Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Cursor Learn Mode Record this workflow for generating a monthly report"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Learn Mode learns workflow intent. It is not a mouse macro recorder.
Record a Windows or macOS demonstration, get a standard Cursor Skill (SKILL.md + workflow.json), then replay in a new chat with tools you already have — desktop computer-use, browser MCP, and the terminal.
Learn Mode does not execute workflows and does not invent a replay engine.
Human demonstration
→ OS Observer (Windows / macOS)
→ events.jsonl + screenshots
→ semantic analysis
→ workflow + SKILL.md
→ Cursor Agent uses existing toolsSay /learn in Cursor.
Start Learning and demonstrate the process.
Stop (Ctrl+Shift+L / Cmd+Shift+L).
Review title, inputs, and steps (edit if needed).
Optionally Teach Another Example with different values.
Save Skill → ~/.cursor/skills/<name>/.
Later, ask Cursor to perform the task with new inputs. The agent reads the Skill and chooses existing tools.
Related MCP server: atomic-computer-mcp
Windows 10/11 or macOS
Node.js 20+
Cursor
Windows: .NET 8 SDK
macOS: Xcode Command Line Tools (swiftc) + Accessibility permission for Cursor
From a clone:
git clone https://github.com/liad07/cursor-learn-mode.git
cd cursor-learn-mode
npm install
node scripts/cli.mjs installOr:
npm run install-mcpThe installer detects the OS, builds the observer, and merges a learn-mode entry into ~/.cursor/mcp.json without removing other MCP servers.
Reload MCP in Cursor. On macOS, enable Cursor under System Settings → Privacy & Security → Accessibility.
Manual MCP entry (if you prefer): see mcp.json.example.
Publishing as
npx cursor-learn-mode installis prepared (bin+files). The package stays"private": trueuntil you intentionally publish to npm.
Passwords and secret-looking fields are redacted. Screenshots are skipped when the focused control is sensitive.
Optional environment flags (passed through to the observer):
Variable | Default | Effect |
|
| Set |
|
| Set |
Session metadata may include recordingOptions (screenshots, clipboard, privacyMode).
After the first Stop + preview, choose Teach Another Example.
Demo 1 and Demo 2 with different field values help the analyzer mark those fields as inputs and keep shared navigation as structure. All demonstration IDs are stored on the Skill.
Recorded sessions/workflows include platform: "win32" | "darwin". Generated Skills recommend the matching desktop tools for the platform where the demo was recorded, while staying intent-based.
learn_open, learn_start, learn_status, learn_pause, learn_stop, learn_add_demonstration, learn_save, learn_list, learn_discard
Password / secure fields → [REDACTED]
Tokens, Bearer headers, and similar clipboard secrets are not kept as plaintext
Skills never embed credentials
Do not commit .learn-recordings/
MIT © liad07
This server cannot be deployed
System-of-record notebook for AI coding agents: pages, datastores, tasks, skills over MCP.
Search & install 6,500+ AI agent skills from skills-hub.ai inside any MCP tool.
Eyes and hands on real Windows PCs — observe, click, type via Glasswarp API.
Run multi-step tasks in a real Chrome browser: persistent environments, live view, human takeover.
