Skip to main content
Glama
dengfeng-0908

web-gpt

by dengfeng-0908
OverviewSchemaRelated ServersScoreDiscussions
Python
Local

Codex Web GPT

通过 CLI 或 MCP 向 ChatGPT 网页发送文本任务、等待答案并继续追问。固定程序负责浏览器收发，Codex 只提交任务并读取结果，不需要逐步操作页面或查看截图。

这是一个非官方的实验项目。当前支持 macOS、已安装的 Google Chrome，以及带“即时／中／高／极高／Pro”五档菜单的中文 ChatGPT 网页。其他系统、语言或网页布局尚未验证。

功能

  • 调用时自动启动专用 Chrome，使用独立且持久的浏览器配置目录。

  • 首次手动登录后复用登录态；关闭标签页或退出浏览器后，按保存的会话网址恢复。

  • 新建对话、同一对话追问、超时后继续读取，不自动重复发送结果不明的消息。

  • 支持默认 ChatGPT 项目，也可为新对话单独指定项目。

  • 发送前选择并核对推理档位；默认 xhigh，调用方可按任务难度降低或选择 pro。

  • 提供 CLI 和两个 MCP 工具，等待与页面轮询在程序内部完成。

Related MCP server: ChatGPT Delegate

安装与首次登录

需要 macOS、uv 和 Google Chrome。项目使用 Python 3.12，依赖由 uv.lock 锁定。

git clone https://github.com/dengfeng-0908/codex-web-gpt.git
cd codex-web-gpt
uv sync --locked
.venv/bin/web-gpt browser

在专用 Chrome 中登录自己的 ChatGPT 账号，然后检查并发送一个简单问题：

.venv/bin/web-gpt doctor
.venv/bin/web-gpt ask --effort instant --prompt '请只回复：连接成功'

后续可直接运行 ask 或 read。浏览器未运行时会自动启动，正常重启不要求重复登录；如果网站让登录态失效，仍需本人重新登录。此项目使用网页会话，不需要 OpenAI API key。

CLI

# 默认使用极高档位，新建对话。
.venv/bin/web-gpt ask --prompt '请分析这个问题：……'

# 使用返回的 session_id 继续原对话。
.venv/bin/web-gpt ask --session '<session_id>' --prompt '继续解释第二点'

# 等待未完成的回复；不发送新消息。
.venv/bin/web-gpt read '<session_id>' --timeout 600

# 大段上下文通过明确指定的 UTF-8 文件传入。
.venv/bin/web-gpt ask --effort pro --prompt-file /absolute/path/task.txt

# 将项目首页的完整 URL 设为默认项目。
.venv/bin/web-gpt project 'https://chatgpt.com/g/g-p-.../project'

# 为此次新对话选择其他项目。
.venv/bin/web-gpt ask --project 'https://chatgpt.com/g/g-p-.../project' --prompt '任务内容'

# 查看或取消默认项目。
.venv/bin/web-gpt project
.venv/bin/web-gpt project --clear

项目 URL 示例中的 ... 需要替换为网页实际地址。更改默认项目不会移动已有对话；追问始终使用原会话。

返回 JSON 包含 status、session_id、url、effort 和 effort_label；完成时包含 answer。只有 completed 表示已观察到回复完成，running 表示等待到时，应使用同一编号继续 read。回复等待参数最多为 840 秒；网页启动和加载另有等待时间。

工具只发送 --prompt 或 --prompt-file 指定的文本，不自动上传本地文件。运行中遇到未发送草稿时会停止，不覆盖草稿。

推理档位

effort

网页档位

建议用途

instant

即时

极简单的转换、格式整理，速度优先

medium

中

简单且边界清楚的问题

high

高

较简单但需要一定推理的问题

xhigh

极高

默认；通常的分析、设计和代码审查

pro

Pro

非常复杂的推理或综合分析

由调用方 Codex 理解任务难度，在同一次请求中选择 effort，不另调用模型做路由。人工直接运行 CLI 时，省略参数固定使用 xhigh。

程序操作网页可见的档位滑块，核对实际标签后才发送。所选档位不可用或布局不匹配时返回 effort_unavailable，不静默降档。底层模型版本和可用额度由网页账号决定。

