Giving GPT-3.5 a search engine, a calculator, and a voice

What it does Ex-ChatGPT is a Python web service that sits in front of the OpenAI API and orchestrates external tools. It prompts GPT-3.5 to generate API calls, then routes them to Google Search, WolframAlpha, or WikiMedia, feeds the results back, and asks the model to summarize. The project also includes WebChatGPTEnhance, a browser extension that patches the ChatGPT web UI with similar external API support. A handful of modes—Web, WebDirect, Detail, and Keyword—let you trade speed for thoroughness, or skip API keys entirely in Keyword mode (though the README warns that accuracy suffers).

The interesting bit The system treats the LLM as both planner and executor: it generates the API request, validates the first response against the original question, and decides whether to supplement with more calls. That loop is what the project calls its “ToolFormer” approach, and it goes beyond a single search-then-summarize pass. The project also bundles terminal-style shortcuts, prompt auto-completion with pinyin search, and Azure-powered voice chat—features that suggest the author actually uses this as a daily driver rather than a weekend demo.

Key highlights

• Multiple search modes: Web (iterative verification), WebDirect (faster single-pass), Detail (extra passes for missing info), and Keyword (DuckDuckGo, no extra keys).
• Voice conversation via Microsoft Azure, with TTS and speech recognition in multiple languages.
• API key pooling and cooldown logic, including automatic cooling for Google 403 errors.
• Chat history auto-compression, redundant backup, and a visible token counter to keep costs honest.
• Browser extension (WebChatGPTEnhance) that adds external API calls and custom prompts to the official ChatGPT web interface.

Caveats

• The README is primarily in Chinese; English documentation exists but appears secondary.
• OAuth2.0 multi-user auth and some other features live on the webTest branch, not main.
• Keyword Mode explicitly sacrifices accuracy to avoid API keys.

Verdict Worth a look if you want a self-hosted, Bing-like augmentation layer for GPT-3.5 and do not mind managing a small zoo of API keys. Skip it if you need a polished, mobile-ready product or a fully open-source stack with no external dependencies.