Loom

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me screen recording that replaces Loom — and know exactly why this is ALMOST.

The round trip is the product: hit stop and the link is already ready. What makes it hard is the two halves you are taking on. **A desktop recorder** that handles system audio, several monitors and the permission dialogs on both macOS and Windows is fiddly platform work. **A transcoding pipeline and a CDN** is where the recurring cost lives — this is a video hosting business now, and video is expensive.

STACK
- A browser recorder using getDisplayMedia and getUserMedia, which avoids the desktop problem entirely at the cost of system audio on some platforms — start here
- A desktop application later, if system audio and multi-monitor matter enough
- Node 20+ with Fastify, SQLite through better-sqlite3
- ffmpeg on a worker for transcoding
- Object storage with a CDN in front
- Caddy in front

THE ROUND TRIP, WHICH IS THE WHOLE POINT
- Allocate the video's identifier before recording starts, so the link exists the moment you stop
- Upload in chunks as they are recorded, not after. By the time somebody presses stop, most of the file is already there and the link is ready in a second or two
- A crash loses seconds, not the whole recording, because chunks are already stored
- The link is copied to the clipboard automatically on stop. That single behaviour is why people use this instead of attaching a file
- Playback begins from the original recording while the transcode runs, with the better renditions swapped in when ready. Never make somebody wait for a transcode to share

THE DATA MODEL
- videos: id, slug, owner_id, title, description, duration_ms, status, source_path, poster_path, bytes, width, height, created_at, expires_at
- chunks: id, video_id, index, path, bytes, received_at
- renditions: id, video_id, height, bitrate, path, bytes, status
- transcripts: id, video_id, language, vtt_path, is_edited
- shares: id, video_id, token, password_hash, expires_at, allow_download, allow_comments, revoked_at
- views: id, video_id, viewer_hash, started_at, watched_ms, max_position_ms, completed, bot
- comments: id, video_id, at_ms, author, body, created_at — anchored to a moment, which is what makes feedback on a recording useful
- reactions and emoji at a timestamp, if you want them

RECORDING
- Screen, camera, or both, with the camera as a circle in a corner
- Microphone with a live level meter, and a warning if it has been silent for several seconds. A recording with no audio is discovered at the worst possible moment
- System audio where the platform allows it, and an honest note where it does not — this is the main thing the browser cannot always do
- A countdown, a visible indicator while recording, pause and resume, and a keyboard shortcut to stop
- A cursor highlight and click ripples drawn from a separately recorded pointer track, which costs almost nothing and makes a demo far clearer

TRANSCODING AND COST
- A ladder of renditions packaged as HLS so a player adapts to the connection
- Queue it with a concurrency cap. This is CPU-hours per video and it is the bill
- Storage and egress are the other bill. Measure bytes per minute early, set a retention policy, and put the numbers in the README. This is the honest reason the subscription exists

WATCHING
- A share page that is fast and plain: the player, the title, the transcript, and a comment box
- Viewer counts and how far people watched. Where people stop is the only analytic that matters — it tells you which minute to cut
- Speed control, captions on by default, and a plain video element rather than a framework
- Optional password, expiry and download permission per share

TRANSCRIPTS
- A local speech model producing WebVTT, editable, searchable, and used as the page's text so a recording is findable later
- Never send the audio to a third party without saying so

OPERATIONS
- .env: DATABASE_PATH, STORAGE_PATH, BASE_URL, SESSION_SECRET, CDN_URL, RENDER_CONCURRENCY, RETENTION_DAYS
- Migrations on boot, each once; disk watchdog and a working sweeper
- Nightly backup of the database; source videos per their own policy
- Health endpoint reporting queue depth and free space

WHAT MATTERS MOST
Chunked upload during recording, so the link is ready at stop. Everything else here is a video site; that one behaviour is why the product replaced attaching a file to an email.