he core challenge in modern AI-driven research is the inefficient consumption of tokens when agents interact with the web. Claude's native web tool often processes excessive formatting code, which consumes tokens and forces the model to rely on summaries rather than the full source text. This inefficiency is not just a cost issue; it limits the model's ability to ground its answers in the actual content of the page, as the context window becomes cluttered with navigation menus and boilerplate code.
The proposed solution is the implementation of 'Only CLI' (OC), a tool designed to provide a structured, numbered outline of web pages. By enabling agents to selectively read only the necessary parts of a page, OC reduces token usage by approximately 21% while ensuring the model accesses the actual text rather than a summary. This approach is particularly effective for deep research tasks where accuracy and context are paramount. The tool also provides a mechanism to bypass site-specific blocks, such as those found on npm or Stack Overflow, which often reject standard automated requests.
However, the utility of OC is context-dependent. For websites that provide clean 'llm.txt' files, the native web tool remains highly efficient and sometimes superior to using an external scraper. Therefore, the optimal strategy involves configuring agents to use OC by default while allowing them to leverage native capabilities when a site is specifically optimized for AI ingestion. The setup process involves installing the tool via GitHub and updating agent configuration files, such as 'claude.md', to prioritize OC for web interactions.
Ultimately, the shift toward agentic research requires a more surgical approach to data ingestion. As agents become more autonomous, managing the cost and quality of their web interactions will become a critical differentiator for developers. By adopting tools like OC, users can maintain high-quality research workflows without incurring the prohibitive costs associated with inefficient token consumption.





