Repository navigation
Open Edit v1.3.1
Open Edit is an AI-native video editor driven over the Model Context Protocol (MCP): agents inspect projects, apply edits, render, and QC through a local MCP stdio server. Everything is local-first — your media and edit history stay on your machine, with no cloud accounts or uploads. v1.3.1 makes the one-command installers provision the entire render stack, not just the Python MCP server, so a fresh download can actually render motion-graphics video out of the box.
What's new
- Installers provision the full render stack.
install.sh(Linux/macOS) andinstall.ps1(Windows) now go beyond clone + venv + pip: they detect (and, where possible, auto-install user-local — never with sudo) every runtime the render pipeline needs, then verify it. - Node.js + HyperFrames: the installers detect Node ≥ 22 (the pinned hyperframes
0.7.65engine requirement; auto-installing a user-local copy when missing), runnpm install, and verify the pinned motion-graphics engine lands atnode_modules/.bin/hyperframesand runs--version. - ffmpeg: detected on PATH; when absent, Windows auto-installs via winget (
Gyan.FFmpeg) and Linux prints exact package-manager commands — never a silent gap. - melt (MLT): checked on PATH on Linux (with apt/dnf/pacman/brew hints when missing). On Windows there is currently no packaged melt (no winget/chocolatey package; official builds are source-only), so the installer detects it and, if missing, ends with an explicit warning covering WSL (
apt install melt) — overlay/motion-graphics-only renders work without melt. - Chrome/Chromium for puppeteer-core: common browser paths are probed; if none is found the installer prints how to fetch one for HyperFrames, e.g.
npx @puppeteer/browsers install chrome. - Runtime readiness summary: every install ends with a clear status table naming which of ffmpeg / melt / node / hyperframes / chrome are ready vs. which need a manual step — no more post-install guessing. The
open-edit-mcp --helpverification still runs and must pass. - GPU (NVENC) encoder support: final exports can encode on the GPU. With
OPEN_EDIT_RENDER_BACKEND=gpu(the default), the renderer probes NVENC → AMF → QSV → VAAPI and falls back tolibx264when no GPU encoder is available; setOPEN_EDIT_RENDER_BACKEND=cputo always use software encoding. NVENC supports h264/hevc/av1 quality tiers and overrides (crf → cq, bitrate, preset, scale, codec). - Docs updates:
INSTALL.mdrender-stack prerequisites,docs/MCP.mdrender pipeline (native HyperFrames graphics → melt → ffmpeg composite/encode), and the agent install prompt indocs/agent-install.mdnow reflect the full render stack and GPU encode path.
Proof point
The release's end-to-end validation: a 60-second HyperFrames logo intro rendered through the pipeline at 1080p30 with GPU (NVENC) encode — HyperFrames overlay capture, melt base track, ffmpeg composite and GPU encode, QC gate passed. That's the same path a fresh install.sh / install.ps1 download now provisions automatically.
Install
- Linux / macOS:
bash install.sh - Windows:
powershell -ExecutionPolicy Bypass -File install.ps1 - One-command downloads:
curl -fsSL https://raw.githubusercontent.com/AH64-dll/OpenEdit/v1.3.1/install.sh | bash·irm https://raw.githubusercontent.com/AH64-dll/OpenEdit/v1.3.1/install.ps1 | iex - Full setup instructions: see
INSTALL.mdand the agent prompts indocs/agent-install.md/docs/agent-configure.md.