
Agent-screenplay:一句提示词即可自动生成精美屏幕录制视频
来源标题:Show HN: Agent-screenplay – Create beautiful screen recordings with one prompt
Agent-screenplay 发布,用户只需一句提示词,AI 代理会探索网站或桌面应用并编写演示脚本,由 Screenplay 在浏览器中自动录制并完成平滑光标、自动缩放、点击特效、字幕等后期编辑。该工具本身不含 AI,需配合 Claude Code、Codex、Cursor 等 MCP 客户端使用,通过 npx agent-screenplay init 安装。
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Create beautiful screen recordings with one prompt. Any website, yours or someone else's, or any desktop app. Works with Claude Code, Codex, Cursor and any MCP client.
Ask your agent: "Make a video that shows how human handoff works". The agent explores the site or app and writes a short scenario; Screenplay plays it in a browser (or a desktop app), films it, and edits the video like Screen Studio would: smooth cursor, automatic zooms that follow the action, click effects, callouts, spotlight, captions, title cards, and an optional voice-over.
Screenplay contains no AI and needs no model key: your agent does the thinking, Screenplay does the filming and editing.
npx agent-screenplay init
The npm package is agent-screenplay; installed globally (npm i -g agent-screenplay), its command is screenplay.
init downloads Chromium (once), adds the Screenplay MCP server to the agents it finds (Claude Code, Codex, Cursor), and installs the guide that teaches them to make good videos. Restart your agent after it.
In Claude Code, type /screenplay and say what you want:
/screenplay Make a 30-second walkthrough of https://askli.co /screenplay Film how to invite a teammate in our app, with a French voice-over
In Codex, Cursor or any agent with the MCP server, just ask: "Make a screen recording of …". The agent explores the site, asks whether you want a voice-over, writes the scenario, films it, checks the storyboard and gives you the MP4. To change the look afterwards: npx agent-screenplay style <scenario.json>.
Without MCP, any agent that can run commands can use the CLI directly:
npx agent-screenplay guide # how to write a good scenario npx agent-screenplay check demo.json # plays the steps fast, without filming npx agent-screenplay record demo.json # films and edits → demo.mp4 + demo.storyboard.png npx agent-screenplay render demo.json # re-edits the last take (style, zoom, text, voice) without filming
Requirements: Node 20+. ffmpeg and the drawing library ship with the package.
{ "url": "https://app.example.com/inbox", "style": { "background": "dawn" }, "steps": [ { "title": "Human handoff, in one click", "subtitle": "New in 2.4" }, { "click": "text=Camille Laurent", "say": "When the AI agent gets stuck, the conversation lands in your inbox." }, { "callout": "role=button[name='Take over']", "text": "Take over from the AI agent" }, { "click": "role=button[name='Take over']", "say": "One click, and the conversation is yours." }, { "type": "Hi Camille, I'm on it.", "into": "placeholder=Write a reply…" }, { "click": "role=button[name='Send']" }, { "highlight": "text=Assigned to you", "text": "Your teammates see it's yours" }, { "end": "Available today on every plan" } ] }
Targets name elements the way people see them: text=…, role=button[name='…'], label=…, placeholder=…, testid=…, css=… (web only), at=x,y (desktop apps without accessible elements only).
Actions: goto, click, type, press, hover, scroll, select, wait, pause.
Direction: title and end cards, callout (a bubble pointing at an element), highlight (spotlight), zoomTo, and on any step say (caption and voice), zoom, hold.
Style: background preset (soft light gradients: dawn, iris, citrus, apricot, coral, blaze, cloud, glacier, tide, meadow), your own gradient, a color or an image; padding, corner radius, shadow, the cursor (arrow by default, a rounded arrow in the accent color; macos, dot) and its click effect (ripple by default; rays, confetti, sparkle), maximum zoom, browser frame, accent color. See Style.
Output: 1920×1080 at 30 fps by default; any size up to 4K, 24 to 60 fps, optional GIF.
The full format is the JSON schema in schema/scenario.schema.json (npx agent-screenplay schema), and the agent guide in skills/screenplay/SKILL.md. Examples are in examples/.
check and record really perform the steps (a "Send" click sends): point scenarios at a demo account or a staging environment.
Set the look by eye: npx agent-screenplay style demo.json opens an editor in your browser, on a frame of the last recording. Pick a gradient (or make your own), a color or an image (drop it on the page), set the padding, corners and shadow, then save to this video or to all videos, and render without filming again.
