Skip to main content
Glama
yyyqt

gpt-web-agent

by yyyqt
OverviewSchemaSourceRelated ServersDiscussionsAdmin
JavaScript
Local

What is this?

When you code with ChatGPT today, it hands you a snippet; you copy it, paste it, run it, and paste the error back.

GPT Web Agent removes that loop. Install it on your computer and ChatGPT on the web can read your project, edit files and run tests itself, then tell you what happened. You just talk in the chat:

You:   Look at my blog project and tell me what could be improved. Analysis only.
GPT:   [reads package.json, src/…] Found 3 issues: ① homepage images aren't compressed ② …
You:   Fix the first one and run the tests.
GPT:   [edits src/image.js] [runs npm test] Tests pass. Changed 2 files: …

It's a small program (an MCP server) running on your computer, connected to your ChatGPT through OpenAI's official private tunnel. ChatGPT does the thinking; this program does the work on your machine.

Related MCP server: Coding Tools MCP

Why use it?

Copy and paste

GPT Web Agent

Editing code

You copy and paste

GPT edits the files

Running tests

You run them and paste errors back

GPT runs them and reads the results

Understanding a project

You pick files to paste

GPT finds and reads them

Generated images

Download, then move them yourself

One sentence saves them into your project

  • Uses the ChatGPT you already have: this project calls no model API, adds no per-token charges, and doesn't need Codex or any other AI tool.

  • No project setup: say "now look at the YYY project" and it switches. No paths to register.

  • No accidental overwrites: every write checks a content fingerprint (SHA-256) and refuses if you changed the file in the meantime.

  • Keeps running when you close the tab: tests and builds already started keep running on your computer; ask for the result later.

  • Leaves your account alone: no ChatGPT proxying, no browser cookies, only OpenAI's official tunnel.

Quick start

IMPORTANT

Use apersonal account and check two things first: ChatGPT Settings → Security and login has Developer mode, and Platform → Tunnels lets you click Create tunnel. A ChatGPT subscription does not by itself grant tunnel access; company or school accounts usually need an admin.

1. Install (Node.js 22+, macOS, Linux, or native Windows)

npm install --global gpt-web-agent@latest

Windows PowerShell:

npm.cmd install --global gpt-web-agent@latest
gpt-web-agent.cmd --help

Use the same install command to upgrade. gpt-web-agent --help should report 0.6.0 or newer. Restart the connection after upgrading; existing tunnel and key configuration is preserved.

2. Create a tunnel and a key in OpenAI Platform

  • Tunnels → Create tunnel, link your ChatGPT workspace, and copy the tunnel_... ID.

  • API keys → Create new secret key, set permissions to Restricted, and grant only Tunnels → Read and Use.

3. Connect your computer

gpt-web-agent connect

In Windows PowerShell, use gpt-web-agent.cmd connect.

Paste the tunnel ID and key when asked; type YES at the second prompt if you want GPT to run tests. Keep the terminal open once you see tunnel-client started. After every restart, this one command is all you need.

4. Add the connection in ChatGPT

Turn on developer mode → on the plugins page, click + → Create app → Create MCP app → choose Tunnel and No Authentication, tick the risk checkbox → create. In a new chat, select it from the + menu next to the message box, and you're ready.

Screenshots of every screen, success checks and troubleshooting are in the setup guide.

What it can do

Capability

Details

Browse and search

List folders, read files, full-text search

Edit files

Create, rewrite or patch, with overwrite protection

Run commands

Tests, builds, installs; check progress, time out or cancel (enable during setup)

Save images

Store images generated or uploaded in the chat directly in your project

Task notes

Progress checkpoints for multi-step work that survive reconnects

Delegate to Codex

Optional: hand a task to local Codex when you explicitly ask

Safety

WARNING

With commands enabled, anything GPT runs hasthe same power as commands you type in your own terminal. It is not a sandbox. Use it only with your own connection; for sensitive projects you can leave commands off and use file tools only.

  • The runtime key lives in the macOS Keychain, Linux system password store, or a Windows DPAPI CurrentUser-encrypted file, never as plaintext in config files or the repository.

  • The connection goes only through OpenAI's official private tunnel; no port is opened to the internet. Never expose the local server to the internet yourself.

  • See the security notes for details.

FAQ

The project is free and open source and calls no model API. You need a ChatGPT account that can use developer mode and OpenAI Platform tunnel access; any tunnel-related charges are shown in your OpenAI account.

Yes. Native Windows uses Windows PowerShell for opted-in commands, DPAPI for the tunnel key, and the official Windows tunnel-client build. Local Codex delegation and install-service are not yet available on native Windows; the file and command workflow does not require either feature.

The connection drops and ChatGPT can't reach the tools. Run gpt-web-agent connect again. On a Mac, gpt-web-agent install-service connects automatically at login.

The chat doesn't have your connection selected. Click + next to the message box, search for your connection's name, select it, and ask it to call workspace_info first.

Create a new key in Platform, run gpt-web-agent set-key to paste it, then run gpt-web-agent connect again.

More in the setup guide's troubleshooting table.

Documentation

License

MIT

This server cannot be deployed

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Related MCP Connectors

Related MCP Servers

  • VibeLens

    soumyachk101
    A
    license
    A
    quality
    B
    maintenance
    Enables AI coding assistants to visually inspect local web pages by providing screenshots, console errors, and sanitized DOM snapshots.
    1
    4 npm
    MIT

  • Coding Tools MCP

    R-HUA
    F
    license
    Not graded
    quality
    A
    maintenance
    Enables AI agents to develop within a local project workspace by reading and modifying files, running commands and tests, checking Git state, and persisting progress as history sessions that can be restored in later conversations.
    -

  • local-agent-mcp

    ezra-y
    A
    license
    B
    quality
    C
    maintenance
    Enables ChatGPT to work directly on local projects by reading files, editing code, running tests and commands, inspecting Git state, and managing durable workflows, with optional delegation to local Codex agents for complex tasks.
    28
    MIT

  • local-mcp

    shworks-25
    A
    license
    Not graded
    quality
    A
    maintenance
    Enables AI assistants to securely inspect, modify, and manage local projects with trusted filesystem access, safe Git operations, controlled task execution, and developer runtime process management over stdio or HTTP.
    3
    Apache 2.0