hyperresearch: an agent-driven research vault for Claude Code
Agent-driven research knowledge base. Agents collect, search, and synthesize web research into a persistent, searchable wiki.
At a glance
- What is it?
- hyperresearch installs a 16-step research pipeline into Claude Code and stores every source it reads in a persistent markdown-plus-SQLite vault. The design is unusual and the benchmark claim is self-reported, so the decision turns on whether you want a growing local corpus or a one-shot answer.
- Who is it for?
- Adopt hyperresearch if you already work inside Claude Code, your research repeats across sessions, and you want the corpus to accumulate rather than disappear into a chat log. Do not adopt it if you need a supported, stable release: pyproject.toml classifies it as Development Status 3 - Alpha, and the README's benchmark claim is described there as a forward-looking projection from a stratified pilot with third party validation pending.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
What hyperresearch solves, and for whom
A chat session that answers a research question leaves nothing behind. The sources are gone when the window closes, the same paper gets fetched again next month, and there is no record of which claim came from which document. hyperresearch attacks that specific loss. The README describes a "persistent, searchable vault" in which "every source it reads lands," and states that each session starts smarter than the last because the next run reuses the vault before it fetches anything new.
The audience is narrow and clearly stated: people who already use Claude Code. The package turns Claude Code into what the README calls a deep research agent, driven by a skill invoked as /hyperresearch. If you do not run Claude Code, the pipeline has no host. The intended user is doing recurring, source-heavy work (literature reviews, competitive or policy research, anything where a citation has to be traceable) and is willing to trade wall-clock time for provenance.
The 16-step pipeline and how tiers and gears scale it
The architecture is a router plus step skills. The README states the entry skill "is a thin router": it pins down the canonical research query, then invokes one step skill per phase through Claude Code's Skill tool. Each step's procedure loads into context only when that step runs, which the README frames as the reason a long pipeline does not quietly drop steps as its context rots.
The numbered steps run from Decompose (1) through a width sweep (2), a contradiction graph (3), loci analysis (4), depth investigation (5), cross-locus reconciliation (6), source tensions (7), a corpus critic that asks "What source would overturn this?" (8), an evidence digest of top claims and verbatim quotes (9), a triple draft (10), synthesis (11), four adversarial critics in parallel (12), a targeted gap-fetch wave (13), and a patcher that applies surgical Edit hunks to the draft (14). Step 14.5 is a cite-check that verifies citation-sentence bindings, and steps 15 and 16 are a polish pass and a readability audit. For dissertation runs, an extra step 1.5 partitions atomic items into 4 to 10 chapters, and steps 2 through 10 loop per chapter.
Two levers control scale. Tiers route per query: step 1 auto-classifies light versus full, and dissertation is opt-in only, meaning you have to ask for it in your prompt. Light covers bounded factual queries, surveys and comparisons through steps 1, 2, 10, 15 and 16, quoted at roughly 30 to 40 minutes. Full is the default and runs all 16 steps plus cite-check, quoted at roughly 1.5 to 2.5 hours at full gear. Dissertation targets 300 to 450 sources across 4 to 10 chapters and 25K to 80K words, quoted at roughly 4 to 8 hours. Gears set the source targets and depth budget for the standard pipeline; the README excerpt cuts off mid-sentence at that point, so the gear names and their numeric targets are not documented in what is available here.
Installing hyperresearch and running a first query
The README gives a two-command install from inside your project directory. The first installs the package from PyPI, the second wires the skill into Claude Code.
cd your-project
pip install hyperresearch && hyperresearch installAfter that, the skill is reachable in Claude Code as /hyperresearch. The README's example invocation passes free text:
/hyperresearch <anything>Python version matters. The README supports Python 3.11 through 3.13 and states that 3.14 is not yet supported, suggesting pyenv install 3.13, uv venv -p 3.13, or py -3.13 -m venv .venv as ways to get a compliant interpreter. pyproject.toml encodes the same bound as requires-python = ">=3.11,<3.14".
