HyperResearch is a simple Claude Code skill harness that outperforms every deep research framework.
HyperResearch surpasses OpenAI, Google, and NVIDIA's offerings in the agentic search space based on DeepResearch Bench. It's open-source, installable with a single command, and uses your CC subscription, so you don't have to pay for OpenAI or Gemini Pro.
It uses a 16-step pipeline that creates a searchable, persistent knowledge store during each session that can be built upon in later searches. I designed it to align with the original user prompt as closely as possible, while incorporating built-in fact-checking, adversarial review, and breadth and depth-investigating capabilities.
This is a generalized framework; you can use it for any large-scale research task, from developing a trading strategy for a specific stock to analyzing competitor products to understanding the current state of the art in LLM architecture.
It uses crawl4ai (an open-source LLM search tool) to capture a wider breadth of information than the standard websearch tool is capable of. You can also configure authenticated sessions, meaning that LinkedIn, Twitter, etc., are now fair game for agentic search.
I got really tired of bloated knowledge bases with folders of unstructured MD files, so, inspired by the "Zettelkasten" and Karpathy's recent tweet, I created an open-source agent-forward CLI tool.
The goal here is to create expansive repositories of information that are fully linked, tagged, and quality assured. This way, when you tell an agent to "find this info" you can actually be certain it'll go to the right place.
Docs are auto-injected into your agents/claude.md files so you can plug and play.
I was able to do a massive deep research session (7+ hours) where Claude Code with web search indexed all major LLM architecture advancements in 2025/2026, and auto-organized into a beautiful wiki spanning over 1 million words and 600 notes.
Built this for fun; uses OpenRouter for easy access to a variety of models.
The joy of seeing these AIs flail and argue is unmatched and reassuring; they're unbelievably bad at the game. However, it is an interesting view into their ability to deceive.
This costs a ton to run, and I don't currently have the funds to properly benchmark across frontier models, so I'd love it if we could source some data from the community!
HyperResearch surpasses OpenAI, Google, and NVIDIA's offerings in the agentic search space based on DeepResearch Bench. It's open-source, installable with a single command, and uses your CC subscription, so you don't have to pay for OpenAI or Gemini Pro.
It uses a 16-step pipeline that creates a searchable, persistent knowledge store during each session that can be built upon in later searches. I designed it to align with the original user prompt as closely as possible, while incorporating built-in fact-checking, adversarial review, and breadth and depth-investigating capabilities.
This is a generalized framework; you can use it for any large-scale research task, from developing a trading strategy for a specific stock to analyzing competitor products to understanding the current state of the art in LLM architecture.
It uses crawl4ai (an open-source LLM search tool) to capture a wider breadth of information than the standard websearch tool is capable of. You can also configure authenticated sessions, meaning that LinkedIn, Twitter, etc., are now fair game for agentic search.
If anyone wants to port it to Codex, be my guest!