OpenScreen: Record, Auto-Edit & Export PR-Ready Demos with Local AI & GPU Encoding

OpenScreen is a desktop application (Windows, macOS, Linux) that records the screen and then edits the recording into a polished product-demo video. It positions itself as the open-source counterpart to Screen Studio, which defined the category of "record-then-auto-beautify" tools but charges a subscription. OpenScreen is MIT-licensed, has 4.3k GitHub stars, and its latest release is v2.0.0.

Core Architecture: A Custom Rendering & Encoding Pipeline

The project's differentiating value lies in its export pipeline. Instead of shelling out to ffmpeg, it implements a native compositor in Rust that performs GPU encoding on each platform: Metal on macOS, D3D11 on Windows, and Vulkan on Linux, with automatic CPU fallback. This pipeline handles the final render of all edits — zooms, backgrounds, motion blur, 3D device frames — at high speed without watermarks.

Feature Layers

Recording Layer

Full-screen or window capture

Microphone + system audio capture

Webcam picture-in-picture with 8 snap positions, adjustable corner radius, mirroring, and shape options

Real cursor capture (shape, type, clicks) with post-recording path smoothing

Built-in teleprompter (Notes teleprompter mode)

Editing Layer

Auto-zoom (follows cursor) and manual zoom with depth, duration, easing, and pixel-level position control

Local Whisper transcription (~260 MB model, downloaded once on first use of a voiced segment); fully offline, no upload. Transcript is editable; deleting a sentence from the transcript cuts it from the video.

AI editing assistant: chat interface describes changes (cut, zoom, speed, annotations, framing) and applies them directly to the timeline. Supports Claude, OpenAI, Gemini, Mistral, OpenRouter, MiniMax, or any OpenAI-compatible endpoint; disabled by default.

Backgrounds: wallpaper, solid color, gradient, custom image

3D effects: cursor-following camera, modeled cursor, click tilt, depth of field, 3D device frames (laptop, phone, browser, monitor)

Webcam background processing: remove, blur, replace — runs locally via ONNX on CPU

Motion blur, crop, trim, per-segment speed change

Text, arrow, image annotations with text animation presets

Timeline snap guides + audio waveform

Export Layer

MP4 or GIF, multiple aspect ratios and resolutions

GPU render and encode with automatic CPU fallback

Automation Layer (Headless CLI + MCP Server)

Headless CLI with three commands: record, export, info; scriptable and CI-friendly; --json outputs NDJSON.

MCP server integrates OpenScreen into coding agents like Claude Code.

UI available in 15 languages including Simplified and Traditional Chinese.

End-to-End Scenario: AI Records a Demo and Attaches It to a PR

The most distinctive workflow connects OpenScreen to an coding agent. In Claude Code, you install an OpenScreen skill (with consent), which sets up the MCP server. You then start a new conversation and say "continue the OpenScreen demo" — a step the agent cannot do alone. After that, you can simply say "Add a video of this change to the PR." The agent then: plans shots → records the app → uses OpenScreen's own editing tools to cut the clip → exports MP4 → frame-checks the result → attaches it to the pull request, where reviewers watch the video directly in the PR. For release demos, it can draft an X (Twitter) post and wait for your "send" confirmation. This pipeline works because editing is exposed as a programmable API: the three CLI commands plus read/write of the .openscreen project JSON make "record-to-finished-asset" fully scriptable.

openscreen record --duration 20 --project demo.openscreen --json
openscreen export demo.openscreen -o demo.mp4 --json

Installation

Windows

Preferred: Microsoft Store via winget install --source msstore OpenScreen — Microsoft-signed, no SmartScreen warning, auto-updates. Alternative: download .exe from Releases (unsigned, SmartScreen blocks; ARM64 devices must use ARM64 build, x64 under emulation records poorly).

macOS

Download .dmg from Releases, drag to Applications. Since v1.9.0: Developer ID signed + Apple notarized, Gatekeeper allows. First launch requires granting Screen Recording and Accessibility permissions in System Settings; both are mandatory.

Linux

Four package formats per distribution:

sudo apt install ./Openscreen-Linux-*.deb  # Debian/Ubuntu
sudo dnf install ./Openscreen-Linux-*.rpm  # Fedora/RHEL
sudo pacman -U Openscreen-Linux-*.pacman  # Arch
./Openscreen-Linux-*.AppImage  # Any distro

NixOS/Nix users can run nix run github:getopenscreen/openscreen or use the built-in flake module.

Hardware Requirements

8 GB RAM minimum, 16 GB recommended. Windows 10 1903+, Intel 8th gen / Ryzen 2000 series or newer. macOS 13 Ventura+ (requires ScreenCaptureKit). Linux needs xdg-desktop-portal + PipeWire.

Current State & Known Issues

The project explicitly states it is in active development with rough edges: the .openscreen project format and CLI may see breaking changes. The public roadmap lists concrete issues: Windows stop-recording occasional hang leaving empty file; macOS long recordings may truncate; Linux on niri (Wayland) cannot record at all; AppImage PipeWire capture reports "error alloc buffers"; some older AMD GPUs on Linux produce invalid H.264; Windows game recording stutters on certain machines. Most have been fixed in recent versions; maintainers await user confirmation.

Who It's For

People creating product demos, documentation, PR demo videos — its core use case; complete record-to-finish pipeline, only free open-source implementation with this feature set.

Users who want Screen Studio parity without a subscription — visual effects (zoom, backgrounds, motion blur, 3D device frames) are all present.

Teams wanting agents to produce demo videos autonomously — MCP server + CLI + project JSON is a unique automation chain.

Privacy-conscious users — transcription, subtitles, webcam background removal run locally; nothing leaves the machine unless you enable an LLM.

Not for: users needing zero-rough-edge, out-of-the-box stable commercial tools. Rapid iteration means occasional crashes and format changes are documented trade-offs. Linux users especially should expect Wayland limitations (e.g., click effects require evdev access via udev rules or input group).

Project History: Community Fork

Original author Siddharth Vaddem archived the repo after v1.5.0. The current getopenscreen/openscreen is a community continuation with his blessing — same name, same MIT license, development moved to a new repo. Contributor stats show maintainer EtienneLescot with 2,157 commits vs. original author's 504. This explains the rapid 4.3k stars and steady release cadence in half a year, and why the project emphasizes "official/trusted links" to avoid phishing sites that often target popular community forks. For an open-source project, willing community succession is a stronger signal than star count alone.

GitHub: github.com/getopenscreen/openscreen