SkillFoxx

WebGPT skill

A Codex skill that hands tasks to web ChatGPT with project terminal access, saving Codex quota; modes from Medium to Pro.

  • 138 ★ on the repository
  • Last commit
  • MIT
Skill

Broad access

Why this level

  • The Worker gives the ChatGPT web assistant full terminal access as the OS user, and the project folder does not limit it.
  • Access travels through a public HTTPS tunnel address whose link is handed to an outside service.
  • The skill drives browser tabs signed in to ChatGPT and permanently deletes chats.
All reasons and checks

nhahan/webgptAuthor:Nhahan

Install

In your terminal, with SkillFoxx CLI

npx skillfoxx add skills/webgpt

Detects the agents on your machine, shows the access level and pins the version.

Other ways to install

Access and checks

How we review

Broad accessThe tool can change data, run commands or handle secrets.

  • The Worker gives the ChatGPT web assistant full terminal access as the OS user, and the project folder does not limit it.
  • Access travels through a public HTTPS tunnel address whose link is handed to an outside service.
  • The skill drives browser tabs signed in to ChatGPT and permanently deletes chats.

This is third-party code. Review the repository files before installing.

What it does

WebGPT delegates bounded tasks to web ChatGPT while Codex works on something else. Codex opens a chat in the browser where the user is already signed in to ChatGPT, sends the assignment in the chosen mode (`m`, `h`, `xh` or `p`) and collects the outcome. Every task goes to a new chat that is deleted after the result is reviewed. For project work the skill starts a local Worker with terminal access and forwards it over HTTPS without a tunnel account. The address is attached to ChatGPT as a tool, and the web assistant runs commands in the project folder. `webgpt open` opens such a chat for manual terminal work on the project.

Who it is for. ChatGPT subscribers who work in Codex and want to pass some tasks to the web version to spare their Codex quota.

Good fit when

  • The Codex quota is nearly spent and several independent project tasks remain on the plan.
  • A topic research with a final summary goes to Pro mode via `webgpt p`.
  • Planned features are built in parallel, each in its own chat with its own files.
  • You want ChatGPT with the current project's terminal and work by hand through `webgpt open`.

Not a fit when

  • The web assistant must not get a terminal: the Worker runs commands as your OS user.
  • Your home folder holds secrets that cannot be shown to an outside service: the project folder is not an isolation boundary.
  • You need API access without a browser: the skill drives a ChatGPT tab, not an API.

Example request

Use webgpt xh as subagents to develop this project's planned features in parallel.

Limitations

The skill is built for Codex. It needs Node.js 22 or newer, a browser already signed in to ChatGPT and browser control available to the agent. Chats are always created in Chat mode, never Work. To link ChatGPT with the Worker the skill raises an HTTPS tunnel that needs no account or own domain, and the Worker stays running as a service after Codex exits. Commands run as the OS user and the project folder is not a sandbox. The task and code fragments are handled by an outside service, so account for that on private projects. Check OpenAI terms for automating the web interface yourself.

How to disable. Stop the Worker service and tunnel, delete the `webgpt` skill from the Codex skills folder and disconnect the WebGPT Worker tool in ChatGPT. Data sits in `~/.local/share/webgpt` and settings in `~/.config/webgpt/config.json`; delete them separately.

FAQ

Which modes does WebGPT have?

Four: `m` (Medium), `h` (High), `xh` (Extra High, the default) and `p` (Pro). Name the mode in the request: `webgpt p Research this topic and summarize the findings.`

How do I install WebGPT?

The README gives a ready Codex prompt: it asks to install `skills/webgpt`, follow `references/setup.md` and explicitly authorizes the Worker, the tunnel and terminal access. Read the prompt before pasting it.

What does webgpt open do?

It opens ChatGPT with terminal access to the current project or the directory given as an argument: `webgpt open /path/to/project`. The skill never deletes such chats.

