What is hyperresearch?
hyperresearch, by Jordan Gibbs, turns Claude Code (and, more recently, OpenAI Codex) into a deep-research agent. One prompt starts a pipeline of up to 16 steps: decomposing the question, parallel source-fetching waves, contradiction mapping, three drafts, four adversarial critics, a tool-locked patcher that may only make surgical edits, and a citation check before the report ships. Every source it reads is saved in a Markdown-plus-SQLite vault that later runs search first.
It installs with pip install hyperresearch and hyperresearch install, or as a Claude Code or Codex plugin. Scholarly search covers OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR and FRED, and it looks for legal open-access copies of paywalled papers. An MCP server exposes the vault to other clients.
The repo is MIT-licensed with about 4,200 GitHub stars; the PyPI package is at 0.13.0 (27 October 2026). The author also sells a hosted version at hyperresearch.ai, priced per report.
Is hyperresearch worth it?
hyperresearch is the most rigorous open deep-research harness we have read about, and it earns that on mechanism rather than marketing. Commercial research modes hand back a fluent report and ask you to trust the citations. This pipeline is built to distrust itself: quoted text must appear verbatim in a saved source or the report cannot ship, citing a retracted paper without saying so is a hard error, an independence check counts five reprints of one press release as one source, and the final editing passes are locked to Read and Edit so they cannot quietly rewrite the draft. Fetched pages are wrapped as untrusted data, a sensible defence against prompt injection from hundreds of pages the agent did not choose.
The headline claim needs a caveat. The README puts it first on the DRACO deep-research benchmark at 77.7, ahead of Claude Research at 73.0 and ChatGPT deep research at 55.2, but that is a 10-task run published by the project itself, and the Codex port has not been benchmarked at all. Treat it as a sign the pipeline is serious, not as a settled ranking.
The cost is time and tokens. A light run takes about 30 to 40 minutes, the default full tier about 1.5 to 2.5 hours, and the premier gear 3 to 5. Each run spawns many subagents, so on a Claude subscription it eats into limits and on an API key it costs real money; run init --budget caps estimated spend and blocks the run rather than letting it overrun. The vault is the long-term payoff: sources accumulate, and the next question in the same field starts from what is already there.
This is a young project, created in April 2026 and at version 0.13, with one principal author and an experimental dissertation tier whose targets are not yet wired in. The README is explicit that the lint gate catches structural failures, not factual ones. Against the hosted version at $9 or $49 a report, running it yourself trades money for setup and supervision; against a chat product's research mode, it trades speed for an audit trail you can actually check.
Where it earns its place
An equity analyst at a small fund covering restaurant software
Produce a competitor analysis built from SEC filings with every claim traceable
Run /hyperresearch in Claude Code with a budget cap; it pulls EDGAR filings into the vault, drafts, critiques and cite-checks. Expect 1.5 to 2.5 hours on the full tier, or $49 for a Deep report on the hosted version.
A PhD student writing a literature review in health policy
Gather papers and trials, including paywalled ones, without citing abstracts as if read
Scholarly search hits OpenAlex, Crossref, CORE and ClinicalTrials.gov, and open-access recovery fetches legal full texts with the substitution disclosed. The vault keeps every source for the next chapter.
A platform engineer giving a team's agents shared research memory
Let Cursor or Claude Desktop query sources gathered by earlier runs
pip install hyperresearch[mcp] and run hyperresearch mcp; 13 tools including search_notes, read_note and fetch_url work over the same vault. Free apart from model usage.
Before you commit
- Setup
- Python 3.11 to 3.14, then pip install hyperresearch and hyperresearch install in the project (--target codex for Codex), or install the Claude Code plugin. Codex sessions need workspace write and network access enabled.
- Integrations
- Claude Code and OpenAI Codex; web providers builtin, crawl4ai, Exa, Tavily, Parallel and Serply; optional Voyage or OpenAI embeddings; an MCP server with 13 tools for Claude Desktop, Cursor and others.
- Data and privacy
- Runs locally and stores sources as Markdown plus SQLite in your project; page content goes to your agent's model provider and any search API you configure. Fetched text is fenced as untrusted input. MIT licence.
- Cost at scale
- No licence cost; each run consumes many subagent calls, so heavy use means a higher Claude plan or API spend, capped per run with --budget. The hosted version charges $9 per Light and $49 per Deep report.
- Limits
- Cannot read paywalls you are not logged into unless a legal open-access copy exists; lint checks structure, not truth; the dissertation tier is experimental; on Codex there is no browser lane and tool locks are instructions only.
What’s good
- Verbatim-quote, retraction and citation checks block a report from shipping
- Persistent local vault reused by later runs and exposed over MCP
- Eight scholarly and filing sources searched in one query
- Spend cap and resumable runs for multi-hour jobs
What’s not
- Full runs take 1.5 to 2.5 hours and many subagent calls
- Benchmark lead is self-reported on 10 tasks
- Codex port is new, unbenchmarked and lacks tool locks
- Single principal maintainer; fast-moving 0.x releases
Best for
Analysts and researchers already on Claude Code who need cited reports they can audit and are willing to wait hours for them.
hyperresearch alternatives
hyperresearch: common questions
- Is hyperresearch worth it?
- Deep research with real verification gates instead of a polished summary, paid for in hours of agent time and a pipeline you have to trust one maintainer to keep working. We rate it 3.9 out of 5.
- How much does hyperresearch cost?
- Free and open source (MIT); you pay for the Claude Code or Codex usage it consumes. Hosted version: pay as you go, $9 per Light report and $49 per Deep report; Enterprise on request.
- What are the downsides of hyperresearch?
- Full runs take 1.5 to 2.5 hours and many subagent calls; benchmark lead is self-reported on 10 tasks; codex port is new, unbenchmarked and lacks tool locks; single principal maintainer; fast-moving 0.x releases.
- Who is hyperresearch best for?
- Analysts and researchers already on Claude Code who need cited reports they can audit and are willing to wait hours for them.
- How much does hyperresearch cost for a team or in production?
- No licence cost; each run consumes many subagent calls, so heavy use means a higher Claude plan or API spend, capped per run with --budget. The hosted version charges $9 per Light and $49 per Deep report.
- Is hyperresearch safe for company data?
- Runs locally and stores sources as Markdown plus SQLite in your project; page content goes to your agent's model provider and any search API you configure. Fetched text is fenced as untrusted input. MIT licence.
- What are the best hyperresearch alternatives?
- The closest alternatives we have reviewed are notebooklm-py, Agent Reach and Crawl4AI.
Answers checked Oct 31, 2026.