jordan-gibbs/hyperresearch

Hyperresearch is a deep research harness that turns coding agents like Claude Code and OpenAI Codex into powerful research assistants. Instead of a simple Q&A, it runs a structured 16-step pipeline that gathers hundreds of sources, verifies every citation, and produces a fully referenced report. It's designed for serious research tasks—competitor analysis, academic literature reviews, market research—and claims to outperform dedicated research tools like ChatGPT and Gemini on the DRACO benchmark.

How it works:

• You give it a question (e.g., /hyperresearch "Compare Toast vs Square vs Clover for restaurants").
• It decomposes the question into atomic sub-questions, then runs a multi-phase pipeline: broad web searches, deep dives into specific topics, contradiction detection, adversarial critique, and finally a synthesized report.
• Every source it reads is stored in a persistent, searchable vault (Markdown + SQLite). Future sessions reuse this vault, so research compounds over time.
• Citations are verified before the report ships—a skeptical checker ensures each claim is actually supported by the cited source, blocking hallucinated quotes.
• The pipeline is modular: it uses multiple specialized AI subagents (fetchers, critics, patchers) with different models (Sonnet, Opus) for different roles.

Key features:

• Scale: Can pull 250+ sources in a single run; optional 'premier' profile targets 100–130 sources.
• Adversarial review: Four critics attack the draft in parallel; a tool-locked patcher can only make surgical edits, not rewrite the whole report.
• Scholarly sources: Integrates OpenAlex, Crossref, CORE, SEC EDGAR, FRED, and more—so it can pull papers, filings, and data.
• Paywalled papers: Automatically finds legal open-access copies via Unpaywall, Europe PMC, and CORE.
• Resume support: If a run crashes, it can resume from the exact step.
• Flexible output: Supports different 'registers' (teach, survey, analyze, advocate) and inference depth levels.

Installation:

• Requires Python 3.11+ and either Claude Code or OpenAI Codex.
• Install via pip install hyperresearch, then run hyperresearch install to set it up in your project.
• Also available as a hosted service at hyperresearch.ai.

Why it stands out: It's not just a wrapper—it's a carefully engineered pipeline with explicit quality controls (cite-checking, independence audits, adversarial critics) and a persistent memory (vault). The benchmark results suggest it's genuinely effective for deep research tasks.