Do I need a tunnel account or my own domain?

No. The README asks to proceed without a new tunnel account or domain; the address is forwarded over HTTPS without registration.

What happens to the chat after a task?

The skill keeps the result, then permanently deletes the task chat and closes the tab. `open` chats and unrelated chats stay untouched.

Discussion

No questions or reviews yet.

Skills

An MCP server that gives plain ChatGPT chat a local Codex toolbox: files, commands, Chrome and Edge, live Excel

MCP serverBroad access125

A skill that links two to four open Codex CLI windows: one splits the work while the others execute, cross-check and deliver.

SkillExtended access105
Needs a VPN

A desktop app with a local MCP server that lets ChatGPT work with files, commands, skills and other MCP servers

MCP serverBroad access394
Editors’ pick

OpenAI's coding agent that runs locally in your terminal, with IDE and desktop versions

CLIBroad access135.1K
Needs a VPN

A Go command-line client for working with OpenAI, Azure, Perplexity and LLaMA language models right from the terminal

CLIBroad access966
Needs a VPN

A plugin that lets Claude Code draw images with gpt-image-2 and hand bulk work to GPT-5 subagents, all on your ChatGPT plan

PluginBroad access243

README in short

The README covers one-prompt installation, the `m`, `h`, `xh` and `p` modes and two scenarios: delegating a task and working directly in ChatGPT. Setup details live in `references/setup.md`.

README badge

Original SKILL.md

---
name: webgpt
description: Delegate work to signed-in Web ChatGPT when the user requests WebGPT (m/medium, h/high, xh/xhigh or p/pro), or open a user-controlled project terminal chat with webgpt open.
---

# WebGPT

Delegate bounded work to save Codex usage. User instructions about division of work, model,
monitoring, permissions, follow-ups and retention override defaults. Never silently take over
WebGPT's assignment; Codex may perform separately assigned work concurrently.

Keep Codex ↔ WebGPT coordination in English, preserving source quotations and the requested
deliverable language. Reply in the user’s language unless they request another; leave browser UI language unchanged.
**Every conversation must use Chat, never Work**, including setup tests, browser-only work and
`open`. Verify Chat before composing or selecting a connector; return to Chat if a shortcut opens Work.

## Route

- `webgpt open [path]`: follow [open.md](references/open.md); skip delegation below.
- Worker delegation: read [workspace.md](references/workspace.md).
- Explicit browser-only work: use [browser-reply.md](references/browser-reply.md); skip Worker commands.
- Read [setup.md](references/setup.md) only for installation or missing capability, and
  [recovery.md](references/recovery.md) only for actionable exceptions or same-chat follow-ups.
  Client tunnel/connection errors require setup's origin recovery before dispatch.

For Worker tasks, register first and use the returned `connection.name` and `connection.registrationId`.
With `connection.status: ready`, skip configuration/setup reads, directory searches and Plugins
rediscovery. Missing/invalid metadata routes to recovery.md; keep the registered task, do not register again.

## Delegate

Assign one bounded deliverable with its relevant tests per new chat. Investigate only enough to
specify the objective, context, constraints and success criteria; leave routine commands to WebGPT.
Avoid open-ended projects, arbitrary duration limits and separate chats for tiny actions. Assign
file ownership/worktrees for concurrent changes. No duplicate task documents or tracking files.

Prepare one concise English prompt before opening the tab: task, necessary context/constraints and
success criteria. Reference accessible project briefs by path instead of restating them.
For Worker tasks, add the named connector/task token and “Save the final outcome
through the Worker.” Tool schemas supply the completion protocol; do not repeat it. Never send the
controller key or connection URL. Retain task and owned chat/tab identifiers in context.

Use documented browser initialization; reuse known browser identity without scanning all surfaces.
For Worker tasks, select the existing connector in Chat by verified registration ID, or
use a verified Chat launch URL. Mentioning a connector in prose does not select it.

Report a problem