Turn AI coding tools into deep researchers and build a persistent personal knowledge base
Throw a single topic at it, and it scours the internet to gather hundreds of sources, verifies their content, and writes a reliable report. This tool transforms AI coding tools like Claude Code or OpenAI Codex from simple chat partners into researchers that independently find, read, and critically review materials. Collected data is not discarded but remains in the repository, allowing subsequent research to dig deeper based on already accumulated knowledge.
Top of the DRACO benchmark and Codex support
The tool ranked first in the DRACO deep research benchmark, outperforming Claude Research, ChatGPT, Gemini, and Perplexity. Recent updates added support for OpenAI Codex to function identically. Codex users can install the entire pipeline with a single command without a separate installation process, sharing the same repository and verification standards as Claude Code. A hosted version is also available, allowing users to access the same features without installing the tool locally.
16-stage pipeline and adversarial verification
User-input questions go through a 16-stage automated process.
- Large-scale data collection A single run collects more than 250 sources. Using the 'premier' profile targets 100 to 130 sources in the broad search phase alone, with additional collection resulting in more than double that amount entering the actual repository.
- Independence audit Cases where multiple media outlets reprint the same press release or article are treated as a single source. Even if five media outlets publish the same content, if the underlying original is one, only one weight is assigned.
- Paywalled paper resolution For paywalled papers, it does not just read the abstract but finds a legal open-access copy and stores the full text. This process clearly records which source was used as a substitute.
Permanent repository and reusable knowledge
Most research tools discard remaining materials after exporting a report. This tool stores all read materials in a repository (vault) composed of Markdown files and a SQLite database. This repository is not just a bundle of files but serves as a base to search first before bringing in new materials in the next session. Materials can be filtered by DOI, journal, citation count, and retraction status, and PDF files are automatically converted to text and stored. The repository is in Markdown format, so it can be opened in any editor and version-controlled with Git.
Installation and execution environment
It supports Python 3.11 through 3.14. If using Claude Code, run pip install hyperresearch followed by hyperresearch install. If using OpenAI Codex, run hyperresearch install . --target codex and enable file write permissions and network access permissions when starting a session. Since Codex blocks network and file writing by default, research will not proceed if these permissions are not granted. Using the global installation option makes commands available in all sessions, but project-specific installation is recommended.
Pre-use considerations
This tool only works in environments where an AI coding agent (Claude Code or Codex) is installed. If you do not want to run it locally, you must use the hosted version. Codex users must change sandbox settings when starting a session; since this relaxes security settings, it should only be used in trusted projects. Additionally, deep research requires significant time and API costs. The 'light' tier takes approximately 30 to 40 minutes, the 'full' tier takes 1.5 to 2.5 hours, and the experimental 'dissertation' tier can take 4 to 8 hours. It provides a feature to set a budget cap to stop execution if expected costs are exceeded.
