Capsulate

@capsulate/mcp

The Capsulate Model Context Protocol server. Give Claude Code, Claude Desktop, Cursor, Windsurf, Antigravity, any MCP client, the ability to launch and drive isolated cloud computers ("capsules"), including real Linux and Windows desktops that a person can watch live and take over.

Setup

1. Create an API key in the Capsulate dashboard: Settings → API & Keys.
2. Add the server to your MCP client config.

Claude Code

claude mcp add capsulate -e CAPSULATE_API_KEY=cap_live_... -- npx -y @capsulate/mcp

Claude Desktop / Cursor

{

"mcpServers": { "capsulate": { "command": "npx", "args": ["-y", "@capsulate/mcp"], "env": { "CAPSULATE_API_KEY": "cap_live_…" } } } }

To keep the key out of the config file, put it in a file and point at it with CAPSULATE_API_KEY_FILE instead. The file holds either the bare key or a CAPSULATE_API_KEY=cap_live_... line.

For a self-hosted or non-default deployment, also set "CAPSULATE_BASE_URL": "https://api.your-domain".

Tools

Desktop computer use

The input verbs accept screenshot: true (and screenshotDelayMs) to return the frame after the action. Every tool works the same on Linux and Windows desktops; a multi-line editor is a document on Windows and a text on Linux.

Browser fast path

The capsule's own Chrome, driven through the page's structure at page speed, on the same screen the person watches: capsulate_browser_open, _navigate, _tabs, _snapshot (an outline with refs), _click, _fill (the value is never echoed), _select, _press, _scroll, _wait, _text, _eval and _screenshot.

Live: a person watches, takes over, hands back

A desktop capsule has one shared screen. The AI drives it through the tools above while a person watches the same screen in a browser, sees the AI's pointer, clicks and keys drawn over it, reads its narration, and can take the screen at any time.

While a person holds the screen, the AI's input verbs answer human_in_control and nothing reaches the capsule; the tool result says to wait with capsulate_wait_for_handback. Reading the screen (screenshot, UI tree, OCR) keeps working. Each AI input holds the screen for 45 seconds; the console is view-only for everyone else during that time and turns interactive again when the AI goes quiet or releases.

Skills: learn once, replay in minutes

The server's instructions carry the learning protocol: explore with a goal, note what you learn, finish, read the digest, save the skill; next time launch its environment and run its procedure.

Try it

"Spin up an ubuntu-22.04-headless capsule, run my test suite, and show me the output."

The model calls capsulate_launch, then capsulate_exec, and reports back.

"Launch an ubuntu-22.04-desktop capsule, give me a live link, and find the latest Firefox release

notes. If anything needs a login, hand it to me."

The model launches the desktop, calls capsulate_console_link and gives you the URL, narrates with capsulate_say as it works, and calls capsulate_handoff when it needs you. You press Done, hand back in the live console when you finish, and it carries on.

Develop

npm run build                          # from the repo root: sdk, ai-tools, cli, mcp

node apps/mcp/scripts/smoke-live.mjs # live tools end to end against a fake API