Hyperresearch: Verifiable Deep-Research Reports for Claude Code
A Claude Code deep-research tool that stores web evidence in a searchable vault and verifies citations individually.
GitHub jordan-gibbs/hyperresearch Updated 2026-09-12 Branch main Stars 2.6K Forks 258
Python Deep research Claude Code SQLite MCP Markdown knowledge base

🧭 Decision Guide

Why trending now: Cannot be determined from the provided materials.

Try it if you

  • You already use Claude Code and Python 3.11–3.13, and need research reports with citation verification.
    The README's “Install” and “Requirements” sections require Python 3.11+ and Claude Code; “Why it wins” says the cite-checker verifies every citation.
  • Your question requires searching OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR, and FRED together.
    The README's “Why it wins” section lists the eight scholarly sources supported by hpr scholar search and deduplication by DOI and title.
  • You want repeated research to accumulate in a Markdown-plus-SQLite vault and resume after interruption.
    The README's “Why it wins” and “Runs: resumable, budgeted, verified” sections say sources enter a persistent vault and run resume continues at the exact step.
  • You need to scale from a 30-minute fast path to a 25K–80K-word dissertation run with 300–450 sources.
    The README's “Why it wins” section specifies the 5-step fast path, 25K–80K-word output, and 300–450 sources.

Skip it if you

  • Your environment is Python 3.14 and you cannot switch to Python 3.11–3.13.
    The README's Install section states that “3.14 [is] not yet supported.”
  • You do not use Claude Code and do not plan to access the vault through MCP or the Web UI.
    The README's Requirements require Claude Code; “Use the vault outside Claude Code” lists only MCP and the local Web UI for external use.
  • You require the DeepResearch-Bench result to have third-party confirmation already.
    The README explicitly states, “Third party validation is pending.”
  • Your task only needs a short one-off answer and does not need the 16-step pipeline, 250+ sources, or cite-check.
    The README's pipeline contains 16 steps; “Why it wins” describes 250+ sources and citation-by-citation verification.

Requirements

  • Python 3.11+; the README specifically says Python 3.14 is not yet supported.
  • Claude Code.
  • The project installation command is: pip install hyperresearch && hyperresearch install.
  • Optional MCP dependency: pip install hyperresearch[mcp].
  • run init --budget 50 can cap estimated API-equivalent spend; the README does not specify the exact model or provider.

First step (verbatim from README)

pip install hyperresearch && hyperresearch install

Watch out

  • A global installation adds about 15 lines to every Claude Code session's system reminder; a per-project installation avoids affecting unrelated sessions.
    The README's Install section describes the approximately 15-line system-reminder cost of hyperresearch install --global.
  • When run init --budget exceeds the budget, the run is blocked instead of allowing spend to grow silently.
    The README's Runs section says crossing the budget blocks the run.
  • For closed papers, the tool first seeks legal full text through Unpaywall, Europe PMC, and CORE; substitutions are disclosed in the note, frontmatter, and CLI output.
    The README's Why it wins section describes the full-text substitution and disclosure locations for paywalled papers.
  • The local Web UI uses a stdlib HTTP server on port 8080, with no build step or JavaScript dependencies.
    The README's Use the vault outside Claude Code section specifies port 8080 and the operating model.

Alternatives

  • Hyperresearch MCP：Use this when you mainly work in Claude Desktop or Cursor rather than the Claude Code command entry point, since those clients can access the same vault through stdio.
    Use the vault outside Claude Code
  • Hyperresearch local web UI：Use this when you only need note browsing, tags, search, and an interactive link graph without connecting Claude Code.
    Use the vault outside Claude Code

Not stated in the README

  • The README does not provide the actual LLM model names, API providers, real per-run cost, or latency.
  • The README provides no publicly reproducible experiment record for the 250+ source claim; it only says third-party DeepResearch-Bench validation is pending.
  • The README does not specify supported operating systems, minimum hardware, or the capacity limit for large SQLite vaults.
  • The README does not specify the accounts, keys, or exact support scope required by Web providers, Authenticated crawling, and the browser lane.
  • The README does not describe encryption, access control, or remote synchronization for research data, API keys, or the local vault.
  • The README does not provide the package's detailed locked dependencies, test coverage, or the content of the 10 recent commits.

💡 Deep Analysis

6
Yes I maintain several Claude Code projects but want Hyperresearch enabled only in one Python 3.13 research directory, without adding system reminders to other sessions. Which installation approach should I use?
For: A developer maintaining multiple Claude Code projects who wants Hyperresearch enabled only in one research project on Python 3.11–3.13

Yes. Use the project-level installation from the README rather than a global installation.

