v1.0.0

The first stable release. CursorCam now works with any AI agent, and npm i cursorcam and npx cursorcam get it without @beta.

• New cursorcam mcp command: an MCP server, so Cursor, VS Code, Codex, Gemini CLI, Windsurf, Claude Desktop and any other MCP app can make videos. Add it with npx -y cursorcam mcp. Its tools match the CLI commands. Recordings run in the background and the agent follows them with wait, so no call times out. inspect, stills and failed runs return their frames as images, so the agent can see its own video. Secrets passed to check, record and run never reach a file, a log or the tool output.
• Native installs for every major AI agent. Most add the MCP server and the skill that teaches the whole flow:
  • Codex: codex plugin marketplace add Avijit07x/cursorcam, then codex plugin add cursorcam@cursorcam.
  • VS Code and GitHub Copilot: add Avijit07x/cursorcam to chat.plugins.marketplaces, then install it from @agentPlugins.
  • Gemini CLI: gemini extensions install https://github.com/Avijit07x/cursorcam, which also adds a /cursorcam command.
  • Cursor: a one-click MCP install link.
  • Claude Desktop: a one-click extension, cursorcam.mcpb, attached to each release. It brings its own Node.js.
  • Claude Code keeps its plugin.
• The MCP server works in the project you have open: it asks the app for its folders, so videos, steps files and relative paths land in your project, even when a plugin starts it from its own folder.
• The skill works with the MCP tools or the CLI, whichever the agent has.
• On Windows, CursorCam finds the browser even when the app that starts it passes only a few environment variables.
• The README and the docs now cover any AI agent, not only Claude Code, and every command uses npx cursorcam without @beta.

v1.0.0-beta.4

Patch Changes

• npm i cursorcam and npx cursorcam now get the newest beta. Until the first stable release, each beta becomes npm's latest as well as beta, so a plain install no longer falls back to the first beta.

v1.0.0-beta.3

Patch Changes

• Clicking something in a fixed header or a sticky bar no longer scrolls the page first. The recorder used to scroll again and again to move it away from the screen edge, which showed in the video as a slow scroll before the click.

v1.0.0-beta.2

Minor Changes

• New scene backgrounds: "background": { "scene": "glow", "colors": "indigo" }. Pick from 14 scenes, like mesh, glow, dunes and glass, and color them with one of 9 palettes or any two hex colors. Each scene fits every video shape and window size.

Patch Changes

• Opening a page now gets at least 30 seconds, even when the steps set a shorter timeout. A short timeout meant for steps no longer fails a slow first page with "Could not open".

