Visual context for AI coding agents
Capture anything. Paste everything.
Take screenshots the way you always do. They fall into a bucket. Press ⌘V in your terminal and Claude Code attaches every one as an image.
Open source, MIT. Runs on macOS 14 Sonoma or later.
Try it
Show Claude Code
what you see.
Capture from any window, then press Paste in the terminal. The paths land on one line, Claude Code attaches them as images and the bucket empties itself.
The bucket is empty. Press Capture.
How it works
Capture. Collect.
Paste. Empty.
A clipboard with memory. Not a screenshot editor, not a document manager.
-
01
Capture
Press ⌘⇧4 and select an area. The selection is the native macOS one.
-
02
Collect
Take as many as you need. The menu bar icon and the panel show what is in the bucket.
-
03
Paste
⌘V in the terminal pastes the paths of every screenshot on one line. Claude Code attaches them as images. No saving files, no dragging them in.
-
04
Empty
When the paste happens, the bucket empties itself. ⌥⇧V puts the last batch back on the clipboard.
Made for the terminal
Built around Claude Code.
You work with a terminal agent and show it what you see: an error, a layout, a chart, a settings screen.
Bucket pastes plain file paths on one line, separated by spaces, and Claude Code turns every one into an image. A path with spaces or shell characters is quoted the way Finder does it, so it cannot run anything. Tested in iTerm2 with Claude Code.
Memory
Every screenshot stays findable.
Ask your agent for the screenshot where it said “invoice total”, and it finds it. The OCR runs on your Mac.
A Claude Code skill in the repo teaches the agent to run Bucket search, a local command that only reads. It searches by the text in the screenshot or by the app that was in front, and returns the file path so the agent can open the image itself. No MCP server.
Outside the terminal, too.
Bucket lays out the clipboard for the app in front. Browsers get one image per screenshot, and other apps get each screenshot as a file and as an image. Tested with claude.ai and chatgpt.com in Safari, and with the Claude and ChatGPT desktop apps.
In those desktop apps the bucket does not empty itself: they read the clipboard before you paste. Clear it from the menu or the panel. Other terminals (Terminal, Ghostty, Warp, VS Code, Cursor…) are expected to behave like iTerm2, and other terminal agents may accept plain paths; neither is tested or promised yet.
Install
Get Bucket.
One line in your terminal.
brew install geheime/tap/bucket
- First launch
- Open Anyway, if you used the zip.The zip is not notarized, so macOS blocks it the first time. Go to System Settings › Privacy & Security and click Open Anyway.
- Allow Screen Recording.macOS asks the first time you capture. Turn it on and capture again.
- Free ⌘⇧4.macOS owns that shortcut. Untick it in System Settings › Keyboard › Keyboard Shortcuts › Screenshots, or pick ⌘⇧2 or ⌥⇧2 in the Bucket menu.
Open source
Trust is part of the product.
An app that sees your screen should say exactly what it does with it. Here it is, short.
No MCP server. No network. Just the clipboard.
Bucket hands your agent file paths through the clipboard, and the agent searches with a local command. No connections, no accounts, no analytics, no update checker. The OCR runs on the device.
Read every line.
MIT licensed, the code is short and has no third-party dependencies. Building from source is the real anchor of trust.
Screen Recording, nothing else.
No Accessibility, no Input Monitoring. Bucket does not read your keystrokes; the shortcuts use the system hotkey API.
Everything stays in ~/Bucket.
~/Bucket.Nothing is deleted on its own. The OCR text is stored in the clear, so whatever was on screen ends up there. Turn on FileVault.
Not notarized. Said plainly.
Signed ad-hoc, with no Apple Developer ID. The first launch needs Open Anyway, and the zip should match the hash in the release text.
Your settings stay yours.
Bucket does not edit macOS settings. You turn off the native ⌘⇧4 yourself.