"All videos" writes screenplay.config.json, which applies to every scenario in its folder and below; a scenario's own style wins, setting by setting. The same settings by hand:
{ "style": { "background": { "gradient": ["#0ea5e9", "#6366f1"], "angle": 135 }, "padding": 96, "radius": 20, "shadow": true } }
Title and end cards (title and end steps) are off by default: turn them on with "cards": true, or with the switch in the editor.
background is a preset name (dawn by default; earlier presets such as aurora or ocean still work), a color ("#0f172a"), an image path (relative to the file that sets it), or { "gradient": [2 to 4 colors], "angle": degrees }.
npx agent-screenplay login https://app.example.com opens a browser window: sign in, then close the window. The session is saved to .screenplay/auth.json (while you go, and when the window closes; Enter in the terminal works too). Use it with "auth": { "storageState": ".screenplay/auth.json" }. Keep that file out of git.
Bot checks (Cloudflare's "Just a moment…"): Screenplay's browser presents itself as an ordinary Chrome, so the checks that need no person pass by themselves. When a site asks a person to prove they are human, check says so: run login on that site, pass the check in the window, close it, and set auth.storageState. Some sites ask again on every visit; those cannot be filmed automatically.
Agents without a screen can sign in with steps: { "type": "${EMAIL}", "into": "label=Email" }, { "type": "${PASSWORD}", "into": "label=Password", "secret": true }. ${VAR} is read from the environment; secret masks the text in captions and logs.
Voice-over
Off by default. Turn it on per scenario:
"voice": { "enabled": true, "language": "fr" }
Each step's say is spoken, and the video waits for each sentence to finish, so the voice sets the pace. Three ways to get the voice:
Narration is cached: re-rendering with unchanged text costs nothing. voice.id picks another voice, voice.speed its pace. Without voice, say still shows as a caption.
Use "app": "<name>" instead of "url". The steps are played through the operating system's accessibility APIs by agent-desktop (installed with Screenplay), and the window is filmed with ffmpeg without the system cursor; the editing is the same as for websites.
Desktop recordings need a real screen: they cannot run in a headless container. Websites can.
While Screenplay drives the computer, a pill at the top of the screen says so ("Screenplay is recording", or "rehearsing" during a check), with a Stop button that ends the run at once. The pill is left out of the capture, so it is never in the video. macOS and Windows; SCREENPLAY_INDICATOR=0 turns it off.
"screen": true films the whole main screen instead of one window, for videos that go from one app to another or use the menu bar. { "open": "<app>" } launches an app or brings it to the front, a step's "app" acts in another app, and "surface": "menubar" (or menu, sheet, popover, alert) finds elements outside the app's window. Everything on screen is filmed, desktop icons and notifications included: tidy the screen first; on macOS "hideOthers": true hides the other apps.
scenario.json ──▶ play + film ──────────────────▶ take (frames + timeline of actions) web: Playwright, headless Chromium │ desktop: agent-desktop + ffmpeg ▼ edit: time remap (long waits sped up), camera (spring zooms), cursor path, overlays, captions, voice track │ ▼ draw each frame (Skia) → ffmpeg → MP4
Screenplay knows exactly where and when each click happens, and the box of every element it acts on, so zooms are framed on the element instead of guessed from mouse movement. The cursor is drawn during editing, along a smooth human-like path.
npm install npm run build npm test # unit tests npm run test:e2e # records a real scenario in headless Chromium and talks to the MCP server
Releasing
npm version patch --no-git-tag-version # or minor / major; also updates src/version.ts
Merge that change, then tag the commit on main:
git tag v0.1.4 origin/main && git push origin v0.1.4
The tag starts .github/workflows/release.yml. It checks that the tag matches the version and is on main, runs the tests, publishes to npm (with provenance) and creates the GitHub release with notes from the merged pull requests. A version like 0.2.0-beta.1 goes to npm's next tag. Creating a release in GitHub's UI with a new tag works too.
examples/demo-app is a small web app for trying scenarios; examples/desktop-app/notes.py is a GTK app for desktop recording on Linux.
MIT. Bundled: Inter (SIL Open Font License), ffmpeg through ffmpeg-static (GPL), agent-desktop (Apache-2.0).
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。