oracle-gpt

name

oracle-gpt

description

Route questions to ChatGPT via the Oracle MCP tool. Use this skill whenever the user wants to ask ChatGPT/GPT something, consult GPT, or get a second opinion from GPT. Trigger on: '问GPT', '问ChatGPT', 'ask GPT', 'ask ChatGPT', '用Oracle', 'oracle', '咨询GPT', '让GPT', 'GPT帮我', '问一下GPT', '让ChatGPT看看', 'consult GPT', '问问GPT', '发给GPT'. Also trigger when the user says things like 'what does ChatGPT think', 'get GPT's opinion', 'ChatGPT怎么说', or any request that clearly intends to forward a question to ChatGPT rather than have you answer it yourself. Do NOT trigger for general questions the user wants YOU to answer — only when they explicitly want ChatGPT involved.

Oracle GPT — Ask ChatGPT via MCP

This skill lets you send questions to ChatGPT through the Oracle MCP server and return the answers to the user. Oracle automates a real ChatGPT browser session on a remote server, so responses take 20-60 seconds.

Prerequisites

The Oracle MCP must be configured in ~/.claude/.mcp.json. If the mcp__oracle__consult tool is not available, guide the user through first-time setup:

1. Install Oracle CLI: npm install -g git+https://github.com/orange4664/oracle.git
2. Ask the user to provide their Oracle server address (host:port) and authentication token
3. Write the MCP config to ~/.claude/.mcp.json:

{
  "mcpServers": {
    "oracle": {
      "command": "oracle-mcp",
      "args": [],
      "env": {
        "ORACLE_REMOTE_HOST": "<user-provided-host:port>",
        "ORACLE_REMOTE_TOKEN": "<user-provided-token>"
      }
    }
  }
}

4. Tell the user to restart Claude Code for MCP to take effect

Recommended Oracle CLI config in ~/.oracle/config.json:

{
  browser: {
    remoteHost: "<user-provided-host:port>",
    remoteToken: "<user-provided-token>",
    modelStrategy: "ignore",
  },
}

browser.modelStrategy: "ignore" keeps Oracle from touching the ChatGPT model picker DOM. This is the safest default for remote/manual-login browser sessions.

If the MCP is already configured (the mcp__oracle__consult tool is available), skip setup entirely.

Conversation State

You need to track three pieces of state across the session:

• chatgpt_url — the ChatGPT conversation URL to reuse. Starts empty. Gets set after the first successful call.
• session_initialized — whether the user has already chosen a conversation mode. Starts false.
• use_remote — whether to use a remote Oracle server or the local browser. Starts unset. Gets set on first use.

These are conceptual variables you maintain in your working memory throughout the conversation — not files or environment variables.

Workflow

First time using Oracle in this session — choose execution mode

Before anything else, on the very first Oracle call in the session, use AskUserQuestion to ask the user whether to run Oracle locally or on a remote server:

Question: "Oracle 使用哪种运行模式？"

Options:

1. 远程服务器（推荐） — "通过远程服务器执行浏览器自动化，不会打开本地浏览器。需要已配置远程服务器地址。"
2. 本地浏览器 — "在本机启动浏览器执行操作。会打开一个 Chrome 窗口，执行期间请勿关闭。"

Then:

• If 远程服务器: Set use_remote to true. Read ORACLE_REMOTE_HOST and ORACLE_REMOTE_TOKEN from ~/.claude/.mcp.json env vars. If they are not configured, ask the user to provide the server address and token, then update the MCP config accordingly. When use_remote is true, always use the Oracle CLI with --remote-host and --remote-token flags instead of the mcp__oracle__consult MCP tool, because the MCP tool may not correctly route to the remote server. Remote oracle serve hosts now default ChatGPT requests to modelStrategy=ignore; if the host is older, fall back to local ~/.oracle/config.json with browser.modelStrategy: "ignore" or add --browser-model-strategy ignore on the CLI.
• If 本地浏览器: Set use_remote to false. Use the mcp__oracle__consult MCP tool directly.

This choice persists for the entire session. The user does not need to choose again unless they start a new session.

First time the user asks GPT something — choose conversation mode

When session_initialized is false, use AskUserQuestion to ask the user which conversation mode they want:

Question: "你想以什么方式和 ChatGPT 对话？"

Options:

1. 继续对话 — "继续一个已有的 ChatGPT 对话（需要提供对话链接）"
2. 从项目新建对话 — "在某个 GPT 项目中新建对话（需要提供项目链接）"
3. 新建对话 — "直接开始一个全新对话"

Then:

• If 继续对话: Ask user for the conversation URL (like https://chatgpt.com/c/xxx). Set chatgpt_url to that URL. Set session_initialized to true.
• If 从项目新建对话: Ask user for the project URL (like https://chatgpt.com/g/g-p-xxx/project). Set chatgpt_url to that project URL. Set session_initialized to true. After the first call succeeds, update chatgpt_url to the returned tabUrl so subsequent messages stay in the newly created conversation.
• If 新建对话: Leave chatgpt_url empty. Set session_initialized to true. After the first call succeeds, update chatgpt_url to the returned tabUrl.

