plonk-mcp
Mac window manager AI agents can drive: see every monitor and window, apply multi-monitor layouts in one call, save and relaunch whole workspaces, snap zones, keep-awake, and annotated screenshots. Backed by a native macOS menu bar app over a local-only loopback API — no cloud, no telemetry.
README
<h1 align="center">Plonk</h1>
<p align="center"><strong>The Mac window manager that puts your desk back together.</strong><br> <sub>To plonk is to set a thing down exactly where it belongs. This menu bar does it to your windows — you drag them there, or your agent says where.</sub></p>
<p align="center"> <img alt="Version" src="https://img.shields.io/badge/version-0.0.5-58a6ff?style=flat-square"> <img alt="macOS 13+" src="https://img.shields.io/badge/macOS-13%2B-111?style=flat-square"> <img alt="Swift 6" src="https://img.shields.io/badge/Swift-6-F05138?style=flat-square"> <img alt="MCP" src="https://img.shields.io/badge/MCP-18_tools-8957e5?style=flat-square"> <img alt="No dependencies" src="https://img.shields.io/badge/dependencies-0-2ea043?style=flat-square"> <img alt="MIT" src="https://img.shields.io/badge/license-MIT-blue?style=flat-square"> </p>
<p align="center"> <img src="docs/demo.gif" alt="An agent is told where the windows go, arranges them, saves the setup as a workspace, and launches it back onto an empty desktop" width="720"> </p>
Drag a window, the zones light up, drop it in. Or skip the dragging and say it:
browser on the left 60%, terminal top right, notes bottom right
save that as a workspace called "review"
keep the screen awake for the next hour
screenshot the screen and tell me what looks off
Everything runs on your Mac. No account, no cloud, no telemetry.
Install
macOS 13+.
brew install --cask ostapondo/plonk/plonk
Or download the latest release, unzip, and drop Plonk.app into Applications.
First launch takes one extra click. Plonk is code-signed, but with a self-signed certificate rather than an Apple Developer ID, and Gatekeeper does not trust one of those — so macOS holds the app the first time, whichever way you installed it. Open Plonk, dismiss the warning, then go to System Settings → Privacy & Security, scroll to Security, and click Open Anyway. That is once, for good.
What that warning means is that macOS cannot vouch for who built this. Fair. So rather than ask you to take it on trust, everything below is checkable — starting with proving the file you just downloaded was built from the source in this repository.
Grant Accessibility when asked, then relaunch. Screen Recording is asked for separately, the first time you capture. Nothing else — no Full Disk Access, no Automation, no Keychain.
If you later move or rename Plonk.app (or its folder), macOS quietly ties the old grant to the old path: windows of newly launched apps stop being seen. Remove Plonk from Privacy & Security → Accessibility and grant it again.
To let an agent drive it (Node 18+):
claude mcp add plonk -- npx -y plonk-mcp # Claude Code
codex mcp add plonk -- npx -y plonk-mcp # Codex CLI
Any MCP client works the same way — give it npx -y plonk-mcp as a stdio
server. One-pagers: Cursor (with a one-click
install button), Zed, Cline.
Several clients at once is fine; see Agents below.
A client that cannot spawn a process connects over HTTP instead:
npx -y plonk-mcp --http serves Streamable HTTP at
http://127.0.0.1:43918/mcp (loopback only, many clients per process,
--port to change).
Or build everything from source: clone the repo, run ./scripts/build.sh, and
point claude mcp add plonk -- node …/mcp/dist/server.js at a locally built
server (cd mcp && npm install && npm run build).
Check it yourself
Accessibility is the only way macOS lets one app move another's windows, and Screen Recording is what a screenshot costs. That is a lot to hand something you installed a minute ago, so none of this is a promise — it is all checkable.
The binary comes from the source. Releases are built, signed and zipped by a workflow on GitHub's runners, never on a laptop, and ship with a provenance attestation GitHub signs:
gh attestation verify Plonk-<version>.zip -R ostapondo/plonk
That prints the commit and the workflow run the archive was built by. It is the step that makes reading the rest of this repo worth anything — without it, the code here and the app on your Mac are two separate claims. (Releases up to and including 0.0.4 were zipped by hand and carry no attestation, so the command fails on those. That is the whole reason it exists now.)
