MCP 服务器

chatgpt-webui-mcp

MCP server that drives chatgpt.com via web UI automation using a session token, enabling long-running tasks, deep research, and image generation.

README

mcp server for querying chatgpt (chatgpt.com) via webui session token (the image above was generated with this!)

quick start

install from npm:

npm i -g chatgpt-webui-mcp

manual run:

CHATGPT_SESSION_TOKEN="your_token_here" chatgpt-webui-mcp

from source:

npm install
npm run build
CHATGPT_SESSION_TOKEN="your_token_here" node dist/index.js

important: this uses chatgpt's internal webui api with a session cookie. for personal/local tinkering only - not affiliated with openai.

overview

chatgpt-webui-mcp is a standalone MCP server that drives chatgpt.com via camofox (UI automation).

it is built for long-running tasks (gpt-5.2 pro runs that take 1h+), deep research, and image generation mode.

getting your session token

open https://chatgpt.com and log in
open devtools
application -> cookies -> https://chatgpt.com
copy the value of __Secure-next-auth.session-token

configuration

because this server uses stdio or sse, you configure it as a local command (or remote url) and pass the token via env.

mcp client config (claude desktop, opencode, etc)

{
  "mcpServers": {
    "chatgpt-webui": {
      "command": "node",
      "args": ["/absolute/path/to/chatgpt-webui-mcp/dist/index.js"],
       "timeout": 7200000,
        "env": {
          "CHATGPT_SESSION_TOKEN_FILE": "/path/to/session-token.txt",
          "CHATGPT_BROWSER_BASE_URL": "http://127.0.0.1:9377",
          "CHATGPT_WAIT_TIMEOUT_MS": "7200000"
        }
     }
   }
 }

legacy CHATGPT_CAMOFOX_* env vars are still supported for compatibility. CHATGPT_TRANSPORT=httpcloak is intentionally unsupported (use camofox). when model/thinking are omitted, requests default to gpt-5-2 (auto), not pro.

opencode workflow (the natural language style)

if you want to type commands like:

with chatgpt webui on gpt 5.2 pro extended thinking: <prompt>
do deepresearch with chatgpt webui on <topic>

use this tool:

tool	what it does
`chatgpt_webui_command`	parses your sentence into the right call and runs it

example:

{
  "name": "chatgpt_webui_command",
  "arguments": {
    "command": "with chatgpt webui on gpt 5.2 pro extended thinking: write a 1-page memo about X",
    "mode": "auto"
  }
}

long runs (recommended)

use the unified tools:

tool	description
`chatgpt_webui_prompt`	main tool. `mode=auto` chooses wait vs background
`chatgpt_webui_run`	check/wait for background runs (`run_id`)

why: deep research and gpt-5.2 pro can take a long time and may exceed a single client timeout. mode=auto returns a run_id for long jobs.

image generation

set create_image=true to switch chatgpt into image generation mode before sending the prompt.

notes:

image_urls is best-effort (derived from page links + visited urls) and may be empty depending on how chatgpt renders images in the webui.
fallback screenshot output is returned in image_data_url (not image_urls) when enabled and size-capped.
enable fallback with CHATGPT_IMAGE_SCREENSHOT_FALLBACK=1.
cap fallback size with CHATGPT_IMAGE_SCREENSHOT_MAX_BYTES (default 2097152, 2 MiB).
for reliable retrieval, you can also use the conversation_id and open the chatgpt UI.

self-test

# env
CHATGPT_SESSION_TOKEN="your_token_here" npm run self-test

# cli flag
npm run self-test -- --token "your_token_here"

# file
echo "your_token_here" > ~/.config/chatgpt-webui-mcp/session-token.txt
npm run self-test -- --token-file ~/.config/chatgpt-webui-mcp/session-token.txt

remote deployment over tailscale (optional)

if you want background runs to survive for a long time, run this server as an always-on SSE service.

copy templates from this repo:

deploy/systemd/chatgpt-webui-mcp.env.example
deploy/systemd/chatgpt-webui-mcp-sse.sh
deploy/systemd/chatgpt-webui-mcp.service

install and enable service (user service):

mkdir -p ~/.config ~/.config/systemd/user ~/.local/bin ~/.local/share/chatgpt-webui-mcp
cp deploy/systemd/chatgpt-webui-mcp.env.example ~/.config/chatgpt-webui-mcp.env
cp deploy/systemd/chatgpt-webui-mcp-sse.sh ~/.local/bin/chatgpt-webui-mcp-sse.sh
cp deploy/systemd/chatgpt-webui-mcp.service ~/.config/systemd/user/chatgpt-webui-mcp.service
chmod 600 ~/.config/chatgpt-webui-mcp.env
chmod 755 ~/.local/bin/chatgpt-webui-mcp-sse.sh
systemctl --user daemon-reload
systemctl --user enable --now chatgpt-webui-mcp.service

point opencode (cloud host) to the endpoint:

{
  "mcp": {
    "chatgpt-webui": {
      "type": "remote",
      "url": "http://<tailscale-ip>:8791/sse",
      "enabled": true,
      "timeout": 7200000,
      "oauth": false
    }
  }
}

tools

tool	description
`chatgpt_webui_session`	validate token and return session payload
`chatgpt_webui_models`	list available models
`chatgpt_webui_command`	natural-language command wrapper
`chatgpt_webui_prompt`	unified prompt tool (wait/background)
`chatgpt_webui_run`	check/wait for background runs
`chatgpt_webui_ask`	direct wait-style prompt tool (legacy/simple)

project structure

chatgpt-webui-mcp/
├── deploy/
│   └── systemd/
│       ├── chatgpt-webui-mcp.env.example
│       ├── chatgpt-webui-mcp-sse.sh
│       └── chatgpt-webui-mcp.service
├── src/
│   ├── index.ts               # MCP server
│   └── chatgpt-webui-client.ts # WebUI automation client
├── package.json
├── tsconfig.json
├── .env.example
├── .gitignore
├── LICENSE
├── INSTALL.md
└── README.md