AI director

This is the most hands-off way to use the screenify CLI: it builds on web demos by letting an AI both drive the demo for you — no action script needed — and direct the cinematic effects on the moments that matter. You bring your own AI key; screenify does the rest.

Bring your own AI key

Agent features call a language model, so you bring a key from an AI provider — screenify doesn't resell AI. Set one of these environment variables and screenify auto-detects the provider:

You only pay your own provider for the tokens used — a typical demo is a few cents.

Setting the key

The key is read from your environment. Pick one of three ways:

# 1. Persist it (recommended) — add to your shell profile, then open a new terminal
echo 'export OPENROUTER_API_KEY="sk-or-..."' >> ~/.zshrc
source ~/.zshrc

# 2. For the current terminal only
export OPENROUTER_API_KEY="sk-or-..."

# 3. Inline, just for one command
OPENROUTER_API_KEY="sk-or-..." screenify web record https://your-app.com -p "demo it"

Swap ~/.zshrc for ~/.bashrc if you use bash. Once set, screenify doctor --json reports that agent mode is ready, and any agent command picks the key up automatically.

Prefer not to touch your shell? In the desktop app, open Settings → Command Line → AI API key, paste your key, and Save. The CLI uses it automatically — no env var needed. An env var of the same name still takes priority if you set one.

Let the AI drive the demo

Instead of an action script, give a plain-language task with -p:

screenify web record https://your-app.com \
  -p "demo the landing page: scroll through the main sections and click the primary call-to-action" \
  --model openai/gpt-4o-mini \
  -o demo.mp4 -- --wallpaper orange --padding 16

The agent reads the page, decides the next action, performs it, and repeats — then screenify finishes it with the same smooth cursor, auto-zoom, and idle-trim as a scripted demo. As it goes, it can also emphasize the moments it judges important with a cinematic effect.

Model tip: pick a model that's reliable at tool use — openai/gpt-4o-mini (via OpenRouter) is a good, cheap default. Add --vision to also send the agent screenshots of the page (slower and more tokens, but better on visually complex sites).

Useful options: --provider, --model, --max-steps <n>, --vision, and --no-agent-effects (let the agent drive but add no effects).

Add cinematic effects

The effects command layers cinematic moments onto any recording — most naturally a web demo, where the click points are already known. It's non-destructive: it renders a styled copy and leaves your project untouched unless you pass --save.

There are two ways to direct it.

Plain recipe

Spell out exactly what you want with --plan "<effect> [preset] [count|all] [param=value …]" — param=value tunes any knob listed in effects schema (e.g. follow amount=2 for a 2x follow-cursor zoom):

# A 3D camera tour across the 3 most important clicks
screenify effects demo.screenify --plan "zoom3d tour 3" -o cinematic.mp4

# A glowing callout on the 2 key moments
screenify effects demo.screenify --plan "callout neon 2" -o cinematic.mp4

AI director

Add --ai and describe the intent — a language model reads what each moment is and picks which to emphasize and which effect fits each:

screenify effects demo.screenify --ai \
  --plan "make this feel cinematic and draw attention to the main call-to-action" \
  -o cinematic.mp4

--dry-run shows what would be added without changing anything; --save writes the effects back into the project.

The effect catalog

Run screenify effects schema --json for the full, always-current catalog — every effect, preset, and tunable parameter.

Two AI roles, one tool

It's worth separating the two ways AI shows up here:

• web record -p — the AI performs the demo (clicks, scrolls) and can emphasize moments live.
• effects --ai — the AI directs the effects on an already-recorded demo, choosing which moments deserve a flourish.

Use them together (drive with -p, then refine with effects --ai) or on their own. And remember the whole CLI is itself drivable by an outer agent like Claude Code — see For agents & CI.