Install

cargo install cloche
# or from git
cargo install --git https://github.com/escoffier-labs/cloche

What it does

Job What you get
Capture Any window or region Desktop frames without a browser extension stack
Present Polished share cards README- and social-ready framing, not a raw dump
Emit Stable JSON stdout Shells, hotkeys, and agents can consume the result
Serve Optional MCP Expose capture to coding agents when you want it

Quick Start

Capture the active app or window as a Shot:

cloche capture --target active --out-dir /tmp/cloche-shot-$(date +%s) --format json

Style an existing screenshot into a presentation card without recapturing:

cloche polish /tmp/diff.png --palette violet-haze --format json

By default the card sits on a procedural deep-space scene. Pick one instead of letting the seed choose:

cloche polish /tmp/diff.png --scene jwst --style-seed 12345

See Space backdrops for the full list of looks.

Render an existing short recording through the experimental Remotion reel engine:

cd remotion && npm install && cd ..
cloche reels render \
  --input raw.mp4 \
  --out demo.mp4 \
  --cues cues.json \
  --title "Create a project"

The cue file follows the AppReels timeline shape:

{
  "titleCard": { "text": "Create a project", "ms": 900 },
  "captions": [{ "startMs": 1000, "endMs": 2600, "text": "Open the project menu" }],
  "zooms": [{ "startMs": 1200, "endMs": 2400, "scale": 1.35, "x": 0.3, "y": 0.4 }],
  "outroCard": { "text": "Done", "ms": 700 }
}

Zooms ease in and out on their own; x/y (0 to 1 across the footage, optional) pick the focus point, defaulting to center.

Preview the latest capture:

cloche preview

Create a self-contained HTML gallery of recent Shots:

cloche gallery --root /tmp --html /tmp/cloche.html --title "My Shots" --open

Generate a Codex turn/start payload from a Shot:

cloche codex-payload --thread-id "$THREAD_ID" /tmp/cloche-shot-123

Space backdrops

Every card is framed on a procedural deep-space scene generated from the style seed. Nothing is a stock image: the nebulae, starfields, and objects are drawn from noise at render time, so each seed is a different sky and the same --style-seed reproduces one exactly.

Eight palettes are color-sampled from astrophotography of real objects, and a scene layers gas, dust lanes, starfields, and one or more focal objects on top:

  • Nebulae: domain-warped clouds with bright ionization fronts and dark dust lanes. Palettes: orion-emission, carina-hubble, pleiades-reflection, rho-ophiuchi, milkyway-core, andromeda-haze, horsehead-flame, lagoon-trifid.
  • Telescope looks: roughly half of scenes take the JWST signature (6-point diffraction spikes, clumpy dust, blue-cored galaxies). Others echo ALMA protoplanetary discs, SDO extreme-UV suns with coronal loops, and Chandra supernova-remnant shells.
  • Objects: spiral and edge-on dust-lane galaxies, ring and bipolar planetary nebulae, open clusters, gravitational-lensing arcs, and rare ultra-deep-field and Planck-CMB frames.

Random styling picks a space palette. The five original gradient palettes (violet-haze, ember-glow, aurora-teal, rose-noir, midnight-sky) are still available by name with --palette.

Pin a look with --scene instead of rolling the seed:

cloche polish shot.png --scene alma
cloche polish shot.png --scene cmb --style-seed 42

Scenes: nebula, jwst, hubble, galaxy, alma, ring, butterfly, edge-on, sun, sdo, cluster, deep-field, lensing, veil, remnant, cmb. --scene applies only to space palettes. On a gradient palette it is a no-op warning. The same flag is available on the MCP polish tool.

Command Reference

cloche doctor --format json
cloche list-windows --format json
cloche capture --target active --presentation both --out-dir /tmp/cloche-shot --format json
cloche capture --target active --style-seed 12345 --out-dir /tmp/cloche-shot --format json
cloche capture --target screen --out-dir /tmp/cloche-shot --format json
cloche capture --target window --title Firefox --out-dir /tmp/cloche-shot --format json
cloche capture --target region --presentation both --clipboard --out-dir /tmp/cloche-shot --format json
cloche polish /tmp/diff.png --format json
cloche polish /tmp/diff.png --out /tmp/diff-card.png --palette ember-glow --style-seed 12345
cloche gallery --limit 10
cloche gallery --root /tmp --html /tmp/cloche.html --title "My Shots" --open
cloche latest
cloche preview
cloche open /tmp/cloche-shot
cloche schema
cloche schema --for polish
cloche codex-payload --thread-id THREAD_ID /tmp/cloche-shot
cloche mcp
cloche setup
cloche setup --print
cloche setup hotkey
cloche setup agent --client claude-code
cloche setup verify --format json

The old appshots command remains as an alias for the same code path.

