can-see

MCP server that lets AI agents see and interact with terminal/CLI applications through virtual terminals and PNG screenshots.

Built for Claude Code and any MCP-compatible agent.

Why?

Some things are easier to show than describe. When debugging a TUI app, an interactive CLI wizard, or anything with visual terminal output, can-see lets the agent see exactly what you see — colors, layout, cursor position, and all.

Related MCP server: textual-mcp-server

How it works

1. Launch a CLI app in a virtual terminal (node-pty + @xterm/headless)

2. Screenshot the terminal as a PNG image (rendered via node-canvas)

3. Send keys/text to interact with the app

4. Screenshot again to see the result

5. Close the session when done

Installation

npm install -g can-see

Prerequisites

can-see depends on node-canvas (Cairo) and node-pty, which require native compilation. Most systems will need:

• Windows: Visual Studio Build Tools (C++ workload) — npm install --global windows-build-tools or install from Visual Studio Installer

• macOS: Xcode Command Line Tools — xcode-select --install

• Linux: sudo apt install build-essential libcairo2-dev libjpeg-dev libpango1.0-dev libgif-dev librsvg2-dev

Configuration

Claude Code

Add to your project's .mcp.json:

{
  "mcpServers": {
    "can-see": {
      "command": "npx",
      "args": ["-y", "can-see"]
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "can-see": {
      "command": "can-see"
    }
  }
}

Other MCP clients

can-see uses stdio transport. Point your MCP client at the can-see binary or npx -y can-see.

Tools

Supported keys

Enter, Tab, Escape, Backspace, Space, Up, Down, Left, Right, Home, End, Delete, PageUp, PageDown, Ctrl+A through Ctrl+Z.

Environment variables

Example usage

From an MCP-connected agent:

Agent: I'll launch your app to see what's happening.
→ launch("node", ["app.js"])  → sessionId: "abc-123"

Agent: Let me wait for the app to start.
→ wait_for_text("abc-123", "Ready")  → Found "Ready" after 1200ms

Agent: Let me read the current output.
→ read_text("abc-123")  → "Welcome to MyApp\nReady\n> "

Agent: I can see the prompt. Let me select option 2.
→ send_keys("abc-123", ["Down", "Enter"])

Agent: Waiting for the screen to settle.
→ wait_for_idle("abc-123")  → Terminal idle for 520ms

Agent: Let me check the result.
→ screenshot("abc-123")  → [PNG image showing result]

Agent: Done, closing the session.
→ close("abc-123")

Changelog

0.5.0

New tools:

• wait_for_exit — wait for process exit, get exit code and signal

• close_all — kill all active sessions at once

• get_process_status — distinguish "app is idle" from "app has exited"

• screenshot_text_region — find text in viewport, capture surrounding area as PNG

Enhancements:

• launch accepts env parameter for custom environment variables

• wait_for_idle supports excludePattern (regex) for dynamic row exclusion in stableMs mode

• stop_recording returns frameCount and durationMs metadata alongside GIF

• get_cell_info supports compact option for reduced output ({char, fg, bold} only)

Bug fixes:

• Fixed wait_for_text and wait_for_color race condition where text/color present in the final buffer was missed when the process exited simultaneously

• Added mutual exclusion validation when both stableMs and idleMs are passed to wait_for_idle

License

MIT