ChatGPT Delegate

ChatGPT Delegate lets Codex hand off substantive work to ChatGPT Web/App while Codex only orchestrates.

It is useful for deep research, PRD drafting, design thinking, comparison writing, community best-practice scans, and quick current-information lookups where you want ChatGPT to spend the thinking/search effort and save the final Markdown back to the local filesystem.

Project page: https://longbiaochen.github.io/chatgpt-delegate/

Boundary

• Codex prepares the task, starts the local MCP connector, opens ChatGPT, sends the prompt, and waits for a saved file.
• ChatGPT Web/App performs the research, thinking, writing, or lookup.
• The connector only saves ChatGPT's Markdown result under artifacts/chatgpt-delegation/.
• The connector does not call Codex, the OpenAI API, private ChatGPT APIs, session tokens, cookies, or reverse-engineered endpoints.
• By default, Codex should return only status and a file link, not paste the full result back into chat.

Install

git clone https://github.com/longbiaochen/chatgpt-delegate.git
cd chatgpt-delegate
python3 -m venv .venv
./.venv/bin/pip install -e ".[chatgpt,test]"

CLI

Prepare a task:

chatgpt-delegate prepare "调研一下 MCP 社区最佳实践" --kind auto

Start the MCP server:

chatgpt-delegate serve --host 127.0.0.1 --port 8000 --public-host <public-host>

Check status:

chatgpt-delegate status <task-id>

Read metadata without loading the Markdown body:

chatgpt-delegate read <task-id> --metadata-only

List recent results:

chatgpt-delegate list

The old console entry remains available as a compatibility alias:

chatgpt-research-connector --help

Profiles

--kind auto routes common Chinese trigger words to one of three prompt profiles:

Intent	Kind	Profile	Behavior
调研, 深度研究, 社区最佳实践, 研究一下	research	research_pro	Thinking pro style, web research, long Markdown report, source list
设计, 构思, 思考, 比较, 撰写, 方案, PRD	think	thinking_high	Thinking high style, structured proposal/PRD/comparison/writing
查一下, 找一下, 确认, 最新	lookup	instant_web	Instant web style, short answer with evidence and links

Data Flow

1. In Codex, ask: 让 ChatGPT 调研一下 xxx.

2. Codex runs chatgpt-delegate prepare ... and gets task_id plus prompt.md.

3. Codex exposes the local MCP server through an HTTPS tunnel and configures ChatGPT Developer connector URL as https://<public-host>/mcp.

4. Codex opens ChatGPT Web/App, selects the connector, and sends prompt.md.

5. ChatGPT performs the work and calls:

   save_task_result(task_id, title, markdown, summary=None, overwrite=True)
   

6. The connector writes:

   artifacts/chatgpt-delegation/<task_id>/task.json
   artifacts/chatgpt-delegation/<task_id>/prompt.md
   artifacts/chatgpt-delegation/<task_id>/result.md
   artifacts/chatgpt-delegation/<task_id>/result.json
   

7. Codex polls chatgpt-delegate status <task-id> and returns the saved file path.

MCP Tools

• save_task_result(task_id, title, markdown, summary=None, overwrite=True)
• list_results(limit=20)
• read_result(task_id)
• connector_status()
• save_markdown_report(...) for older prompts

connector_status() must report:

{
  "codex_execution": false,
  "openai_api": false,
  "private_chatgpt_api": false
}

Codex Skill

This repo includes a reusable skill at:

skills/chatgpt-delegate/SKILL.md

Install or copy it into your Codex skills directory, then trigger it with requests such as:

让 ChatGPT 调研一下 100M 上下文大模型的社区方法
用 ChatGPT 想一下这个 PRD
请 ChatGPT 查一下最新进展

Tests

python3 -m pytest -q
git diff --check