MCP（按需启用）

CLI 可以独立使用，不需要配置 MCP。如果希望在 Codex 中直接看到工具，在可信项目的 .codex/config.toml 中添加 配置示例，将 command 改为本机的绝对路径；已有配置时只合并对应服务段。

[mcp_servers.web-gpt]
command = "/absolute/path/to/codex-web-gpt/.venv/bin/web-gpt"
args = ["mcp"]
startup_timeout_sec = 20
tool_timeout_sec = 1020
enabled_tools = ["ask_chatgpt", "read_chatgpt"]

工具

用途

ask_chatgpt(prompt, session_id?, timeout=600, project_url?, effort="xhigh")

选择档位，发送一次并等待；提供编号则恢复并追问

read_chatgpt(session_id, timeout=600)

恢复并读取已有任务，不发送新消息；timeout=0 在页面加载后读取一次状态

工具说明包含按难度选档、精简任务上下文和超时续读的规则。Skill 可用于包装 CLI 的调用方法，本仓库目前未提供单独的 Skill 安装包。

本机数据与运行限制

  • 专用浏览器配置、默认项目与会话映射位于 ~/Library/Application Support/codex-web-gpt/，不在源码目录内。可以通过 WEB_GPT_DATA_DIR 明确指定其他本机数据目录。

  • 会话映射保存网址、回复 ID 和状态，不保存完整提示词或答案；Chrome 自身仍会保存普通浏览器缓存、历史和站点数据。

  • 不复制日常浏览器 Cookie，不读取 ChatGPT 私有 API，不自动绕过验证码或用量限制。请保管好专用浏览器目录，不将其提交到仓库。

  • 调试连接仅使用本机 127.0.0.1:19231。同时只允许一个请求使用浏览器。

  • 空闲时可以退出 Chrome；正在收发时应保持浏览器与网络可用。电脑睡眠会影响本机执行。

  • 如果首次发送后尚未取得会话 URL 就关闭浏览器，工具可能无法恢复；不会自动重发原任务。

  • 本工具不下载 GPT 权重。Chrome 自身可能下载浏览器组件或内置 AI 模型；当前启动器尚未禁用这类下载。Chrome 模型管理说明

该接口用于分配任务和减少模型逐步操作网页的开销，不增加订阅额度，也不保证固定的节省比例。网页账号自身的权限与限制仍适用。普通 Chat 与 Work 的用量规则应分别核对；官方说明 ChatGPT Work 和 Codex 共享用量。官方说明

开发与验证

uv sync --locked
.venv/bin/python -m unittest discover -s tests -v
git diff --check

测试使用本机 Chrome 和独立临时上下文，通过本地页面夹具验证收发、恢复、项目绑定、五档选择及真实 MCP 协议，不调用真实 ChatGPT。无需运行 playwright install。

2026-09-06 已完成 17 项本地测试，并分别做过真实 CLI 收发、独立 MCP 追问、关闭标签页后的只读恢复、退出浏览器后的免重新登录恢复和五档切换验证。即时档做过真实收发；Pro 仅验证切换生效，未以长任务验证其推理效果。真实账号验证不代表所有账号或后续网页版本兼容。

网页选择器集中在 src/web_gpt/chatgpt.py。报告问题时请附上系统、Chrome 版本和脱敏状态，不附 Cookie、浏览器配置目录或完整私人对话。

参考与许可证

连接方式参考了 mcp-web-llm 的架构方向，本项目聚焦 ChatGPT 独立实现。调研资源还包括 coding-tools-mcp、mybolide/coding-tools-mcp、webcodex 和 ChatCodex。

依赖与接口参考：Playwright CDP、MCP Python SDK、Codex MCP 配置。

本项目使用 MIT License。

Available Tools

2 tools
ask_chatgptA

Send one task and wait. Choose effort: normally xhigh; instant/medium/high for simpler faster tasks; pro for very complex tasks. Chrome starts with saved login. Omit session_id for a new chat; reuse it to restore/continue. project_url overrides the default project for new chats. timeout: 1–840 seconds.

