Developer gives a web AI chat hands: an MCP server with 14 tools and 21,000 lines of code
A web AI chat like Notion could not build a project: its sandbox had no internet access to install dependencies. Using MCP, a protocol that lets a chat call tools on your own machine, the author built ZeroIDE, a Windows server with 14 tools and about 21,000 lines of agent-written code. The chat can run commands and see the screen, but by default writes and commands wait for your click. This is not a sandbox: commands run with your user's privileges.
AI-processed from Habr AI; edited by Hamidun News
Imagine asking an AI in a web chat to build a project, and it throws up its hands. Its sandbox has no internet, so it cannot install dependencies or run the project. A developer who hit this problem in Notion solved it with MCP (Model Context Protocol).
MCP lets you give a model tools that live not on its side but on yours: the chat says "call run_command with these arguments," your server executes the command and returns the result. That idea grew into ZeroIDE MCP: 14 tools (commands, file reads and writes, screenshots, a task journal) and about 21,000 lines of code that the author cannot write by hand. For you, it is both a ready-made tool for Windows 10/11 and a practical case study in how to hand an AI control of a computer safely.
How a script grew into 14 tools
The first version was tiny: an MCP server on a VPS with three capabilities — run a command, read a file, write a file. The VPS became a sandbox with internet access for the model. Notion's sandbox now has internet, but the author kept developing the project. New tools came from annoyances, not from a plan:
- two agents work in parallel, and you need an asynchronous connection so they don't get in each other's way;
- a build takes three minutes and the chat times out, so commands run in the background and their status is checked by ID via job_status and job_output;
- an agent took ten steps while you were away, and you can't tell what it did. That led to journal_entry and journal_log for the journal, task_status for task status, and a user_message channel so the agent can write to you directly.
Every call comes with a short "why" note. The author considers this one of his best decisions: the confirmation window shows not only the command but also the agent's explanation.
How a browser chat reaches your home PC
When the VPS was no longer enough, the project moved to Windows and turned from a "script for myself" into a product. The app lives in the system tray, the server is off by default, and it shuts down automatically when idle. File reads are allowed immediately, while writes and commands wait for your button press (human-in-the-loop, HITL). The take_screenshot tool lets the agent see the screen, but it must state why it needs the screenshot.
The server listens only on 127.0.0.1, which is enough for Claude Code, Cursor or Codex on the same PC. ChatGPT, Claude in the browser and Notion need an internet-facing address, and the app sets up three routes:
- ngrok for web chats (the recommended option): you paste in a key, and the app downloads ngrok, verifies its signature and launches it together with the server;
- Tailscale for your own computers: a private network with no public address. An agent on a laptop controls the home PC, or one agent works on a task across several machines at once;
- Cloudflare Tunnel, if you have your own domain.
Provider keys are entered only in the app window, never in the chat. Before the tunnel is enabled, the address is checked: the link must lead to this specific server. Web chats use OAuth with confirmation in the tray. Other clients use named tokens with a scope (read, read and write, full), a folder the token cannot leave, an expiration date and a request limit. Any token can be revoked with one button, and the audit log records every call.
Vibe coding by the numbers, and lessons learned
The first commit is dated October 24, release 0.3.0 came out on November 11, and 0.4.0, with an install wizard, Tailscale and uninstall support, on November 15. Under the hood, the server runs on Node.js, the tray app is C# and WinForms, the installer and helpers are PowerShell, and CI runs automated tests (about 250 checks). Different agents wrote the code, from Muse Spark 1.3 to Fable 5.1, while the author set tasks, reviewed the results and scolded them.
The main lessons for anyone doing the same:
- AI confidently writes insecure code. In an early version, the token could be passed in the query string, and if the token file was unreadable, authorization was simply switched off. A separate agent, tasked with auditing from an attacker's perspective, found this. The server now works on a fail-closed principle: no credentials means a 401 response to everything.
- A claim of "done" needs checking. The installer crashed at step two on a clean machine because the first failed attempt had already created the folder. The author ran the installation on a clean PC about five times and admits he probably did not catch everything.
- A product for strangers is a different kind of work: clear error messages, uninstalling without leftovers, and an honest README.
How safe is it?
The author warns plainly: this is not a sandbox. An approved command runs with the privileges of your Windows user, and the app says so on its first screen. He himself uses the "trusting" profile and has had no incidents so far, but the default settings are cautious, and confirmations can be configured separately for each tool.
The practical takeaway is simple: security is only as good as your willingness to read the calls. If you connect an agent to a work computer, restrict the token's folder, keep confirmations on for writes and commands, and look at the audit log from time to time. The author will add Linux and macOS if there is demand, but he has nothing to test on, so he is looking for volunteers. The code is on GitHub: Danywalery/ZeroIDE-MCP.
- Nov 18, 2026OpenAI releases GPT-6.2 Astra: second place in the rankings and a 20% lower price
- Nov 18, 2026Three "token-saving" utilities raised the AI agent bill by 18–46%
- Nov 18, 2026Meituan open-sources 1.6-trillion-parameter agentic model trained without Nvidia
- Nov 12, 2026Jensen Huang steps down as Nvidia CEO: company loses $1.4 trillion in two weeks
Key Questions Answered
How safe is it?
The author warns plainly: this is not a sandbox. An approved command runs with the privileges of your Windows user, and the app says so on its first screen. He himself uses the "trusting" profile and has had no incidents so far, but the default settings are cautious, and confirmations can be configured separately for each tool. The practical takeaway is simple: security is only as good as your willingness to read the calls.
How a script grew into 14 tools?
The first version was tiny: an MCP server on a VPS with three capabilities — run a command, read a file, write a file. The VPS became a sandbox with internet access for the model. Notion's sandbox now has internet, but the author kept developing the project.
How a browser chat reaches your home PC?
When the VPS was no longer enough, the project moved to Windows and turned from a "script for myself" into a product. The app lives in the system tray, the server is off by default, and it shuts down automatically when idle. File reads are allowed immediately, while writes and commands wait for your button press (human-in-the-loop, HITL). The take_screenshot tool lets the agent see the screen, but it must state why it needs the screenshot.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.