Documentary Montage is here — cinematic stock-footage edits from a single prompt, for $0

[calesthio]

Documentary Montage is here. Zero dollars. Real footage. No AI slop.

We just shipped a pipeline that takes a single natural-language prompt and cuts a real documentary-style montage from actual stock footage — Pexels, Archive.org (Prelinger), NASA — with rhythm, juxtaposition, music sync, and editorial polish.

No generative video models. No per-minute API bills. No API keys for the video side at all. Just you, your coding assistant, and a CLIP-indexed corpus of free footage.

Total cost of the example below: $0.00.

Under the hood

• Retrieval-first, not generation-first. We build a corpus, index it with CLIP ViT-B/32, and semantically search for the exact shot each slot needs.
• Pexels + Archive.org (Prelinger) + NASA. All free, all properly licensed, zero API keys on the video side.
• Agent-driven editorial intelligence. The director skills (scene-director, asset-director, edit-director, compose-director) teach your coding assistant how to think like a documentary editor — rhythm tables, hero slots, L-cuts, adjacent-diversity checks, silence windows, register continuity across mixed-era footage.
• The intelligence is in the skill files, not the Python. Your coding assistant IS the orchestrator. No LLM_API_KEY inside OpenMontage itself.

Try it with these prompts

Drop any of these into Claude Code (or any Claude-Agent-SDK-compatible assistant) in an OpenMontage workspace:

"Make a 90-second documentary-style montage about what a city feels like at 4am. No narration. Elegiac tone. Use real stock footage only."

"Create a 60-second Adam-Curtis-style tone poem about consumer optimism in the 1950s. Prefer archival footage. Music yes, narration no."

"Cut together a 2-minute real-footage montage about coming home in the rain. Dreamlike, slow, cinematic, YouTube landscape."

The agent picks the pipeline, builds the corpus, indexes it with CLIP, picks the clips, cuts the edit, and renders the final. Watch it think.

Real example — home is a verb.

We made this with the pipeline. 60 seconds. Real stock footage. Zero dollars. Editorial tagline overlay rendered via Remotion.

The prompt:

"Cut together a 60-second documentary montage about the universal act of walking home. Elegiac, list-shape — one idea, many angles. No narration. Stock footage only. YouTube 16:9."

The pipeline generated the brief, planned 12 slots (3 hero-flagged), built a CLIP corpus from Pexels + Prelinger, picked the shots, cut the rhythm, added an L-cut on the rain-ambient handoff, synced to a music bed, and rendered a 60-second deliverable. We then layered the "home is a verb." editorial tagline as a Remotion polish pass.

final_titled_web.mp4

Pipeline shape

1. idea → brief (duration, tone, shape, thematic question)
2. scene_plan → ordered slots with hero flags and search queries
3. assets → CLIP corpus + ranked picks per slot
4. edit → in/out, transitions, L-cuts, diversity checks, music sync
5. compose → render via FFmpeg or Remotion

Each stage has a director skill in skills/pipelines/documentary-montage/ that teaches the craft, not the code.

Get started

1. Clone the repo
2. Open it in Claude Code, Cursor, or any Claude-Agent-SDK-compatible assistant
3. Say: "Make me a documentary montage about [anything]"
4. Watch it work.

Full pipeline docs: skills/pipelines/documentary-montage/
Pipeline manifest: pipeline_defs/documentary-montage.yaml

We want to see what you make. Drop your result in Show and tell.

[tvcnet]

I am very impressed but I'm still not clear on the zero cost thing...
I did end up adding some other API's, like fai.ai, etc.

But it's not clear to me how the system is creating the video's "free"

And by free I'm guessing you are referring to prompts that use non-AI imagry (youtube, local png's, et al)?

[calesthio]

For this new feature we shipped please make sure to pull the latest from origin
git pull origin
Then provide a prompt something like "Make a 90-second documentary-style montage about what a city feels like at 4am. No narration. Elegiac tone. Use real stock footage only."
The AI will scour the internet with free footage , write a scene plan, use the footage that most accurately represents the theme, add music and produce the final video

[ClaytonFarr]

@calesthio Still getting up to speed on the tooling, but this looks like a sizable step up. Nice work, thank you.

Is there a path that could work with existing footage (e.g. personal documentary footage that includes multiple shots, subjects, etc)? Where OpenMontage could index these to understand content, related content, and help parse it into clips and an edit?

[calesthio]

Yes absolutely. This use case is definitely supported. Here is an example someone used their footage to create a video using OpenMontage
https://x.com/Arjunjain/status/2041716929145221132?s=20

[calesthio]

final_narrated_v3_github.mp4

Here is another video with just a single prompt ""Make a 90-second documentary-style montage about what a city feels like at 4am. Impactful narration and music. Elegiac tone. Use real stock footage only."