonk

Grant Accessibility when it asks, then relaunch. Screen Recording is asked for separately, the first time you capture. Nothing else: no Full Disk Access, no Automation, no Keychain.

Plonk is signed but not notarized, so macOS holds a copy you download by hand. The cask takes care of that for you.

Running it alongside Rectangle or Magnet is fine, as long as their shortcuts do not collide.

Installing by hand, Intel Macs, tiling managers, and removing it

Why macOS holds a downloaded copy. The certificate is self-signed rather than an Apple Developer ID, because notarizing needs a paid Apple account and this project does not have one. macOS cannot vouch for who built it, and says so. The cask skips that check by clearing the quarantine flag for you.

That is a check skipped on your behalf, so here is a stronger one to run before you open anything:

gh attestation verify $(brew --cache)/downloads/*--Plonk-*.zip \
  -R ostapondo/plonk

It prints the commit and the GitHub Actions run that built this exact archive. Apple's stamp would tell you a build passed a malware scan. This tells you the binary came from the source in this repository, with no laptop in between.

Without Homebrew. Download the latest release, unzip, drop Plonk.app into Applications, then clear the flag yourself, which is all the cask does:

xattr -dr com.apple.quarantine /Applications/Plonk.app

Or do the Gatekeeper detour once: open Plonk, dismiss the warning, then System Settings, Privacy & Security, scroll to Security, Open Anyway.

If you move or rename Plonk.app later, macOS ties the old grant to the old path and windows of newly launched apps stop being seen. Remove Plonk from Privacy & Security, Accessibility, and grant it again.

On an Intel Mac. Releases are built for Apple silicon only, so the download will not run. Building from source ought to work, see Build, but nobody has tried it and a report either way is welcome in issues.

Next to a tiling manager. yabai and Amethyst own every window on screen and will pull windows straight back out of a zone. Run one or the other.

Removing it. brew uninstall --cask plonk, or quit Plonk and drag it to the trash. Then delete ~/Library/Application Support/Plonk/. The login item goes with the app, and nothing was written anywhere else.

What you get

The window manager is four things, and this is the part of each that the sections above left out.

ZonesOverlap them, gap them, hide the numbers, keep a list of apps Plonk never touches. Windows return to their zone after a display is unplugged and plugged back in
WorkspacesFiles, folders or URLs each app should open on the way up, so a desk comes back with the right documents and not just the right apps. Monitors are keyed by UUID, so unplugging one does not scramble them
Focus that follows the layout⌃⌥⇧← goes to the window actually on the left, not the one you used last. ⌃⌥` cycles the windows stacked in one zone
VoiceHold ⌃⌥V and say it. Common commands run in the app, offline. Anything bigger goes to your agent. Recognition is on-device

And the MCP server, which is every one of these as a tool an agent can call.

Seven smaller things, behind the same menu bar icon
Text off the screen⌃⌥T selects an area and copies the words in it: a screenshot, a paused video, a dialog that will not let you select. On-device
A ruler⌃⌥R, then hover: how far the pointer can go each way before it meets an edge, read off the pixels. The width of a row, the height of a bar, the gap between two things. Drag for a straight-line distance. Points and pixels both
Pin part of the screenFloat a live crop above everything else. A build log, a chart, a call, visible in a corner while you work over it
Keep awakeReal power assertions, not a jiggler. Sessions end by themselves: after N minutes, at a wall-clock time, or when a process exits
ScreenshotsRegion, window or screen, then pen, arrow, rectangle, ellipse, highlighter. Saved at native resolution
A shortcut guideEvery shortcut the front app actually has, read from its own menus, so it is never out of date
Pointer toolsFind the cursor, ring every click for a recording, crosshairs, jump to the next display

Longer versions: Zones · Workspaces · Hotkeys · Everything else

For agents

An agent can read the desk, rearrange it, save the result, and read text back off the screen, with no screenshot round trip for anything that is really just words.

browser on the left 60%, terminal top right, notes bottom right
save that as a workspace called "review"
keep the Mac awake until this build finishes
read the error out of that dialog and tell me what it says
how tall is that toolbar, in points and in pixels

Twenty tools across state, layouts, workspaces, zones, keep-awake, screenshots, on-device OCR and measuring. Frames are fractions of a monitor's visible area, origin top-left, so "left 60%" is {x: 0, y: 0, w: 0.6, h: 1}. Several agents can connect at once, each registering itself, with an optional mode that locks changes to the active one.