Modes

Shots are available now. A Shot is a still capture with raw and presentation images, metadata, and optional extracted text.

Reels are planned next. A Reel will be a short desktop recording with the same Cloche presentation system, cursor emphasis, captions, and stable metadata. The existing Appreels prototype is the starting point for this mode.

GIF export is planned after Reels. GIFs will be generated from finished Reels as a delivery format, not recorded as the primary source format.

Why Cloche Exists

OpenAI documents Appshots as a macOS app feature for Codex. The Codex repository can already resume threads that contain local images through app-server v2 turn/start input:

{ "type": "localImage", "path": "/absolute/path/to/shot.png", "detail": "high" }

Linux and Windows users still need a reliable way to create those captures from a normal CLI. Cloche fills that capture side while staying independent of any one agent stack. Use it with Codex, OpenClaw, Claude Code, Hermes, a local MCP client, or a plain shell script.

Reference: https://developers.openai.com/codex/appshots

Output Files

Each successful Shot directory contains:

  • shot.png, the raw captured image.
  • shot-card.png, a presentation image: the screenshot with rounded corners and a soft shadow on a full-bleed gradient backdrop, fully opaque so it survives JPEG and pasting anywhere.
  • metadata.json, the same JSON object printed to stdout.
  • text.txt, optional best-effort accessible text from the focused app.

Capture exits with 0 only when a raw image was written. Text extraction and presentation-image failures are warnings because accessibility support and desktop compositing vary by toolkit, app, desktop environment, and OS. --target screen exists as a fallback and debugging mode. --target active is the default.

Use --presentation raw, --presentation card, or --presentation both to control output image generation. Use --style-seed <number> to reproduce a randomized card style exactly.

--target region opens an interactive selector (Flameshot when available, ImageMagick import drag-select on X11): drag a rectangle and the shot is taken the moment you release. Add --clipboard to copy the finished card straight to the clipboard (wl-copy on Wayland, xclip on X11). Region capture needs a human at the desk; it is not for headless agents. Not yet supported on Windows: use Win+Shift+S, save the file, then cloche polish <file>.

Hotkey Workflow

Bind one key to get a share-ready card on your clipboard: press it, drag a region, paste the polished card anywhere. The capture, polish, and clipboard copy all happen in cloche; only the key binding is set up per desktop.

Cloche hotkey workflow

Generated from docs/assets/workflows/hotkey.json with plating workflow.

The fastest path is one command:

cloche setup

It installs cloche-grab, binds it to Print on GNOME (and prints the exact steps on KDE/sway/i3), registers the MCP server with any agent it detects, then verifies that capture, the hotkey, and the MCP server actually work. Run cloche setup --print to preview every change first, or cloche setup verify any time to re-check. The manual steps below are what cloche setup automates, and the fallback for unsupported desktops.

The repo ships scripts/cloche-grab.sh, which wraps the capture and adds a desktop notification. Install it and bind it:

# 1. Put the script on your PATH (or point the binding at it in place).
install -Dm755 scripts/cloche-grab.sh ~/.local/bin/cloche-grab

# 2. Confirm it works (it opens the region selector):
cloche-grab

Then bind cloche-grab to a key:

  • GNOME: Settings -> Keyboard -> View and Customize Shortcuts -> Custom Shortcuts -> +. Name it "Cloche Grab", command cloche-grab, and set the shortcut (e.g. Print). To move the native screenshot UI off Print first: gsettings set org.gnome.shell.keybindings show-screenshot-ui "['<Shift>Print']".
  • KDE: System Settings -> Shortcuts -> Custom Shortcuts -> Edit -> New -> Global Shortcut -> Command/URL, command cloche-grab, then assign a key.
  • Anything else (sway, i3, ...): bind a key to cloche-grab in your WM config.

Prefer no script? Bind this one-liner directly instead:

cloche capture --target region --presentation both --clipboard --out-dir ~/Pictures/ClocheShots/$(date +%s)

cloche polish writes a single card PNG instead of a Shot directory: <input>-card.png next to the input by default, or the --out <path> you pass (it must end in .png). Its stdout JSON reports input, card, and presentationStyle.

Agent Use

Any shell-capable agent can call:

cloche capture --target active --out-dir /tmp/cloche-shot-$(date +%s) --format json

Then parse image.path from stdout or read the generated metadata.json.

Codex app-server clients can turn a capture into a ready turn/start payload:

cloche codex-payload --thread-id "$THREAD_ID" /tmp/cloche-shot-123

Other agents should treat Cloche as a normal subprocess tool. The core command has no MCP dependency, desktop-app dependency, or agent-specific runtime dependency.

