Your coding agent is blind to your screen — so you describe the bug in words and hope. DeckShot ends that: hit Ctrl+Shift+S, mark up what's wrong, and on-device OCR reads every word of text. A ready-to-paste prompt lands in your clipboard — the annotated shot's path, your marks, the OCR — and your agent opens the actual pixels. Claude Code reads images by path; DeckShot is built around exactly that.
DeckShot is a small, fast Mac app. One hotkey captures, a clean annotator marks up, and everything your agent needs — image, marks, text — is packaged into one prompt before you switch windows.
Press Ctrl+Shift+S (yours to change) and grab a region, window or the whole screen. Shots save to ~/Pictures/DeckShot with clean timestamped names.
Boxes, arrows and labels — point at exactly what's wrong. Marks save in a sidecar file, and a flattened annotated PNG is exported for the agent to read.
A structured prompt is already in your clipboard: the absolute path, your annotations, the OCR text. Paste it — the agent opens the pixels and gets to work.
A screenshot in your Downloads folder does your agent no good. DeckShot's whole job is the handoff: the moment you finish annotating, your clipboard holds a prompt with the image path (agents like Claude Code read images by absolute path), every mark you drew with its coordinates and label, and the on-device OCR of all visible text — so the agent sees the picture AND can quote the error message in it.
DeckShot ships an MCP server, so the pipe runs both ways. Mid-session, your agent can call deckshot_latest to pull in the shot you just took — or deckshot_capture to request a fresh one. You capture (and approve), the agent sees. "Show me what the dialog looks like now" stops being a copy-paste chore.
Same local-first philosophy as the rest of the deck: no account, no telemetry, no cloud. Your screenshots and OCR never leave your Mac — the only thing that moves is the prompt you paste.
Ctrl+Shift+S from anywhere — region, window or full screen. Configurable. Shots land in ~/Pictures/DeckShot with timestamped names your agent can trust.
Boxes, arrows, labels. Saved as a sidecar file next to the shot — re-editable later — plus a flattened annotated PNG so the agent literally sees your marks.
Every word of visible text extracted locally — error messages, console output, UI labels — cached beside the shot. Nothing is uploaded, ever.
After each capture your clipboard holds one structured prompt: absolute path, annotation summary, OCR text. Paste into Claude Code and it reads the image itself.
deckshot_latest and deckshot_capture let the agent pull shots into its own context — including asking YOU for a fresh capture mid-session.
Signed and notarized for macOS, with in-app auto-updates. Free updates within v1.x, delivered through your personal download link.
No subscription, no credits, no AI bill from us. Pay once, screenshot forever.
Coding agents like Claude Code can read image files from disk by absolute path — their file-read tool returns the actual pixels to the model. DeckShot leans on exactly that: it saves the shot locally and puts a prompt in your clipboard that starts with "Read /path/to/shot.png". Paste it, and the agent opens the image itself — plus your annotations and the OCR text, inline in the same prompt.
Two tools: deckshot_latest returns your most recent capture (path, annotation summary, OCR text) so the agent can pull it into context without you pasting anything. deckshot_capture lets the agent request a fresh screenshot mid-session — you take it, and it lands in the agent's context. One line of config for Claude Code; the app shows it ready to copy.
Yes. Text recognition runs locally on your Mac — no cloud calls, no uploads, no account. The recognized text is cached in a small file next to the screenshot so repeat pastes are instant. Your screen contents never leave your machine.
Everything lands in ~/Pictures/DeckShot (configurable): the original PNG, a sidecar file with your annotations (re-editable), a flattened annotated PNG for the agent, and the OCR text cache. Plain files with predictable names — easy to gitignore, easy to clean up, easy for an agent to find.
Claude Code is the first-class target — its file Read tool handles images natively and the MCP server plugs straight in. But the clipboard prompt is plain text + a file path, so any agent that can read files from your disk gets the same benefit, and any MCP client can use the tools.
It can — capture, annotate and save works standalone, with a proper hotkey and clean file naming. But its reason to exist is the pipe: if you never paste a screenshot into an agent, plain macOS screenshots are free and fine. DeckShot is for people who show their agent the screen many times a day.
No trial — instead there's a 7-day, no-questions refund. Buy it, wire it into your real workflow, and if it's not for you, email info@deckspace.dev from the address you paid with and a real person refunds you promptly.
Two Macs per license, tied to your key. Switching machines is self-serve — deactivate the old Mac in the app and activate the new one. Teams: one license per person.
Capture it, mark it up, and paste the whole thing — pixels, marks and text — straight into your agent's context. The bug report writes itself.
DeckSpace