Open source · Hosted or self-hosted · MCP for agents

Long videos in.
Vertical clips out.

Paste a link to a podcast, interview or talk. Reelsaw finds the moments that stand on their own, frames them for 9:16 around whoever is speaking, and burns in captions and a hook. Every pick comes with its reason.

15:27

A 15-minute podcast

clip from 13:49–14:09

Pick #2 of 8

Why this moment: Puts experience over theory, with a strong takeaway for viewers.

One real run

A 15-minute podcast in. 8 clips out.

No prompt, no settings, no scrubbing: this is what a stock Reelsaw kept, in its own order, with the reasons it gave. One clip plays at a time: point at another, or swipe on a phone.

15:27
The whole podcast, 16:9. 8 clips came out of it.
  1. How we launched a podcast! Captures the excitement of starting a podcast and the connection between the two hosts.

  2. We're actually living it! Puts experience over theory, with a strong takeaway for viewers.

  3. Where will we be in a year? Invites viewers to picture what comes next, which builds anticipation.

  4. Podcasting isn't as easy as it seems! Shows how much there is to it, for viewers who think podcasting is simple.

Source: We Started a Podcast with Zero Experience. Here's What 1 Year Taught Us by Truth and Time Podcast, CC BY. This run used cloud models; fully local works the same, slower.

See it

Nothing is a black box

Every automatic decision is shown where you are looking, with a one-click override.

Every pick scored, ranked and explained

The model reads the timed transcript, keeps the moments that stand on their own and ranks them head to head. Each clip arrives with its score, the rubric behind it and a one-line reason, so you can disagree with it in a click.

Project workspace: the source and its transcript on the left, rendered clips with score and layout on the right

Other sources, all CC BY: When is it Time to Quit? by Stand Show · Webinar: Building for Builders by Product School.

How it works

Four stages, each one a resumable job

Every stage checkpoints its output in Postgres. Restart the box mid-render and it picks up where it stopped.

  1. 01

    Ingest

    Paste a YouTube link or upload a file. yt-dlp fetches it, ffprobe reads it, and the source is stored once so every later step can resume from it.

  2. 02

    Transcribe and clean

    Audio is transcribed in parallel chunks with word timestamps. Names are fixed from your vocabulary, fillers and false starts are cut with jump cuts, and every cut is checked by listening to the seam.

  3. 03

    Pick the moments

    Map, pool, rank: the transcript is chunked at long pauses, every standalone moment is collected, scored on hook, value, flow and shareability, and the top picks are compared head-to-head with positions swapped to cancel judge bias.

  4. 04

    Reframe and render

    A 1.2 MB face detector on the CPU picks a layout per shot cut: follow the speaker, stack two people, slides over the webcam tile, or fit the whole frame. ffmpeg burns in captions, hook, intro and outro, fades and crossfades in one encode.

project.ingest
Ingest
project.transcribe
Transcribe
refine.* ×3
Refine
project.analyze
Pick moments
clip.render ×N
Render
One long video in.
Clips out.
Five queues, and each one saves its work.
✓ 5 clips readyfrom 58:12 of source
yt-dlpyoutube.com/watch?v=Qx7f…Founders Pod #42
ffprobe58:121920×108029.97 fpsAAC · 48 kHz
source stored once
▲Worker restarted mid-render. Job stays in the queue.
One video through the five queues: project.ingest → project.transcribe → refine.* ×3 → project.analyze → clip.render ×N. Every box is a pg-boss queue in Postgres; a worker restarts mid-render and resumes from its last save.

Features

Everything between the link and the upload

Built for people who publish every week and want their own pipeline, not a credit meter.

  • Fully local by default. Any vendor by choice.

    The installer sets up Ollama and a local Whisper, so nothing leaves the machine and no key is needed. Language, transcription and alignment models are each a provider:model string you can point at OpenAI, Anthropic, Google, Groq, Deepgram, AssemblyAI, Vercel AI Gateway, or any OpenAI-compatible server. No vendor is baked in; admins switch at runtime.

  • Research-backed clip selection

    No fixed clip count. Weak picks are dropped relative to the run, the best clip is always kept, and a different judge model can score and compare so a model never grades its own hooks. Click a score for the rubric breakdown.

  • Clean transcripts, kept honest

    Disfluencies are located by forced alignment so karaoke highlights land on the syllable. Each fix shows inline with accept and reject, long pauses can be shortened, and your decisions survive re-runs.