v1.0.0-beta.1

Major Changes

• First beta release. Ask Claude for a demo video: the Claude Code plugin adds the cursorcam skill, so /cursorcam http://localhost:3000 — sign up and create a project makes Claude explore the app, write the steps, check them, record and render in a hidden browser, review stills of its own video and hand back the MP4.
  • Install it with /plugin marketplace add Avijit07x/cursorcam, then /plugin install cursorcam@cursorcam.
  • Claude starts the app's dev server when it is not running, and stops only the servers it started.
  • When the steps need something only you have, like a login, a 2FA code or which account to use, Claude asks in two clear lines: what it needs and why, then Reply with: and the format.
  • A login you give is used for that run only. Claude passes it to the one command that needs it and never writes it to a file, a log or memory.
  • The plugin blocks commands that would print every environment variable, like env or printenv, in every session. Everything else runs as before.
• The cursorcam CLI records a click demo of any web app from a steps file and renders a polished MP4: zoom on each click and typed text, a smooth cursor with click ripples, a gradient or image background, rounded corners, a shadow and a browser bar with the page URL.
• Ten commands:
  • doctor checks Node, the browser, screen capture, the video encoder, crash dumps, disk space and fonts, and prints a fix for each problem.
  • inspect <url> saves the first frame of a page and lists what can be clicked, with a target for each that is checked to match exactly one element. --filter and --limit narrow the list, and it warns about things a video can't show, such as native date pickers and tooltips.
  • check <steps> is a dry run that saves one frame per step and stops at the first failure.
  • record <steps> records without rendering, render <run> renders a recording, and run <steps> does both.
  • stills <run> saves key-moment stills at video size and one close-up crop per step.
  • wait [run] waits until a run finishes, fails or needs an answer, and says which in one line.
  • answer <run> passes a reply to a run waiting at an ask step.
  • login <url> opens a browser window to log in by hand, for sites with a captcha or a bot check, and saves the login. Record with --profile <name> from then on. The saved login holds the site's cookies, not the password, in a file only you can read, and --forget deletes it.
• --detach starts check, record or run in the background and prints the run folder at once, and check --until <step> checks only the first steps.
• Recording uses the Chrome, Edge, Chromium or Brave already on the machine, version 120 or newer, in a hidden window. Nothing else is downloaded. Desktop runs capture at 2560×1600 and phone runs at 1170×2532, so zoomed shots stay sharp.
• The page behaves like a real browser on camera: hover effects work, scrolling is smooth wheel scrolling, scrollbars are hidden and pages never flash white.
• Steps support goto, click, dblclick, hover, type, press, select, upload, scroll, drag, waitFor, pause and ask. Every step also takes zoom, speed, pauseAfter and dialog.
• The camera zooms in while the cursor travels, so the zoom lands before the click. It stays zoomed while the next actions are in the same area, and zooms out when the flow moves somewhere else or ends. It never pans across the page while zoomed. "zoom": { "hold": 1000 } in the style keeps it zoomed after each action, and a hover step with a zoom holds on that part for the whole step.
• Each step pauses 700 ms after it, so viewers see what changed. pauseAfter sets a longer or shorter pause, and "zoom": false keeps the whole page in view for a step.
• Targets read like the page: role=button[name='Save'], label=Email, text=Sign in, placeholder=…, testid=… or plain CSS. When a page has two matches, the error lists both, and nth, within or near picks one. Targets inside iframes (frame) and shadow DOM work too.
• Checkboxes and radio buttons that the page draws itself, over an invisible input, can be clicked like any other target.
• Steps never hang. Each step waits for its target to be visible, still, enabled and uncovered, then for page loads, network requests and fonts to settle, with a time limit (timeout, 15 s by default). A cookie banner in the way is named in the error.
• New tabs are followed and left when they close. Alerts, confirms and prompts are answered by the step's dialog option, downloads are saved in the run folder, and uploads work with styled file pickers.
• Typing is exact for emoji, flags, CJK and accented text, and the field value is checked after typing.
• Phone recordings ("viewport": "phone") tap instead of click and show a touch dot instead of a cursor.
• Logins stay private:
  • $secret:NAME reads CURSORCAM_SECRET_NAME from that one command, and $env:NAME reads a variable you already have. Values never reach a file, a log or the terminal output.
  • A secret typed into a field that is not a password field is blurred in the video, and mask blurs any element in every frame.
  • An ask step waits for a one-time code, passed with cursorcam answer. The code is never stored.
• Bot checks and captcha pages stop the run with exit code 10 instead of recording a blocked page.
• Videos are 1920×1080 at 60 fps, H.264 with standard BT.709 color. Each file is read back after encoding to check its size and frame count.
• --for youtube, x, linkedin and discord match each platform's size, frame rate and limits, and --max-size 10MB fits any file size limit.
• --format square makes a 1080×1080 video, and --format vertical makes a 1080×1920 video from a phone recording.
• The camera holds while typing and cuts on page loads and tab switches. zoom sets the level per step, or turns it off.
• Idle waits are sped up, while pause, pauseAfter and the ending play at normal speed. While a page loads, the last good frame is held.
• --style changes the background (six gradients, any color, a custom gradient or an image file), padding, corner radius, shadow, browser bar, URL text, cursor size, ripples, zoom limit, dialog cards and idle trimming. An image path in a style file starts from that file's folder.
• Each render also saves a poster image, the video's first frame, next to the video, like poster.jpg for video.mp4.
• A recording can be rendered again with a new look or preset without recording again. Raw frames are kept for 7 days, or deleted right away with --clean.
• Every run gets its own folder in cursorcam-output/, with status.json while it runs and result.json when it ends, and --json prints the result for scripts.
• Each failure has its own exit code and a clear message with a hint, and a failed step saves failure.jpg.
• A run never leaves a browser running, writes no crash dumps, and stops cleanly when the disk runs low.