ParametersJSON Schema
NameRequiredDescriptionDefault
effortNoxhigh
promptYes
timeoutNo
session_idNo
project_urlNo

TDQS

A4.3/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations are minimal (readOnly false, openWorld true, destructive false), so the description adds meaningful context: Chrome starts with a saved login, session_id restores/continues chats, project_url overrides the default project, and timeout is bounded at 1–840 seconds. It does not contradict annotations and adds behavioral detail beyond them.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is three compact sentences with no filler. It front-loads the core action, then efficiently covers effort selection, session handling, project override, and timeout.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a tool with five parameters and no output schema, the description covers invocation, effort choices, session lifecycle, project selection, authentication context, and timeout. The main gap is that it does not describe the return value or response format, and it does not mention the relationship to read_chatgpt.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 0%, but the description compensates comprehensively: it explains effort options, session_id semantics, project_url behavior, and the timeout range. Even though 'prompt' is not explicitly described, it is clearly implied by 'Send one task'.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose4/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the action as 'Send one task and wait', naming a specific verb and resource (ChatGPT). It does not explicitly distinguish itself from the sibling tool read_chatgpt, so it misses the full differentiation criterion.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description provides practical usage guidance: when to choose instant/medium/high vs pro effort, how to start a new chat, how to continue one with session_id, and how to override the project with project_url. It does not explicitly state when to use this tool instead of read_chatgpt, so it lacks explicit exclusions.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

read_chatgptA
Read-only

Restore and read a previous chat without sending a message. Chrome starts automatically with saved login. timeout: 0–840 seconds; use 0 for a snapshot after loading.

ParametersJSON Schema
NameRequiredDescriptionDefault
timeoutNo
session_idYes

TDQS

A4.4/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

The description adds valuable behavioral context beyond the annotations: Chrome starts automatically, a saved login is used, timeout has a 0–840 second range, and timeout 0 produces a snapshot after loading. This aligns with readOnlyHint and destructiveHint and provides concrete operational expectations.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is compact and front-loaded with the core purpose in the first sentence. The second sentence adds essential behavioral details without unnecessary filler.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a simple read-only tool with two parameters, the description is largely complete: it explains side effects, timeout behavior, and the fact that no message is sent. It could more explicitly describe what session_id should contain or how it relates to prior chats, but the overall picture is sufficient.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

With 0% schema description coverage, the description must compensate for parameter documentation. It explains timeout semantics well, including range and special value 0, but the required session_id is left mostly to inference from 'previous chat' and the property name.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the tool's function: restoring and reading a previous chat without sending a message. This distinguishes it from the sibling ask_chatgpt by explicitly noting that no message is sent.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The phrase 'without sending a message' provides clear context for when to use this tool rather than ask_chatgpt. However, it does not explicitly name the alternative or state when to choose ask_chatgpt instead.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections. Dates show when Glama detected each change.

  1. 2 tool updatesv0.1.0
    • First observedask_chatgpt
    • First observedread_chatgpt
Install Server
RayrunBuild, deploy, and operate MCP servers without becoming an infrastructure team.ray.runAd

TDQS

A4.4/5.0
Disambiguation5/5

ask_chatgpt sends tasks and awaits responses, while read_chatgpt only restores and reads previous chats without sending. Their purposes are distinct and unlikely to be confused.

Naming Consistency5/5

Both tools follow a consistent verb_chatgpt snake_case pattern: ask_chatgpt and read_chatgpt. The naming convention is predictable and clear.

Tool Count3/5

With only 2 tools, the surface feels thin even though both tools are clearly useful. The server covers a narrow scope, but the count is at the low end of the acceptable range.

Completeness4/5

The core workflow of sending a task and reading a previous chat is well covered. Minor gaps exist, such as no explicit delete or list-chats operation, but agents can work around them using session_id handling.

Maintenance

ActivityMaintained
ResponsivenessNo issues

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Related MCP Connectors

Related MCP Servers

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/dengfeng-0908/codex-web-gpt'

If you have feedback or need assistance with the MCP directory API, please join our Discord server