Open source CLI

Every picture lands on its word.

A markdown script and HTML slides in, a narrated mp4 out. Edit one line and only its section rebuilds.

$ curl -LsSf https://decktalk.ai/install.sh | sh

No API key, no sudo. Read it first.

Measured

A cue waits for a word, not a second.

The voice times every word. Each picture names the phrase it waits for. A browser records the slides frame by frame.

7.8s: to build a 50-second film from cold. 4.6 s after one word changes.; The starter project with a click for each word in place of the voice, median of three on an Apple M5 Pro, Chromium and ffmpeg already fetched.
29ms: the furthest any of Follows' 33 pictures landed from its word. Less than one frame. The median is 15 ms.; Measured on the 30 fps film above by decktalk verify, against the time the voice gave each word.
$0.02: to voice the one changed line again after "Share" became "Send". The other line was kept. The first build, which voiced both lines, cost $0.03.; The rebuild Follows shows (run 629b5a01): 60 characters at $0.30 per 1,000, as that build printed it.

Films

Asked first

Why not Remotion or Manim?: They time a picture in frames or seconds, in code. DeckTalk times it from the spoken word, so a rewritten sentence moves its pictures with it, and the slides are your own HTML. For frame-by-frame animation, they are the better tools. The comparison, tool by tool.
What is 29 ms measured against?: decktalk verify decodes the finished mp4, finds the first frame where each picture changes, and subtracts the time the voice gave that word. A build with no voice is measured against the clicks in its own audio. How verify measures.
Is it tied to ElevenLabs?: The voice is ElevenLabs today, a stock voice or a clone of your own. The script, the cues and the slides name no service, and what DeckTalk needs from any voice is a start and an end time for every word.
Do I write JavaScript?: No. A slide is HTML and CSS, and each thing that appears names its cue in a data-in attribute. Headless Chromium records the page against the word times.
What comes out besides the mp4?: Captions as SRT and VTT, chapters, a transcript page, a poster and every word's time as JSON, all built from the same word times. Every file.
Can a coding agent run it?: Yes. Every command prints JSON with --json. Run without a terminal, a build with something to buy refuses unless --spend or --no-spend is passed, so an agent left alone cannot spend by accident.

Before you install

Costs: Free. The voice is your own ElevenLabs key, billed by ElevenLabs per character, and decktalk build --no-spend builds the whole film for nothing, with a click on each word where the voice will be. The build asks before it spends, and --max-cost refuses a run that could cost more than you set.
Status: Alpha The CLI and file formats still change between releases.
Runs on: macOS, and Linux with glibc. On Windows, uv tool install decktalk from PowerShell. The installer brings its own Python through uv.
Downloads: Chromium and ffmpeg on the first build, about 347 MB on a Mac with Apple silicon. It says so before it fetches them.
Accounts: None for DeckTalk. ElevenLabs is the one voice provider today.
Licence: Apache-2.0. Source, issues and changelog on GitHub.

Install

No account, no key, no Python to install first.

Installonce per machine

$ curl -LsSf https://decktalk.ai/install.sh | sh

Installs uv, then DeckTalk. Never runs sudo. Read it first, or add -s -- --dry-run after sh to see what it would do. Have uv already? uv tool install decktalk. Command not found? Open a new shell.
Start a projectfor each film

$ decktalk init my-pitch

Writes a script, its cues and a slide page that already build.
Buildafter every edit

$ cd my-pitch && decktalk build --no-spend

Writes build/final/my-pitch.mp4 with a click on each word where the voice will be. No key.

Your own voice: Put an ElevenLabs key in .env (copy .env.example) and your voice id in decktalk.toml. The build asks before it spends. What a rebuild costs.
Two voices, motion, any shape: decktalk init --example interview or --example motion writes a finished project. One setting makes a film wide, square or vertical.
Coding agents: Every project carries six skills in twelve files. An agent writes the files, reads the JSON every command prints, and fixes what it finds.
Offline, CI and Linux: decktalk install fetches Chromium and ffmpeg ahead of time. On Linux it also installs Chromium's system libraries, the one step that asks for sudo.