humblebanana/open-record-replay
Open-source macOS record-and-replay workflow recorder for computer use agents
Open Record/Replay is a local-first macOS tool for teaching Computer Use agents by demonstration. Record a desktop workflow once — the tool captures it as structured session.json and events.jsonl artifacts — validate the recording quality, and package the evidence as input to an agent's skill-creation flow. Bridges the gap between 'hard to describe as a prompt' workflows (file pickers, multi-app sequences, drag-and-drop) and agent skills that can replay them reliably.
Install
git clone https://github.com/humblebanana/open-record-replay.git && cd open-record-replay && npm install && npm run build:nativeRequires macOS, Node.js 18+, Swift toolchain / Xcode Command Line Tools, Accessibility permission, and Input Monitoring permission. After build, run: node bin/orr.js permissions check
Compatibility
macOS only (alpha). Requires Accessibility + Input Monitoring permissions. Node.js 18+ and Swift/Xcode Command Line Tools required for the native recorder build.
Details
- Repo: humblebanana/open-record-replay
- Category: Web UI Enhancements
- Stars: 139
- Last push: 2026-06-22
- First seen: 2026-06-21
Recent updates
Alpha status — actively developed. Captures action-level evidence for workflows involving desktop UI, browser navigation, file operations, and multi-app sequences. Recording data contract documented in docs/recording-data-contract.md.
FAQ
- How do I install Open Record/Replay?
- Clone the repo, run npm install, then npm run build:native to compile the Swift-based native recorder. Before first use, run: node bin/orr.js permissions check — and grant Accessibility + Input Monitoring permissions in macOS System Settings.
- What can I record with Open Record/Replay?
- Any macOS workflow: sending files through a desktop chat app, creating and sharing documents, browser search-and-navigate sequences, moving between apps, or any repetitive UI process not covered by a clean API.
- What does the recorded output look like?
- Each recording produces a session.json (metadata and structure) and events.jsonl (action-level event stream). These artifacts are validated for quality, then packaged as skill input for a Computer Use agent.
Alternatives
drewnekota/cetus · pulseaiclub/phi