Hyperresearch is a deep research harness for Claude Code — a Python package that drives the agent through a long, structured research pipeline and files every source it reads into a persistent, searchable vault. It is aimed at people who already work inside Claude Code and need research output whose claims can be traced back to something real: engineers doing technology due diligence, analysts, and writers who have to cite what they assert. It is not a hosted service. You install it into your own setup, run it from there, and the corpus it builds stays on your side.
What it does
The pitch is breadth before prose: Hyperresearch reads a large number of sources before it writes a sentence. The README says a single run can take in 250 or more sources, and breaks that down — the premier scale profile targets 100 to 130 sources in the width sweep alone, and citation chasing plus gap-fill fetches more than double what eventually lands in the corpus.
The other half is verification. Before a report ships, a skeptical cite-checker audits whether each cited source actually supports the sentence attached to it, which is aimed squarely at the failure mode where a paywalled paper gets cited from a one-page summary the agent never got past. The finished report carries full source provenance.
Everything read is kept. The vault is persistent and searchable, so a second run on a neighbouring topic starts from the material the first one already gathered rather than from an empty page.
How it works
One prompt goes in and a tier-adaptive 16-step pipeline runs it. Tier-adaptive means the amount of work scales with the profile you choose rather than every question getting the same treatment; the premier profile is the heavy end.
Within the pipeline, the stages named in the materials are a width sweep that casts broadly across the topic, citation chasing that follows references out of the sources already collected, and gap-fill fetches that go after what the first two passes missed. Drafts are then attacked: the channel's own description counts four adversarial critics going at every draft, plus the cite-checker pass, before anything reaches you.
The README claims the harness currently leads the DeepResearch-Bench RACE leaderboard. Read that claim with the repository's own footnote in hand: it was benchmarked internally, the chart is described as a forward-looking projection from a stratified pilot against a leaderboard snapshot, and third-party validation is pending. That is the author being unusually candid, and it is the honest state of the evidence.
Getting started
Hyperresearch is on PyPI and needs Python 3.11 or newer. The installation is a single pip command run inside Claude Code, after which the tool is available as a slash command: type /hyperresearch followed by your topic and the pipeline takes over. There is no separate account, server or dashboard to stand up.
The project is MIT licensed, written in Python, and has drawn roughly 2,400 stars and 248 forks. It was created in April 2026 and was still being pushed to in September 2026, so it is young and actively moving.
When to use it / when not
Reach for it when:
- the question is broad enough that a handful of sources would obviously misrepresent it
- the citations have to survive scrutiny from someone who will click them
- you expect to research the same domain repeatedly, so the vault compounds instead of being thrown away
Skip it when:
- you want a quick answer — a run designed around hundreds of sources costs time and tokens in proportion
- you do not work in Claude Code, since that is the only interface described
- you need an independently verified benchmark position rather than a self-reported one
Alternatives
The README positions Hyperresearch against a field of deep research systems: Grep Deep Research, Cellcog Max, nvidia-aiq, Gemini Deep Research and OpenAI Deep Research. The practical difference is not the leaderboard ordering but the shape of the product. Those are largely hosted services you send a question to. Hyperresearch runs inside an agent you already control, keeps every source it touched, and lets you inspect the pipeline that produced the report.
Anyone whose job involves writing things other people will fact-check should look at this seriously, provided Claude Code is already part of their workflow. The engineering ideas on display — reading wide before drafting, auditing every citation against its sentence, and treating collected sources as an asset that persists rather than context to be discarded — address the exact ways AI research output usually embarrasses the person who relied on it. Treat the leaderboard claim as unproven until someone outside the project checks it, and judge the tool instead on whether the reports it hands you hold up when you open the links. On that measure it is built to be checked, which is more than most tools in this category offer.