OpenMontage: one prompt to a narrated video, for zero pounds
"Turns Claude Code into a full video production pipeline." I ran the free path and kept notes on where the money starts.
- The free path is real: a narrated, captioned, animated explainer for £0, no accounts.
- The cinematic videos in the posts need paid generation (fal.ai, Kling, Veo, ElevenLabs). I did not try those.
- Its tools work as a plain Python library, which makes it a good building block.
What the posts claim
Five TikTok posts, all about OpenMontage: "25,000 stars, turns Claude Code into a full video production pipeline". One shows the cinematic-trailer version, with a fal.ai account topped up with $10 and "$12 for the Claude session". The star count was out of date by the time I looked: about 35,000.
What it is
Not an app. A repo you open in Claude Code (or Cursor, Codex and so on). It holds about 50 Python tools (speech, image and video providers, stock-footage search, FFmpeg and Remotion composition, subtitles, QA), 12 YAML pipelines and hundreds of Markdown skills that tell the coding agent which tool to call. The agent is the orchestrator. There is no server. Licence: AGPL-3.0.
What I ran
A 2.9 GB Docker image with Node 22, a Python 3.11 environment, Piper text-to-speech with the en_US-lessac-medium voice, and the Remotion composer. OpenMontage was cloned inside the image at one pinned commit. I read the install scripts first. No API keys were given to it.
The prompt was one of its own "zero keys needed" examples: Why the sky is blue, 45 seconds.
- A free model, reached through my homelab's model gateway, wrote the script and scene plan in about 90 seconds.
- A small script I wrote cleaned the plan and timed the scenes to the voice.
- OpenMontage's Piper wrapper narrated each scene in about 20 seconds.
- Its Remotion renderer produced 1080p in 5 to 6 minutes on 6 CPUs.
- FFmpeg re-encoded it to 720p H.264.
What happened
A 42-second video, 1.2 MB, with voice and word captions. A free model wrote a correct, well-paced script on the first try. The Remotion scene library (title, stat, charts, comparison, callout, terminal and screenshot scenes) looks professional out of the box. Its own zero-key demo, rendered as a baseline, is 23 seconds and silent by design.
Limits
- Piper's voice is clear but flat next to ElevenLabs.
- Captions are estimated, because Piper gives no word timings. They drift by a word or two. Its Whisper transcriber would fix that at extra CPU cost.
- No music on the free path without a free Pixabay or Freesound key.
- The documentary-montage pipeline needs CLIP indexing of downloaded footage. Too heavy for my laptop, not tried.
- The comparison card colours its right side green whatever the label says.
- Render time is about 8 seconds per second of video on 6 cores.
- AGPL: fine for personal use. Anything offered as a service has to publish its source.
Verdict
Keep it as a building block. I called PiperTTS and the Remotion render straight from Python, with no agent session, and that is the part I will reuse. For the posts' Pixar-style shorts you need paid clips from fal.ai, Kling or Veo, which the posts price at $1.33 to $12 or more per video plus the Claude session. I spent nothing, so I cannot tell you how good those are.
Draft status: one trial, one prompt. If you think I got something wrong, tell me and I will correct the post and say so at the top.