AgentPilot

Drive local chat and task execution via ChatGPT web — no API key, no token limits, unlimited use.

AgentPilot — A lightweight AI agent powered by ChatGPT web. No API, no token limits, unlimited use. It can write code, find files, and much more; explore what you can do with it.

---

Requirements

• Node.js 18+ · Download or install via CLI:

  winget install OpenJS.NodeJS.LTS --source winget
  # If terminal cannot find npm, run first:
  $env:Path = "C:\Program Files\nodejs;" + $env:Path

• Python 3 · Required for the agent (e.g. agent_loop.py, executor). Download or winget install Python.Python.3.12.
• Chrome browser
• Windows (PowerShell)

Install and first run

git clone https://github.com/springcell/AgentPilot.git
cd AgentPilot
npm install
npm run chrome    # Start Chrome, open https://chatgpt.com/ in browser and log in

Commands

Command	Description
npm run run	One-shot run: Chrome → API → Agent (recommended)
npm run run -- your task	One-shot run with task, e.g. npm run run -- find 10 today news and put on desktop
npm run chrome	Start Chrome only (log in at chatgpt.com)
npm run api	Start API server only (port 3000)
npm run agent	Start agent only (Chrome + API must be running first)

Agent

• Connects to ChatGPT web via Chrome; AI plans and executes tasks locally (e.g. write files, run commands, push code).
• Self-learning: learned skills are saved as skills to avoid forgetting.
• Use /new to start a new conversation (e.g. npm run agent "/new your task" or type /new task in interactive mode).
• Example: npm run agent "push this repo to branch AgentPilotCN"; output goes to terminal; some tasks (e.g. "10 news today") write a txt on the desktop.

Usage examples

Run with a task: npm run run -- "your task" or start the agent and type your task when prompted.

Task (English)	What it does
Generate a tiger image and save it to my desktop	AI generates an image in ChatGPT and the agent saves the file to your desktop.
Draw a sunset over the sea, save to D:\Pictures	Same as above; output goes to the folder you specify.
Write a simple alarm clock app and save it to my desktop	AI writes code (e.g. HTML/JS or a small script); the agent saves the file to your desktop.
Create a small countdown timer and put it in D:\Projects	Writes the app and saves it under the given path.
Find 10 today news and put on desktop	Fetches/summarizes news and saves a file (e.g. txt) on the desktop.

Use natural language; you can say save to desktop, save to D:\xxx, or put in folder YYY.

Known issues

npm run run auto-detects and starts Chrome + API. If you still get errors after running it:

• Chrome was closed → Run npm run chrome again and confirm you are logged in at https://chatgpt.com/ in the browser.
• API service died → Restart with npm run api.
• Multiple Chrome instances → Close all Chrome windows, then run npm run run again.

After a system/PC restart, you may see:

Error: AgentPilot request failed (500): {"ok":false,"error":"Protocol error: Connection closed."}

The CDP connection to Chrome is gone. Run npm run chrome (or npm run run) again so Chrome starts with the debug port, then confirm you are logged in at https://chatgpt.com/ before using the agent.

(or your branch name).

Roadmap

• IDE integration for API (e.g. Cursor)
• Vision module
• Voice input

License

MIT

Latest: direct_delivery now requires real file_op write when task asks for a file on disk (no prose-only completion); README usage examples and write_code/validation-loop improvements.