Four sentences typed to an agent: a layout is applied across the desk, the result is saved as a workspace called review and the Mac held awake until a build finishes, everything is then closed to an empty desktop and rebuilt from that workspace, and finally a line of error text is read off the screen

Setup, if you want the plonk CLI or an agent driving it (Node 18+):

claude mcp add plonk -- npx -y plonk-mcp   # Claude Code
codex mcp add plonk -- npx -y plonk-mcp    # Codex CLI

In Claude Code it can also be a plugin: same server, pinned to the release it shipped with rather than to whatever npm serves as latest.

/plugin marketplace add ostapondo/plonk
/plugin install plonk@plonk

For Claude Desktop there is nothing to type. Download plonk-<version>.mcpb from the latest release and open it. The bundle carries the server and its dependencies, so no config file is edited and nothing is fetched at launch.

Any MCP client works, over stdio or HTTP. One-pagers for Cursor, Zed and Cline.

The same package carries a plonk command, for the things that are neither an agent nor a settings window:

plonk state                      # screens, zone sets, workspaces, windows
plonk launch review              # a saved workspace
plonk awake while npm run build  # awake for exactly as long as the build
plonk text | pbcopy              # OCR a region into the clipboard
plonk measure 0.5 0.5            # size of what is mid-screen, in points and pixels

For agents has every tool, the multi-agent rules, the HTTP transport and the rest of the CLI.

Privacy

No account, no cloud, no telemetry. The API binds to 127.0.0.1, refuses anything carrying headers a browser cannot suppress, and is gated on a token only you can read. The one outbound connection is the update check, which carries no identifier and can be switched off.

None of that is a claim you have to take on trust. Releases are built and signed on GitHub's runners and ship with an attestation, so the binary on your Mac ties back to the commit it came from:

gh attestation verify Plonk-<version>.zip -R ostapondo/plonk

Check it yourself is every claim above with the command that tests it. SECURITY.md says where each promise stops.

Under the hood

Claude talks to the MCP server over stdio, which calls the app's loopback HTTP API

  • The app is the single source of truth. The MCP server is a stateless bridge.
  • App/ is the Swift menu bar app, mcp/ the TypeScript MCP server.
  • Config is plain JSON at ~/Library/Application Support/Plonk/config.json.

Build

Five commands, and they are what CI runs on every pull request. Each line is a subshell, so paste the block from the repository root.

(cd App && swift build)                    # the app compiles
./scripts/test.sh                          # the unit suite
./scripts/lint.sh                          # style rules, no dependencies
(cd mcp && npm ci && npm test)             # the MCP server
node scripts/check-zone-sets.mjs           # the layouts in zone-sets/

None of that needs a signing certificate. Zone geometry, config decoding, HTTP routing, MCP tools, voice parsing, the CLI and every document here are reachable from that loop, and most changes need nothing more.

Producing a launchable Plonk.app does need one. Make your own once with ./scripts/make-signing-cert.sh, then run ./scripts/build.sh. macOS ties Accessibility and Screen Recording to the code signature, and an ad-hoc one changes every build, so a stable certificate is what stops rebuilds from resetting permissions.

Contributing

Bug reports, zone sets, client one-pagers and code are all welcome. None of them need a signing certificate.

  • The smallest useful change is one JSON file. zone-sets/ is a gallery of layouts worth copying: an ultrawide split, a rotated monitor, the one built around a recurring meeting. Draw it in the app, read the numbers out of plonk state --json, open a pull request. That folder has its own CI job and answers in about twenty seconds. No build, no signing, no Swift.
  • good first issue issues are written to be picked up cold. Each says where the code is and how to tell it worked, and carries a prompt you can hand to an agent, since AGENTS.md already explains the repo to one.
  • needs-hardware issues need neither Swift nor a certificate, and are the most useful thing anyone can send. A window manager breaks on arrangements the author cannot see, so a report from three monitors or an ultrawide is worth more than a patch.

CONTRIBUTING.md has the rest, including how long a review takes. Questions and half-formed ideas go to Discussions. A security problem goes through SECURITY.md, not a public issue. Everyone taking part follows the Code of Conduct.

License

MIT © ostapondo