  • In the target directory, the documented flow is pip install hyperresearch && hyperresearch install, followed by /hyperresearch in Claude Code.
  • The project requires Python 3.11–3.13 and does not yet support Python 3.14, so Python 3.13 meets the stated requirement.
  • The README says hyperresearch install --global makes the command available in every Claude Code session and adds roughly 15 lines to each session’s system reminder.
  • Project-level installation keeps unrelated Claude Code sessions clean, matching the requirement to enable it in only one research project.
  • README Install: “cd your-project / pip install hyperresearch && hyperresearch install”
  • README Requirements: “Python 3.11–3.13. (3.14 not yet supported.)”
  • README Install: “`hyperresearch install --global` ... at the cost of ~15 lines in every session's system reminder.”
  • README Install: “Per-project install ... keeps unrelated CC sessions clean.”
cd your-project
pip install hyperresearch && hyperresearch install
Not stated in the README：The README does not say whether project-level installation affects existing Claude Code Skills or system prompts in that directory.；It does not confirm whether vault default paths are fully isolated across different Python virtual environments.
Yes I mainly work in Cursor and want to keep local Markdown and SQLite research material while searching, reading, and updating it through MCP. Does Hyperresearch fit this deployment constraint?
For: A developer using Cursor with MCP who wants to reuse a team's local research material without moving it to a cloud knowledge base

Yes. The README explicitly provides an MCP entry point for Cursor and stores the knowledge base locally as Markdown plus SQLite.

  • After installing hyperresearch[mcp], hyperresearch mcp serves MCP over stdio, allowing Cursor and other MCP clients to use the same vault.
  • The 13 tools include search_notes, read_note, get_backlinks, check_source, create_note, and update_note, so the interface is not read-only.
  • The local web UI uses the standard-library HTTP server and requires no build step or JavaScript dependencies.
  • However, the project insight says access control, team permission isolation, and sensitive-data governance for the local service and vault must be evaluated by the deployer.
  • README: “An MCP server. `pip install hyperresearch[mcp]`, then `hyperresearch mcp` speaks stdio, so Claude Desktop, Cursor, or anything else that speaks MCP can work the same vault.”
  • README: “Thirteen tools: `search_notes` ... `update_note`”
  • README: “No build step and no JavaScript dependencies.”
  • Project insight: access control, team collaboration, and permission isolation for MCP and the local web UI require deployment-side evaluation
pip install hyperresearch[mcp]
Not stated in the README：The README does not say whether the MCP service provides authentication, encrypted transport, or a multi-user permission model.；It does not describe conflict handling when multiple Cursor instances modify the same vault concurrently.
Yes I must deliver an auditable policy report where every cited sentence traces back to a source, while retractions and untraceable numbers are detected. Is Hyperresearch more suitable than a conventional research chatbot?
For: A compliance or strategy analyst producing auditable policy reports and especially concerned with sentence-level citation support, retractions, and numeric evidence

Yes, particularly because it treats citation and process checks as publication gates rather than merely generating a text with references.

  • The README says the cite-checker verifies whether each cited source supports its sentence, while hallucinated quotes and undisclosed retractions block shipping.
  • Pre-publication verification includes quote integrity, retracted citations, numeric consistency, and a per-citation binding audit, directly matching the audit requirements.
  • Sources, notes, links, and the run manifest persist in the vault, allowing review of the underlying material and process, not just the final report.
  • It is not a formal proof system: the project insight says models may still misread sources or overstate causality, and inaccessible pages may leave only abstracts or substitutes.
  • README: “Every citation is verified before the report ships.”
  • README: “quote-integrity ... retracted-citations ... numeric-consistency”
  • README: “plus the cite-check step's per-citation binding audit.”
  • Project insight: the system cannot guarantee factual correctness; inaccessible originals may result in abstracts, substitutes, or incomplete evidence
hyperresearch run verify <tag> -j
Not stated in the README：The README does not state the cite-checker's accuracy in judging source quality, study design, or causal claims.；It does not say whether audit records are tamper-resistant enough for regulatory or legal evidence.
It depends I already use Python 3.13 and Claude Code, and need a 25K–80K-word dissertation based on 300–450 sources. Can Hyperresearch serve as my primary research pipeline?
For: A graduate researcher or scholar using Python 3.13 and Claude Code to write a 25K–80K-word dissertation from 300–450 sources

It depends: it is suitable for retrieval, evidence organization, and draft generation, but not as the sole quality guarantee for a dissertation.