MCP Server

cloche mcp runs a minimal stdio MCP server for clients that prefer the Model Context Protocol over direct subprocess calls. It speaks newline-delimited JSON-RPC 2.0 on stdin/stdout and exposes capture, polish, list_windows, doctor, latest, and gallery as tools. Each tool call shells out to the same binary, so the JSON contract is identical to the CLI.

cloche setup agent registers this server with Claude Code, OpenClaw, and Codex CLI automatically (backing up any config it edits, and skipping clients already configured). The manual config below is for other clients or if you prefer to wire it yourself.

Register it like any stdio MCP server:

{
  "mcpServers": {
    "cloche": { "command": "cloche", "args": ["mcp"] }
  }
}

Compatibility config:

{
  "mcpServers": {
    "appshots": { "command": "appshots", "args": ["mcp"] }
  }
}

Linux Backend Notes

  • X11 active/window capture uses xdotool/wmctrl for window metadata and ImageMagick import for PNG capture.
  • Wayland wlroots screen capture uses grim.
  • GNOME/KDE Wayland may block silent active-window capture by design. Use --target screen or run cloche doctor --format json for diagnostics.
  • Text extraction is best-effort through AT-SPI using Python GI when available.

If you are invoking Cloche from SSH, a TTY, or an agent process that did not inherit the desktop environment, Cloche will try to discover the live desktop variables from desktop processes. On GNOME X11 they usually look like:

export DISPLAY=:1
export XAUTHORITY=/run/user/$(id -u)/gdm/Xauthority
export DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/$(id -u)/bus
export XDG_SESSION_TYPE=x11

You can discover the active values from a desktop process:

tr '\0' '\n' </proc/$(pgrep -u "$(id -u)" -n gnome-shell)/environ | grep -E '^(DISPLAY|XAUTHORITY|DBUS_SESSION_BUS_ADDRESS|XDG_SESSION_TYPE)='

Windows Backend Notes

  • Active/window capture uses Win32 foreground-window and top-level-window metadata, then captures the target window with PrintWindow so covered windows are not polluted by whatever is on top. It falls back to .NET CopyFromScreen if PrintWindow is unavailable for that window.
  • Screen capture uses the Windows virtual screen.
  • Text extraction is best-effort through UI Automation.
  • Capture must run in a logged-in interactive desktop session. Plain OpenSSH sessions can build and run doctor, but Windows blocks screen capture from the non-interactive SSH service session.

Gallery HTML Export

cloche gallery --html <path> writes a single self-contained HTML file with each capture's image embedded inline, so the result can be shared without any companion files. Combine with --root, --limit, --title, and --open. The JSON output gains an htmlPath field pointing at the written file.

Release Packaging

Build a local release archive:

bash scripts/package-release.sh

On Windows:

powershell -ExecutionPolicy Bypass -File scripts/package-release.ps1

Archives are written under dist/. Tagged GitHub releases are packaged by .github/workflows/release.yml for Linux and Windows.

Why not other screenshot tools?

  • Flameshot, Spectacle, GNOME Screenshot, Greenshot, ShareX are excellent interactive GUI tools, built for a human clicking and annotating. They do not present a stable command surface for scripts, they do not emit machine-readable JSON, and they do not run headless from an agent process. Cloche is CLI-first and JSON-first; it uses tools like Flameshot for the region-select step but owns the polish, metadata, and contract.
  • grim / scrot / ImageMagick import / maim capture pixels and stop there. You still hand-roll the framing, the gradient, the metadata, and the JSON. Cloche wraps the same low-level capture and does the rest in one command.
  • carbon-now, silicon, ray.so make beautiful cards out of source code or arbitrary images, not live windows. Cloche captures the actual app and then frames it, and cloche polish <image> covers the "I already have a screenshot" case.
  • The macOS Appshots app for Codex is the inspiration, but it is macOS-only and tied to one agent stack. Cloche fills the same capture role on Linux and Windows while staying agent-neutral: use it with Codex, OpenClaw, Claude Code, Hermes, any MCP client, or a plain shell script.

What Cloche is not

Cloche is a local capture tool, not a service or an annotation suite.

It does not:

  • run a background daemon, tray app, or scheduler
  • upload, sync, or phone home with your captures
  • annotate, blur, or redact (it frames what is on screen; review before you share)
  • record audio or do OCR (text extraction is best-effort via the OS accessibility layer only)
  • replace your editor or your sharing host; it writes local files and a JSON receipt, and you take it from there

Roadmap

See ROADMAP.md.