There is a second install mode. The README says hyperresearch install --global makes /hyperresearch reachable from every Claude Code session anywhere, at the cost of roughly 15 lines in every session's system reminder. Per-project install keeps unrelated sessions clean. That is a real trade-off rather than a preference: the global mode taxes every unrelated session you run.
Optional extras exist for the search and crawling providers. pyproject.toml defines extras named crawl4ai, mcp, parallel, exa, tavily, watch, and all. The mcp extra is upper-bounded at mcp>=1.6,<2 with a comment explaining why: mcp 2.0 removed mcp.server.fastmcp, which hyperresearch/mcp/server.py builds every tool on, so without the bound a fresh install resolves 2.x and hyperresearch mcp dies on import while reporting the extra as missing. The parallel extra is kept separate at mcp>=1.8,<2 so that installing the Hyperresearch MCP server does not force an unrelated SDK upgrade.
The vault, the resume manifest, and what the cite-checker actually blocks
Two mechanisms carry most of the project's weight. The first is the vault: a searchable markdown-plus-SQLite store that the README says nothing is thrown away from. Because the next session reuses it before fetching anything new, the value of the tool depends on repetition. A single one-off query gets you a report and an unused vault.
The second is the cite-check at step 14.5. The README states that a skeptical cite-checker audits whether each cited source actually supports its sentence, and that hallucinated quotes and unacknowledged retractions are hard blocks at the gate. A related audit clusters derivative copies of the same source so that, in the README's phrasing, "five reprints of one press release argue with the weight of one source." The patcher at step 14 is tool-locked to Read and Edit, which the README says means it physically cannot rewrite the report; the same lock applies to the polish subagent at step 15.
Paywalled papers get separate treatment. The README argues that a closed paper normally enters a vault as a 1,500-character abstract that a report then cites as though it had been read. hyperresearch asks Unpaywall, Europe PMC and CORE for a legal open-access copy and stores the full text instead, even when the publisher blocks the fetch. The README states every substitution is disclosed in the note, the frontmatter, and the CLI output. That disclosure is the part worth checking in practice, since a substituted full text is not the same artifact as the version of record.
Runs are resumable. Each run keeps a manifest, and the README gives run resume as the command that picks up at the exact step where a run died.
Where hyperresearch is the wrong tool
The clearest limitation is the host dependency. This is a Claude Code harness, and the pipeline runs through Claude Code's Skill tool and subagents. There is no documented path to run the 16 steps under a different agent runtime or as a standalone command that produces a report without Claude Code in the loop. If your team standardizes on a different assistant, none of the pipeline design transfers.
The second limitation is the benchmark claim itself, and the README is unusually candid about it. The headline says hyperresearch "currently leads the DeepResearch-Bench RACE leaderboard (benchmarked internally)." The caption under the chart says it is a "Forward-looking projection from a stratified pilot against the DeepResearch-Bench leaderboard snapshot" and that "Third party validation is pending." A projection from a pilot is not a leaderboard result. Treat the placement ahead of the named systems in that chart as a hypothesis the maintainers intend to test, not as a measured outcome.
The third is maturity. pyproject.toml carries the classifier "Development Status :: 3 - Alpha" while the package version is 0.11.1. Alpha classifiers and a 0.x version line mean interfaces can move. The README also gives no rollback procedure for hyperresearch install, and the available material does not document an uninstall path, so plan on being able to reconstruct what the installer touched before you run it.
Finally, cost. The README quotes roughly 30 to 40 minutes for a light run, 1.5 to 2.5 hours for a full run at full gear, and 4 to 8 hours for a dissertation run. A question you would otherwise answer in a chat turn is not a good fit for a pipeline with those numbers. The README also notes the premier scale profile targets 100 to 130 sources in the width sweep alone, with citation chasing and gap-fill more than doubling what lands in the corpus. That is a lot of fetching for a question that does not need it.