The MCP server is published the same way, which matters more, because npx -y plonk-mcp fetches it every time: npm view plonk-mcp dist.attestations, or the
Provenance panel on its npm page.
One thing dials out, and you can switch it off. Every socket the app has open:
lsof -nP -i -a -p "$(pgrep -f 'Plonk.app/Contents/MacOS/plonk')"
plonk … TCP 127.0.0.1:43917 (LISTEN)
One listener on loopback. The only outbound connection Plonk makes is the
update check: on launch and once a day it asks api.github.com for the latest
release, and sends nothing but a User-Agent naming the app and its version — no
identifier, no account, no analytics, no crash reporter. Turn it off under
Updates and it stops happening — including for agents, which get a 409 rather
than a connection made on your behalf, so the buttons on that page are the only
thing that can trigger one. nettop or Little Snitch will then show a process
that only ever listens. The URLs compiled into the app are that endpoint, the
releases page, and the issue tracker that opens when you click Report a bug —
Release.swift has all three.
A web page cannot drive it. The API is loopback-only and unauthenticated, so it refuses anything carrying headers a browser cannot suppress:
curl -so /dev/null -w '%{http_code}\n' -H 'Origin: https://example.com' \
http://127.0.0.1:43917/state
403
There is not much else to hide. Package.swift declares
no third-party dependencies, so a build from source is this repo and nothing
else. Config is plain JSON at ~/Library/Application Support/Plonk/config.json.
Screenshots go where you send them. There is no account to make.
The MCP server is a separate npm package that depends only on the official MCP
SDK and zod. It speaks to 127.0.0.1:43917 and nowhere else.
SECURITY.md has the rest: the entitlements the bundle ships with (none), every step the updater takes before it replaces anything, what the signing certificate does and does not prove, and where each of these checks stops being one.
Workspaces
<p align="center"> <img src="docs/workspaces.svg" alt="A workspace of four windows, saved, closed to an empty desktop, then launched back into place" width="720"> </p>
A workspace is a desk you can put away. It remembers the apps, the frame of
every window, the monitor each one belongs on, and what each app should open on
the way up. Launching one opens whatever is closed, waits for the windows, and
puts them back — from the Workspaces page, or right-click the menu bar icon.
Rename, recapture or delete from the workspace's ⋯ menu.
| Per app | Files, folders or URLs to open with it: a project folder for an editor, a set of tabs for a browser |
| Per monitor | Windows return to the display they were captured on, keyed by display UUID so unplugging a monitor does not scramble them. Or pull the whole workspace onto one screen |
| Already open | Running apps get moved, not relaunched. Turn that off to leave them alone and only open what is missing |
| The catch | macOS cannot open an app straight into a position, so windows appear first and jump a moment later. A second window of the same app cannot be conjured — give it a file to open instead |
Zones
<p align="center"> <img src="docs/zones.svg" alt="A screen split into three zones, with a window being dragged into the highlighted one" width="720"> </p>
<p align="center"> <img src="docs/zone-sets.svg" alt="Five built-in zone sets and a sixth, irregular one drawn by hand" width="720"> </p>
Five sets ship with it. Everything past that you draw yourself: any number of zones, any size, overlapping if you want — a narrow rail for chat, a wide middle split in two, a strip for the terminal. Or describe it and let the agent build it.
