AI Frontier Post
Get the Briefing
Tutorials

Turn one sentence into an explainer video: a hands-on guide to Frametell's director agent

Frametell (sitor-ai/frametell, AGPL-3.0, ≈365 GitHub stars) is a local-first app whose director agent turns one sentence — or your PDFs and notes — into an explainer video: you approve the storyboard, it writes and voices the narration, draws every page in your chosen style, and renders a 1080p MP4 with subtitles. This guide installs it, adds your keys, runs the studio and the one-command CLI, and ships your first film.

MDMarcus DoyleSenior Technology Writer · 2026-10-21 · 8 min read
TL;DR: Frametell (sitor-ai/frametell, AGPL-3.0, ≈365 GitHub stars) is a local-first app whose director agent turns one sentence — or your PDFs and notes — into a voiced explainer video: you approve the storyboard first, then it writes and voices the narration, draws every page in one of five built-in styles, and renders a 1080p MP4 with subtitles. This guide installs it (Mac app or source), adds your model and ElevenLabs keys, runs the studio and the one-command CLI, and ships your first film.
  • What: Frametell (sitor-ai/frametell) — a local-first app that turns a sentence or documents into an explainer video through a director agent you approve before anything is drawn.
  • License: GNU AGPL-3.0 (Sitor AI; a commercial license without the copyleft network-use obligation is available via [email protected]).
  • Traction: ≈365 GitHub stars since its 2026-10-10 debut; the README’s demo shows a 72-second film finished in 12 minutes 52 seconds for $1.59.
  • Requirements: Node.js 22.19+, pnpm, and Chrome/Edge/Chromium; signed app for Apple Silicon Macs, build-from-source (FFmpeg bundled) elsewhere.
  • Costs: your own model API key (Anthropic, OpenAI, Google, DeepSeek, OpenRouter, or any compatible endpoint); ElevenLabs key for voiceover (optional for silent films).
  • Output: 1080p MP4 with subtitles and publishing notes; five built-in styles: warm-story, study, chalkboard, whiteboard, tech-dark.

What you’ll need

  • A computer with Node.js 22.19 or later and pnpm installed, plus Chrome, Edge, or Chromium — Frametell renders its pages through the browser. (On Apple Silicon Macs you can skip the build entirely and grab the signed .dmg from the latest release.)
  • An API key from a model provider — Anthropic, OpenAI, Google, DeepSeek, OpenRouter, or any OpenAI- or Anthropic-compatible endpoint.
  • An ElevenLabs API key for the voiceover (optional if you only want the storyboard and silent film; Chinese voiceovers use Volcengine TTS instead).
  • About fifteen minutes and a sentence worth explaining.

Step 1 — Get Frametell running

On an Apple Silicon Mac, download the .dmg from the latest release, open it, and drag Frametell into Applications. It’s signed and notarized by Apple. Everywhere else, build from source — FFmpeg is bundled, so there’s nothing else to install:

git clone https://github.com/sitor-ai/frametell.git && cd frametell
pnpm install && pnpm build

You can also hand it notes, documents, PDFs, slide decks, images, or Markdown instead of a single sentence — the pipeline is the same either way.

Frames of character animation in progress, illustrating how Frametell draws each explainer page with a video engine rather than generating footage
Credit: School of Motion (schoolofmotion.com), via image search

Step 2 — Add your model and voice keys

Open Frametell’s Settings and add a key for your model provider (Anthropic, OpenAI, Google, DeepSeek, OpenRouter…). Add an ElevenLabs key under voice settings for the narration. Running from source? Export them instead:

export ANTHROPIC_API_KEY=...        # or OPENAI_API_KEY, GEMINI_API_KEY …
export ELEVENLABS_API_KEY=...       # optional, for the voiceover

Your materials and API keys stay on your computer — the project is explicitly local-first.

Step 3 — Open the studio, or run it from the command line

You have two front ends. The studio is a browser app where you type ideas, approve storyboards, and watch the agent work:

node apps/cli/dist/cli.js serve

Or skip the UI entirely and go straight from a file to a film. This turns notes.pdf into a 60-second English film in the warm-story style:

node apps/cli/dist/cli.js run notes.pdf --style warm-story --duration 60 --lang en
A hand-drawn storyboard with sketched video frames pinned on a planning board, mirroring the storyboard the Frametell director agent drafts for approval
Credit: M&R Group (mandr-group.com)

Step 4 — Give it one sentence and approve the storyboard

Type what you want to explain — “explain the uncertainty principle” is literally enough — or drop in your documents. A director agent turns your material into pages and story beats, and nothing gets drawn until you’ve signed off on the storyboard. If a page is wrong, point at it and say what to change. The agent narrates what it’s doing and why, and the cost and time spent are on screen at every step.

Step 5 — Narration, voice, and pages

Once you approve the plan, the pipeline runs through its stages: it writes the narration script, voices it (ElevenLabs for English, Volcengine TTS for Chinese), then draws every page in the style you chose — warm-story, study, chalkboard, whiteboard, or tech-dark. Each page passes through lint and screenshot checks, so the agent catches its own rendering mistakes before they reach you.

Step 6 — Collect the finished film

The agent runs a final check on the full video, then renders a 1080p MP4 with subtitles, plus publishing notes for wherever you’ll post it. The project’s own demo shows the scale: a 72-second film, from one sentence to finished render, took 12 minutes 52 seconds and cost $1.59. Two complete English films the repo links — “Why is the sky blue?” (87 s, chalkboard style, made with Gemini) and “Why starting early matters” (72 s, warm-story, compound interest, also Gemini) — were made end to end this way.

What you built

A repeatable explainer-video pipeline on your own machine: one sentence (or a pile of notes) in, storyboard approval, a voiced 1080p MP4 with subtitles and publishing notes out. Because the agent drives a mature video engine with styles, components, camera moves, and quality checks — instead of generating video from scratch — the results are stable and the token spend stays small. And since it talks to any model provider through standard endpoints, you’re not locked to whoever is cheapest this month.

Honest limitations

First, the platform gap: Windows and Intel Macs don’t have a desktop app yet — on those you build from source and live in the CLI/studio. Second, the voiceover is a paid dependency: without an ElevenLabs key (or Volcengine for Chinese), you don’t get narration, and the film quality tracks the strength of the model key you feed it. Third, every demo in the repo is a science or finance explainer in one of the five built-in styles; if you want a visual language the engine doesn’t have, you’re extending the engine, not prompting your way there. Finally, the AGPL-3.0 license means you can use and modify it commercially, but if you offer a modified version over a network you must make your modified source available under the same license.

Frequently asked questions

Do I need a GPU or any local models to run Frametell?

No. Frametell calls a model API through your own key — Anthropic, OpenAI, Google, DeepSeek, OpenRouter, or any OpenAI- or Anthropic-compatible endpoint. Your machine only runs the studio and the rendering pipeline; the intelligence comes over the API.

Does the voiceover cost extra?

The voiceover runs on ElevenLabs for English (or Volcengine TTS for Chinese), so it draws on those keys. Frametell’s own README shows a 72-second film costing $1.59 in total, and marks the ElevenLabs key as optional if you want the storyboard and silent film only.

Can I use Frametell on Windows or Linux?

Yes, but from source. The signed desktop app is macOS Apple Silicon only; Windows and Intel Macs build with Node.js 22.19+, pnpm, and the bundled FFmpeg. You need Chrome, Edge, or Chromium installed for page rendering on every platform.

Related coverage