How it differs from qualitative analysis software
People searching for hyperresearch alongside MAXQDA or asking about hyperresearch vs nvivo are comparing two different jobs. MAXQDA and NVivo are qualitative data analysis packages: you bring your own corpus (interview transcripts, field notes, documents you already collected), and the software gives you coding, memoing, retrieval and matrix queries over that fixed set. The corpus is an input.
hyperresearch inverts that. The corpus is an output. The pipeline fetches from the web, chases citations, fills gaps identified by a critic step, and writes what it finds into the vault. Its analysis steps (the contradiction graph, loci analysis, source tensions, the evidence digest) operate on material the tool itself assembled, and the deliverable is a report with citation-sentence verification rather than a coded dataset.
The practical consequence is that hyperresearch does not replace a QDA package for interview-based work, because nothing in the documented pipeline ingests your transcripts and codes them. Conversely, MAXQDA and NVivo do not go and find sources for you. If your problem is "I have 40 interviews and need to code them," hyperresearch is the wrong tool. If your problem is "I need 250 sources on a question and a record of which one supports which sentence," the QDA packages are not addressing it. There is also a search-volume artifact worth naming: hyperresearch terraria and hyperresearch download are unrelated to this project, and the README describes no game, no binary download, and no installer beyond pip.
Licence, maintenance and the cost of upgrading
The licence is MIT, declared in pyproject.toml as license = "MIT" and listed in the repository root as LICENSE. MIT is permissive: it allows commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That is the general shape of the licence, not legal advice, and anything you build on top of the vault or the pipeline is your own compliance question.
On maintenance, the evidence is concrete rather than promotional. The last push to the default branch was on 2026-09-12. The repository is not archived. Three releases landed on 2026-09-11: v0.10.1, v0.11.0, and v0.11.1, with the version in pyproject.toml matching v0.11.1. Three releases in one day suggests rapid iteration, which cuts both ways: fixes arrive quickly, and so do breaking changes in a 0.x line classified as alpha.
Upgrade cost is mostly the dependency bounds. The mcp extra is pinned below 2.0 for a stated reason (the removal of mcp.server.fastmcp), and the parallel extra is pinned at a different floor. If you install with the all extra, you inherit those pins. The README's install flow is pip-based and does not describe a lockfile or a pinned set of transitive dependencies, so an upgrade can move more than the package version. CHANGELOG.md and docs/ exist in the repository root, which is where the upgrade notes would live; the README itself does not summarize breaking changes between 0.10.x and 0.11.x.
Editorial conclusion
Adopt hyperresearch if you already work inside Claude Code, your research repeats across sessions, and you want the corpus to accumulate rather than disappear into a chat log. Do not adopt it if you need a supported, stable release: pyproject.toml classifies it as Development Status 3 - Alpha, and the README's benchmark claim is described there as a forward-looking projection from a stratified pilot with third party validation pending. Before committing, verify three things: that your Python is 3.11 through 3.13, since 3.14 is not yet supported; whether you want the per-project install or hyperresearch install --global, which the README says adds roughly 15 lines to every Claude Code session's system reminder; and what the premier scale profile actually costs you in wall-clock time, because the README quotes 1.5 to 2.5 hours at full gear and 4 to 8 hours for a dissertation run.
Frequently asked questions
How do you use hyperresearch?
Install it from your project directory with pip install hyperresearch && hyperresearch install, then invoke the skill in Claude Code with /hyperresearch followed by your question. The README notes that a bounded factual query auto-routes to a 5-step fast path, while the default full tier runs all 16 steps plus cite-check.
Does hyperresearch require Python 3.14?
No. The README supports Python 3.11 through 3.13 and states that 3.14 is not yet supported, and pyproject.toml encodes the same bound as requires-python = ">=3.11,<3.14". The README suggests pyenv install 3.13, uv venv -p 3.13, or py -3.13 -m venv .venv if you need a compliant interpreter.
What happens if a hyperresearch run crashes partway through?
Each run keeps a manifest, and the README gives run resume as the command that picks up at the exact step where the run died. That avoids restarting the whole pipeline after a failure late in a multi-hour run.
Does hyperresearch lead the DeepResearch-Bench leaderboard?
The README headline claims it currently leads the RACE leaderboard benchmarked internally, but the chart caption describes that placement as a forward-looking projection from a stratified pilot against a leaderboard snapshot, with third party validation pending. Read the chart as a stated intention to validate rather than a confirmed result.
Community notes