Screen Browser

Screen Browser skill

Narrated demo and tutorial videos of your web app, from a guide or your coding agent.

Ask the agent that knows your codebase for a tutorial video of a feature. It writes the walkthrough in the words your buttons and fields use, Screen Browser records a real browser following it against your deployed app, narrates it, adds the on-screen effects, and the agent hands you the MP4 with a GIF, subtitles and chapters. When the app changes, edit a line and ask again.

[The first chapter of a tutorial video Claude Code made from one sentence: a campaign being created in Chatamatic, narrated and captioned]

That is the first chapter of a 66-second video made from the request "make a tutorial video of creating a campaign". The whole session, from that sentence to the finished video with the exact prompt and every screen along the way, is in the Claude Code guide.

This repository gives your agent two things:

• The screenbrowser skill — how to read your codebase and write the two guides Screen Browser needs (the login flow and the walkthrough), which effect fits which moment, how to check a guide for free before recording, and how to review the video before handing it over.
• The MCP server configuration — the tools the agent calls: create the project, upload guides, check them, start the run, poll it, fetch the video. The server itself is part of your Screen Browser account at https://mcp.screenbrowser.com.

Claude Code, Codex, Cursor or any MCP client can use it. Sign-in happens in the browser; there is no key to paste.

Install (Claude Code)

Add this repository as a plugin marketplace and install the plugin:

claude plugin marketplace add screenbrowser/skillclaude plugin install screenbrowser@screenbrowser

The first time the agent reaches for a Screen Browser tool, Claude Code opens your browser: sign in to Screen Browser, approve the connection, and you are done. Nothing to copy. (If it does not prompt, run /mcp and pick screenbrowser → Authenticate.)

Without the plugin system, register the server yourself and copy the skill:

claude mcp add --transport http screenbrowser https://mcp.screenbrowser.comcp -r skills/screenbrowser ~/.claude/skills/

With an API key instead

If your agent cannot open a browser to sign in, create a key under Settings → API keys in Screen Browser and pass it as a bearer token:

claude mcp add --transport http screenbrowser https://mcp.screenbrowser.com \  --header "Authorization: Bearer <your API key>"

Any MCP client works the same way: Streamable HTTP at https://mcp.screenbrowser.com, OAuth 2.1 with dynamic client registration for interactive sign-in, or Authorization: Bearer <key>.

Other agents

The skill is a plain SKILL.md in the open Agent Skills format, and the server is standard MCP over HTTP, so the same two pieces work outside Claude Code.

Cursor, VS Code, GitHub Copilot and other Agent Plugins clients — this repository is also an Agent Plugins package (plugin.json, mcp.json, skills/), so install it as a plugin from the repository URL and sign in when the client offers. Cursor without the plugin: add the server to .cursor/mcp.json (or the global one in Cursor settings), click Sign in next to it, and copy the skill into .cursor/skills/:

{  "mcpServers": {    "screenbrowser": { "url": "https://mcp.screenbrowser.com" }  }}

mkdir -p .cursor/skills && cp -r skills/screenbrowser .cursor/skills/

OpenAI Codex CLI — add the server to ~/.codex/config.toml, sign in once, and copy the skill into Codex's skills folder:

[mcp_servers.screenbrowser]url = "https://mcp.screenbrowser.com"

codex mcp login screenbrowsermkdir -p ~/.codex/skills && cp -r skills/screenbrowser ~/.codex/skills/

Anything else (Gemini CLI, Windsurf, a custom agent) — register an HTTP MCP server with the URL the way that tool documents it; a client that supports OAuth will sign you in, one that does not takes the API key as an Authorization: Bearer header. Give the agent skills/screenbrowser/SKILL.md as instructions (paste it, or point the tool's rules file at it). The MCP server itself also carries the guide syntax as a resource and a write-tutorial-guide prompt, so a client with MCP prompts can work without the skill file at all.