  • Smart 9:16 reframing

    Face, split, stack and fit layouts chosen per shot cut. Webcam-tile stacks, crowds shown whole, hand-set crops, and one clip's framing copied to the project. Override per project, per clip or per section.

  • Captions that read

    AI-picked key words coloured and enlarged, readability limits on line length and time on screen, karaoke word highlight, boxed or outlined. Drag caption and hook positions on a live preview with platform safe zones.

  • A real editor

    Live preview with the clip's framing and captions, a multi-lane timeline with zoom and snapping, frame-accurate trims, layout sections, layers, fades in and out, intro and outro crossfades, undo and redo. Keyboard first, shortcuts listed under ?.

  • Library, templates and export

    A library of intros, outros, logos, stickers, b-roll and music. Layer templates save a look once and apply it to any clip. Export all as a zip with paste-ready text, YouTube Data API and TikTok post bodies, a manifest and a CSV per clip.

  • Accounts, approval and quotas

    The first account is the admin. Open, approval or closed sign-up, per-user quotas, retention, captcha, email notices (including when clips are ready if you've stepped away) and usage per account. Admins switch models and providers at runtime without a restart.

  • Reliable, resumable, scalable

    Every stage is a pg-boss job in Postgres that checkpoints its output. A crashed worker resumes where it left off, a long video on a slow local model keeps the parts it finished (the project shows "13 of 43 parts, about 1 h left"), orphaned work is re-queued, and extra worker containers scale rendering. Local disk or any S3-compatible storage.

Agents and automation

MCP is the UI for agents

Every editing action in the app (projects, clips, transcript fixes, library, layer templates, settings and project defaults, billing links) exists as a tool on the built-in MCP server; only the admin pages for users, providers and models are app-only (Model Context Protocol, Streamable HTTP). Claude Code, Claude Desktop, Cursor or a script can create projects, wait for them, read timed transcripts, edit layers and framing, create and re-render clips, manage the library and layer templates, and export the results. The same tools are on every hosted plan, including Free. The operator CLI mirrors the same service code for people running the box.

  • create_project
  • wait_for_project
  • get_transcript
  • search_clips
  • get_clip
  • update_clip
  • create_clip
  • search_assets
  • apply_layer_template
  • export_project
  • get_media_links
  • create_webhook
Claude Code
claude mcp add --transport http reelsaw https://your-host/api/mcp --header "Authorization: Bearer rs_…"
{
  "mcpServers": {
    "reelsaw": {
      "type": "http",
      "url": "https://your-host/api/mcp",
      "headers": { "Authorization": "Bearer rs_…" }
    }
  }
}

Create an API key under Settings → API keys and MCP, or set MCP_OAUTH=true so claude.ai and Claude Desktop sign in through the browser.

How it compares

Hosted clippers charge by the minute. Self-hosted Reelsaw charges nothing.

Opus Clip, Klap, Vizard, Submagic, 2short.ai and quso.ai meter clips, minutes or credits and watermark or expire their free tiers. The other open-source clipper still needs Gemini for layout picking and silent videos. Reelsaw is AGPL-3.0 software on your own machine, fully local or with any provider you choose: no plan, no cap, no watermark. If you would rather not run a machine, hosted Reelsaw is $12 a month for 100 source minutes — the same list prices as OpenShorts Cloud, and cheaper than Opus Clip Starter at $15. Each comparison says what the other tool does better. Pricing checked 2026-10-24.

Self-host

Run it yourself: one command, no source needed

Needs Docker with Compose and about 10 GB of disk. The script writes ~/reelsaw with a compose file and generated secrets, pulls the published image and a language model sized to your RAM, then waits until the app answers.

macOS, Linux, WSL2

bash
curl -fsSL https://raw.githubusercontent.com/osbytes/reelsaw/main/deploy/install.sh | bash

On macOS install Ollama first. On Linux the script uses a host Ollama if present, otherwise runs it in a container.

Windows (PowerShell)

powershell
irm https://raw.githubusercontent.com/osbytes/reelsaw/main/deploy/install.ps1 | iex

With Docker Desktop running. If script execution is blocked, download install.ps1 and run it with -ExecutionPolicy Bypass.

From source, or with hosted providers

Clone the repository, copy the example environment, set an auth secret and at least one AI provider key, and bring the stack up. Open http://localhost:3000; the first account becomes the admin.

Deployment guide: providers, S3, workers, reverse proxies
git clone https://github.com/osbytes/reelsaw.git && cd reelsaw
cp .env.example .env
# set BETTER_AUTH_SECRET, APP_URL and at least one AI provider key
docker compose up -d --build