NPC

Your browser's NPC. Handles the side quests.
Control your real browser from any IDE - no context switching.

npm i -g npc-agent

What it does

Your IDE agent says what to do ("message Anna on Slack"). NPC does it in your real, logged-in browser. No API keys per service, no OAuth, no bot accounts. Works with any MCP IDE - Cursor, VS Code, Windsurf.

| You want to | Your IDE says | | ------------------------ | ------------------------------------------------------------ | | Message someone on Slack | "go to slack and message #general: deploy is done" | | Reply on Messenger | "open messenger and reply to Tim: sounds good" | | Check Gmail | "take a screenshot of my gmail inbox" | | Fill out a form | "find the email field, click it, type my address, press Tab" | | Do it all in one shot | use npc_batch with an array of actions |

Setup

1. Install: npm i -g npc-agent
2. Load the extension: chrome://extensions > Developer mode > Load unpacked > select extension/
3. Add to .cursor/mcp.json or .vscode/mcp.json:

{
  "mcpServers": {
    "npc": {
      "command": "node",
      "args": ["/path/to/npc/dist/index.js"]
    }
  }
}

Click the NPC icon on any tab. Green badge means connected.

How it works

IDE (Cursor / VS Code)       NPC Server              Browser
 |                            |                       |
 |--- MCP stdio ------------->|                       |
 |    "click Send button"     |--- WebSocket :7221 -->|
 |                            |                       |--- CDP (chrome.debugger)
 |                            |                       |--- clicks in real tab
 |                            |<-- result ------------|
 |<-- tool response ----------|                       |

The IDE handles reasoning. NPC just executes browser actions via Chrome DevTools Protocol. No LLM inside NPC.

MCP tools

| Tool | What it does | | ------------------ | ----------------------------------------------------------- | | npc_screenshot | Capture tab as PNG | | npc_navigate | Go to a URL | | npc_click | Click at (x, y) | | npc_type | Type text into focused element | | npc_press_key | Press Enter, Tab, Escape, arrows | | npc_scroll | Scroll up/down/left/right | | npc_find | Find element by CSS selector or text, returns (x, y) center | | npc_batch | Run multiple actions in one call | | npc_evaluate | Run JavaScript in page context | | npc_extract_text | Get all text from the page | | npc_extract_html | Get full page HTML | | npc_current_url | Get current tab URL | | npc_page_title | Get current tab title |

Batch example

One MCP call instead of four:

[
  {"action": "find", "selector": "Message Tim"},
  {"action": "click", "x": 450, "y": 320},
  {"action": "type", "text": "hey, deploy is done"},
  {"action": "key", "key": "Enter"}
]

Limitations

Chrome and Brave only (uses chrome.debugger API). One active tab at a time per extension instance. Cannot attach to chrome://, brave://, or extension pages. Screenshot coordinates are at device pixel ratio - divide by DPR before clicking on HiDPI displays.

Requires Node.js >= 18.

Contact