Use

Videos can be recorded on the desktop or as a phone or tablet (device: iphone, android, ipad, …), in a phone frame on a vertical 9:16 or widescreen 16:9 canvas or as the bare screen; one run is one device.

Open your app's repository in Claude Code and ask:

Make a tutorial video of creating a campaign.

The agent will check credits, create the project (it will ask you for the deployed URL, a demo user, and the legal attestation), write the guides from your routes and templates, validate them, show them to you, start the run, and give you the video link a few minutes later.

The MCP server also ships a get_guide_instructions tool with the same writing rules, for clients without this skill, a write-tutorial-guide prompt that walks any client through the same flow, and two resources: screenbrowser://guide-syntax and screenbrowser://example-guides.

What your app needs

• A demo user that signs in with email and password: no OAuth, no 2FA, no captcha. Seed it with realistic data.
• The deployed URL to record (staging is fine) and every host the app talks to (API subdomain, auth provider, CDN).
• Confirmation that you are allowed to record the app (yours, a client's with permission, or an internal tool).

What it sends, and where

Everything the plugin does goes to one place: the Screen Browser MCP server at https://mcp.screenbrowser.com, over HTTPS, signed in with your Screen Browser account. The skill itself is instructions for the agent; it runs nothing on your machine and fetches nothing.

What the agent sends to Screen Browser when you ask for a video:

• the project's settings: your app's URL, the hosts it talks to, the sign-in check, narration language and voice;
• the guides it writes from your code: plain sentences naming buttons and fields as they appear on screen. Read them before they are saved; the agent shows them to you;
• the demo user's credentials, as project variables. Screen Browser stores them encrypted, uses them only to sign in during a recording, and never returns them through any tool. Values you give the agent pass through its conversation; to keep them out of chat, set them on the project's page in the Screen Browser dashboard instead;
• run commands: start, poll, cancel, and, only if you agree, a failure report with the recorder's screenshot and logs of a run that failed on Screen Browser's side.

It never sends your source code. The agent reads your repository locally to learn the labels; only the guide text leaves the machine. Nothing is sent to any other service. What Screen Browser keeps, and for how long, is in the privacy policy.

What it costs

Plans start at $49 a month, sized in minutes of finished video. Checking a guide before recording is free, and a run that fails is not charged. Videos are capped at five minutes. There is no free plan, but the first 100 accounts get 10 free minutes with no card. Current prices are on the pricing page.

Layout

plugin.json                        Agent Plugins manifest (Cursor, VS Code, Copilot, …)mcp.json                           Agent Plugins MCP config (the URL; sign-in is OAuth).claude-plugin/plugin.json         Claude Code plugin manifest.claude-plugin/marketplace.json    lets this repo act as its own Claude Code marketplace.mcp.json                          Claude Code MCP server configskills/screenbrowser/SKILL.md      the skillskills/screenbrowser/references/   guide syntax reference, worked exampleexamples/chatamatic/               a real auth guide + main guide pair

What a guide looks like

Plain language, one step per line, naming things as they appear on screen. Screen Browser turns it into the script — the narration, the highlights, the pacing — and resolves each step against the live page:

Go to https://app.acme.test/campaigns"Let's create a spring promotion."Click "New campaign"Type "Spring promo" into the Name fieldClick "Save"Confirm text "Campaign created" is visible

Ids and name= attributes are allowed when a step needs precision; class names generated by a build tool are refused. The full reference, including the optional effect directives, is in skills/screenbrowser/references/guide-syntax.md.

Documentation index for AI agents

The skill carries everything an agent needs to write a guide and run a recording. For anything else — prerequisites, costs, limitations, the API — https://screenbrowser.com/llms.txt is a Markdown index of the public documentation, and https://screenbrowser.com/docs/agents/ is the reference for this integration.

Support

Issues and ideas about the skill are welcome in this repository. Anything about your account or a particular recording: use the contact form at https://screenbrowser.com/contact/.

License

MIT. Screen Browser itself is a hosted service; this repository contains only the agent-facing skill, configuration and examples.