Project / AI Video

ADI Montage

A self-hosted agentic video studio. Type an idea — "a 60-second explainer on how black holes bend light" — and an agent researches it, writes the script, generates the imagery and motion, narrates it, scores it with music, captions it, and renders a finished MP4. Built on the open-source OpenMontage engine with a custom theLAB dashboard, driven by a local LLM and a fleet of swappable creative providers.

Watch the output on YouTube OpenMontage on GitHub

ADI Montage demo — the type-an-idea to finished-video pipeline animating end-to-end — // Demo — ADI Montage producing a video end-to-end

ADI Montage Create view — type an idea, pick providers, and produce a video — // Create — idea in, providers picked, video out

ADI Montage Create view with Kling video mode and the soundtrack library open — // Create — Kling video mode + the soundtrack library

ADI Montage Create view with a generated scene script, soundtrack library, and per-scene images — // Create — generated script, soundtrack library, per-scene images

ADI Montage producing a video — live scene-by-scene log with generated frame thumbnails — // Producing — live scene-by-scene log with generated frames

ADI Montage producing — per-scene stills animating into Kling motion clips — // Producing — per-scene stills animate into Kling motion

A finished ADI Montage render — 'theLAB Commercial' with player, script, and Download MP4 — // Finished render — “theLAB Commercial”, 6 scenes, script + MP4

A finished ADI Montage render — 'NEON RECKONING', a cinematic cyberpunk short — // Finished render — “NEON RECKONING”, a cyberpunk short

OpenMontage is an open-source (AGPLv3) agentic video production system. theLAB runs it self-hosted on adi-cloud-2 behind a purpose-built dashboard — ADI Montage — that turns the whole pipeline into a single "type an idea → watch it produce → download the video" surface. The agent's reasoning runs on a local Devstral model, and every creative step can fall back to a free/local provider, so a full video can be produced for a few cents — or nothing.

OpenMontage (AGPLv3)

FastAPI dashboard

Devstral (local LLM)

FLUX via fal.ai

ComfyUI (local)

Kling video

ElevenLabs TTS

Pixabay music

WhisperX captions

Remotion render

How it works

One prompt kicks off an agent loop that walks the whole production end to end. Each stage writes a checkpoint, so a run can be guided, paused, or resumed:

idea ─▶ research ─▶ script + scene plan ─▶ imagery ─▶ motion ─▶
       narration ─▶ music ─▶ word-level captions ─▶ Remotion composition ─▶ MP4

Script & scene plan — the local Devstral model researches the topic and writes a narrated, scene-by-scene script.
Imagery — FLUX via fal.ai for fast cloud generation, or local ComfyUI for zero-cost stills.
Motion — either a Ken-Burns slideshow over the stills, or real Kling motion clips (Standard / Pro / Master, 5s or 10s per scene).
Voice & music — ElevenLabs narration with voice preview; royalty-free Pixabay tracks, AI music, or an uploaded track.
Captions — word-level, animated subtitles aligned with WhisperX.
Composition — everything is assembled and rendered to MP4 by a Remotion studio (running alongside on :3300).

Fast or free — pick per step

Every creative provider has a paid "fast" lane and a local/free fallback, chosen right in the dashboard. That's what keeps a finished video in the cents range:

Stage	Fast (paid)	Free / local
Images	`FLUX` via fal.ai (~$0.05/img)	`ComfyUI` on local GPU
Motion	`Kling` clips via fal.ai	Ken-Burns slideshow over stills
Music	`ElevenLabs` AI music	`Pixabay` royalty-free / upload
Scripting	cloud LLM (optional)	`Devstral` local on adi-cortex

Cost-aware by design The dashboard estimates each job's price before you produce it (image count × provider rate, plus Kling per-5s pricing) and tracks cumulative spend. A budget layer can warn or hard-cap runs, with approval required for new paid tools — so an agent can't quietly run up a bill.

Soundtrack Library

Rather than re-uploading the same music for every video, ADI Montage keeps a persistent Soundtrack Library. Add a track once and it's available to every render — in both Create and YouTube Shorts Studio — alongside the per-job Pixabay / ElevenLabs / upload options.

Upload once, reuse everywhere — drop in any audio file from the Library tab; it's stored server-side and shows up in the track picker on every job.
Preview & pick — click a track to play it, then select it for the current video; set a start offset so the music drops in where you want it.
Shared across modes — the same library backs the 16:9 Create flow and the 9:16 YouTube Shorts Studio, so a signature track is one click away in either.
Manage it — saved tracks are listed in the Library tab with play and delete; the selected track is passed straight through to the composer.

# the library is a small REST surface on the dashboard
GET    /api/soundtracks            # list saved tracks
POST   /api/soundtracks            # upload a new track (multipart audio)
DELETE /api/soundtracks/{name}     # remove one
# files are served from /soundtracks and reused across every render

YouTube Shorts Studio

A dedicated 9:16 short-form mode. Give it a hook ("3 mind-blowing facts about self-hosted AI") and it produces a vertical video with word-level captions, then writes a ready-to-paste caption and lets you copy the script or download the clip — tuned for the format instead of cropping a 16:9 render.

Straight from the studio

The videos ADI Montage produces are posted to the @adi_onlin3 YouTube channel.

Visit the channel

Running it

ADI Montage runs as a FastAPI service with a Remotion render studio beside it, both bound to the tailnet on adi-cloud-2 — reachable only over Tailscale, never the open internet:

# dashboard API (serves the ADI Montage UI)
uvicorn app:app --host 0.0.0.0 --port 8750

# Remotion composition studio
remotion studio src/index.tsx --host 0.0.0.0 --port 3300

Credit & license The production engine is OpenMontage by calesthio, released under AGPLv3 — github.com/calesthio/OpenMontage. ADI Montage is theLAB's self-hosted deployment: the custom dashboard, the local-LLM (Devstral) and ComfyUI integrations, and the fleet wiring on adi-cloud-2.