Sending a question to ChatGPT

Call the MCP tool:

mcp__oracle__consult(
  prompt: "<the user's question>",
  files: [<file paths if the user wants to attach files>]
)

If chatgpt_url is set, you need to tell the user's question in a way that Oracle will use that conversation. Currently Oracle MCP's consult tool doesn't directly accept a chatgpt-url parameter — so if the conversation URL needs to be passed, include it in the prompt context or use the Bash tool to call Oracle CLI directly:

oracle --chatgpt-url "<chatgpt_url>" -p "<user's question>"

If Oracle is configured as a remote server, add --remote-host and --remote-token flags using the values from the user's MCP config (read from ~/.claude/.mcp.json env vars ORACLE_REMOTE_HOST and ORACLE_REMOTE_TOKEN). Only add --browser-model-strategy ignore explicitly when the remote host has not been updated yet; newer oracle serve builds already apply that default on the host side.

This is the reliable method for continuing conversations.

File Uploads

When the user wants to send files to ChatGPT, use the CLI --file flag. Read ORACLE_REMOTE_HOST and ORACLE_REMOTE_TOKEN from the user's ~/.claude/.mcp.json.

# Single file
oracle --remote-host <host:port> --remote-token <token> \
       -p "帮我review这段代码" \
       --file src/main.ts

# Multiple files (glob pattern)
oracle --remote-host <host:port> --remote-token <token> \
       -p "分析这个项目" \
       --file "src/**/*.ts"

# Binary files (images, PDFs, etc.) — must use --browser-attachments always
oracle --remote-host <host:port> --remote-token <token> \
       --browser-attachments always \
       -p "分析这个文件" \
       --file report.pdf

Combine with --chatgpt-url if continuing a conversation. The MCP tool's files parameter also works for text files: mcp__oracle__consult(prompt: "...", files: ["path/to/file"]).

Image Generation / Downloading

When the user asks ChatGPT to draw, generate an image, or download a generated image, use the Oracle CLI directly and save artifacts locally:

oracle --remote-host <host:port> --remote-token <token> \
       --timeout 600 \
       --download-artifacts /Users/claw/oracle-gpt-artifacts \
       -p "请使用 ChatGPT 网页端图片生成能力生成一张图片：<prompt>。生成后请回复图片已生成。"

After the run finishes, report:

• the ChatGPT answer
• the conversation URL
• every Saved browser artifact to ... local path

Important: for remote Oracle, artifact capture runs on the remote oracle serve host and is returned to the local CLI. The remote host must run an Oracle build that supports --download-artifacts; otherwise only the text sandbox:/mnt/data/... link will be available and the local file cannot be downloaded automatically.

After receiving a response

1. Extract the answer text and the conversation URL (tabUrl or from the Conversation URL: line in CLI output).
2. If chatgpt_url was empty or was a project URL, update it to the returned conversation URL for future calls.
3. Show the user:
   • The ChatGPT answer
   • The conversation URL (so they can open it in browser if needed)
4. Tell the user the response time if available.

Subsequent questions

When session_initialized is true and the user asks GPT another question, skip the mode selection — just send the question using the stored chatgpt_url. This maintains conversation continuity in ChatGPT.

Starting a new conversation

If the user says anything like "新对话", "new chat", "开一个新会话", "换个对话", "重新开始", "start fresh":

1. Reset chatgpt_url to empty
2. Reset session_initialized to false
3. On the next question, the mode selection will appear again

Error Handling

• "busy" — Oracle is processing another request. Tell the user: "ChatGPT 正在处理其他请求，30 秒后重试。" Wait 30 seconds, then retry once.
• "Cloudflare challenge" — The ChatGPT login session has expired. Tell the user: "ChatGPT 的登录已过期，需要重新注入 cookies。"
• "ECONNREFUSED" — Oracle serve may have restarted. Tell the user to wait a moment and retry. The service has auto-restart configured.
• Timeout / no response — Oracle can take up to 60 seconds. Always warn the user before sending: "正在向 ChatGPT 发送问题，大约需要 20-60 秒..."

Example Interactions

User: 帮我问一下 GPT，Python 的 GIL 是什么 You: [First time → show mode selection → user picks 新建对话] You: 正在向 ChatGPT 发送问题，大约需要 20-60 秒... You: [Call Oracle, get response] You: ChatGPT 的回答：[answer] 对话链接：https://chatgpt.com/c/xxx

User: 继续问它，那多线程还有意义吗 You: 正在继续对话... [auto-uses stored chatgpt_url] You: ChatGPT 的回答：[answer]

User: 开个新对话 You: [Reset state, next question will show mode selection again]