| Editor | Click to split, ⇧-click to split vertically, drag a divider to resize neighbours, ✕ to delete and let them heal over the gap |
| Per monitor | Each screen gets its own set, remembered by display, not by index |
| Overlap | Allowed — the smallest zone under the cursor wins |
| Trigger | On drag, or only with a modifier held. Holding it inverts the mode, so a free move stays one keypress away |
| Or none | Edge snapping instead: middles are halves, top is maximize, corners are quarters |
Hotkeys
<p align="center"> <img src="docs/hotkeys.svg" alt="Where each hotkey puts the front window" width="720"> </p>
<p align="center"> All on <code>⌃⌥</code>. Plus <code>⌃⌥Z</code> to flash the zones and <code>⌃⌥S</code> to grab a region. </p>
And the rest
| Keep awake | IOKit power assertions, not a jiggler. Display-on or system-only, pause on battery, auto while charging, timed sessions, and a menu bar icon that glows while it holds |
| Screenshots | Region, window or screen through the native picker, then pen, arrow, rectangle, ellipse and highlighter. Saves at native resolution |
| Notices | A panel in the top-right corner, not Notification Center: no permission to ask for, nothing left in your history, and it can show the screenshot instead of describing it |
| Updates | One button on the Updates page. The download is checked against the checksum GitHub published for it before it is unpacked, and Plonk installs a build only if it is signed with the same certificate as the copy you are running — the same test macOS applies, so your Accessibility and Screen Recording grants carry over instead of being asked for again. Anything that fails is discarded and nothing is replaced. Switch the check off and the app never looks |
For agents
Frames are fractions of a monitor's visible area, origin top-left — which is why
"left 60%" is just {x: 0, y: 0, w: 0.6, h: 1}.
| Tool | |
|---|---|
get_state |
Monitors, every open window and where it sits, zone sets, saved workspaces, awake status |
apply_layout |
Place any set of windows, across any number of monitors, in one call |
save_workspace · launch_workspace · delete_workspace |
Named desktops, launched from nothing |
snap_window |
Drop a window into a numbered zone |
save_zone_set · assign_zone_set · delete_zone_set |
Snap zones, per monitor |
set_awake |
Keep-awake, optionally time-limited |
take_screenshot · annotate_screenshot |
Capture, mark up, hand the image back |
select_agent |
Make an agent the user's active one, optionally the only one allowed to control |
Several agents can be connected at once. Every client registers itself, so
get_state lists who is online; the user picks an active agent from the menu
bar or the settings — or an agent does it with select_agent. An optional
strict mode locks changes to the active agent: everyone else keeps reading
state and taking screenshots, but gets a clear 409 on anything that moves
windows or edits config. Set PLONK_AGENT_NAME in a client's MCP config to
tell two sessions of the same client apart.
Under the hood
<p align="center"> <img src="docs/architecture.svg" alt="Claude talks to the MCP server over stdio, which calls the app's loopback HTTP API" width="760"> </p>
- The app is the single source of truth; the MCP server is a stateless bridge.
- The API binds to
127.0.0.1and refuses anything carrying browser headers, so an open web page cannot drive your desktop — see Check it yourself. - Config is plain JSON at
~/Library/Application Support/Plonk/config.json.
Build
cd App && swift build # the app
./scripts/test.sh # 191 unit tests
./scripts/build.sh # produces Plonk.app
cd mcp && npm run build # the MCP server
App/ is the Swift menu bar app, mcp/ the TypeScript MCP server. Point an
agent at AGENTS.md before it touches either.
build.sh signs with a Plonk Dev keychain identity and stops if it is
missing. macOS ties Accessibility and Screen Recording to the signature, and an
ad-hoc one changes every build, so create that certificate once (Keychain
Access → Certificate Assistant → Create a Certificate → type "Code Signing",
name it Plonk Dev) and rebuilds stop resetting permissions. Set
PLONK_SIGN_IDENTITY to sign with a different one.
If a permission was first granted while the app was ad-hoc signed, the old grant is pinned to a signature that no longer exists and every rebuild looks like a reset. Clear it once and grant again:
tccutil reset ScreenCapture dev.plonk.app
tccutil reset Accessibility dev.plonk.app
Releases: bump MARKETING_VERSION and BUILD_NUMBER in
version.env and version in mcp/package.json, then push a
v<version> tag. The release workflow builds,
signs and attests the app on GitHub's runners, uploads the zip to a draft
release, and publishes the MCP server to npm with provenance. Nothing ships from
a laptop, which is what makes gh attestation verify mean anything.
scripts/release.sh is what that workflow runs, and it works locally too. It
holds every build to the requirement in
scripts/release-requirement — a release signed
with anything else cannot be updated to and takes Accessibility and Screen
Recording away from everyone who installs it by hand. It notarizes when a
Developer ID certificate is in the keychain and says so when there is not;
Plonk's releases are not notarized, which costs a paid Apple account.
License
MIT © ostapondo
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
E2B
Using MCP to run code via e2b.