  • The README explicitly supports opt-in dissertation runs producing 25K–80K words across 300–450 sources.
  • Its 16-step workflow covers decomposition, contradiction analysis, synthesis, four critics, citation checking, and pre-publication checks.
  • Research is stored in a searchable Markdown-plus-SQLite vault, while each run has a manifest for resuming after failure.
  • However, the README says the system cannot guarantee factual correctness. Open-access substitutes may differ from publisher versions, and research methods, statistics, and academic ethics remain outside its validation scope.
  • README: “Opt-in dissertation runs write 25K–80K words across chapters, from 300–450 sources.”
  • README: “A tier-adaptive 16-step pipeline”
  • README: “Every source lands in a searchable markdown-plus-SQLite vault”
  • README: “Crashed runs resume”
  • Project insight: citation verification cannot fully judge research design, statistical correctness, or discipline-specific argument quality
cd your-project
pip install hyperresearch && hyperresearch install
Not stated in the README：The README does not say whether generated chapters conform to a target journal, university, or dissertation template.；It does not provide actual runtime, API cost, or quality differences for dissertation mode across model configurations.
Yes I use Claude Code for multi-round web research, must stay below an estimated API-equivalent budget, and may be interrupted by network or process failures. Can Hyperresearch prevent budget overruns and resume from the exact step?
For: An independent researcher who must control Claude Code research costs and frequently encounters network interruptions or process crashes

Yes. Budget caps, isolated runs, and step-level recovery are built into its run mechanism.

  • run init --budget 50 sets an estimated API-equivalent spending cap and blocks the run when the cap is crossed instead of allowing silent cost growth.
  • Each run has an isolated workspace and manifest recording steps, spend, and source yield, so concurrent runs do not collide.
  • hyperresearch run resume -j reports the exact next step and Skill invocation; the README explicitly says crashed runs resume from where they stopped.
  • A budget is only a blocking mechanism, not a completion guarantee. The project insight notes that multiple agents, gap filling, and review can create substantial model calls and network requests.
  • README: “`run init --budget 50` caps estimated API-equivalent spend”
  • README: “Each run owns an isolated workspace ... and a manifest”
  • README: “`run resume` picks up at the exact step where it died.”
  • Project insight: the full workflow can generate substantial model calls, network requests, and local storage requirements
hyperresearch run status -j
Not stated in the README：The README does not state the error range between estimated spend and the actual provider bill.；It does not explain whether failed requests, partial-step retries, or duplicate calls increase recorded spend.
Yes I need papers, books, clinical trials, SEC EDGAR filings, and FRED data in one research list. Can Hyperresearch reduce the work of maintaining multiple search interfaces and deduplication workflows?
For: A policy or financial research analyst who needs one search layer across OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR, and FRED

Yes, because it directly targets unified discovery and deduplication across multiple scholarly and professional source types.

  • The README says hpr scholar search accesses OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR, and FRED through one client layer.
  • Results are deduplicated by DOI and title and tagged by type, including papers, books, trials, and filings.
  • Sources then enter a searchable vault, allowing later sessions to reuse existing material before fetching more.
  • It still depends on external API availability, rate limits, and metadata quality. The README does not promise complete coverage or describe access to complex licensed databases.
  • README: “Eight scholarly sources, one query”
  • README: “returns a single list deduplicated by DOI and title”
  • README: “Nothing is thrown away”
  • Project insight: external API availability, rate limits, and metadata quality may affect retrieval results
pip install hyperresearch[mcp]
Not stated in the README：The README does not list the exact fields, pagination limits, or retry behavior for each provider.；It does not say whether SEC EDGAR or FRED queries support all advanced filters and time-series operations.

✨ Highlights

  • A single run can cover 250+ sources and produce 25K–80K-word dissertation-scale output.
  • The cite-checker verifies citations individually; retractions and hallucinated quotes block shipping.
  • OpenAlex and seven other scholarly sources are unified and deduplicated by hpr scholar search.
  • A Markdown-plus-SQLite vault persists research, and crashed runs continue with run resume.

🔧 Engineering

  • Claude Code invokes a 16-step pipeline to generate adversarially audited reports with full provenance.
  • hpr scholar search queries eight scholarly sources, including OpenAlex, Crossref, and CORE.
  • The Markdown-and-SQLite vault stores sources so later sessions can reuse prior research.
  • Four critics attack drafts in parallel, while the tool-locked patcher can only make surgical edits.

⚠️ Risks

  • The project requires Python 3.11–3.13 and Claude Code; Python 3.14 is not yet supported.
  • The DeepResearch-Bench leadership claim comes from an internal benchmark; third-party validation is pending.
  • The premier profile targets 250+ sources, while run init --budget only caps estimated API-equivalent spend.
  • The full 16 steps include four critics, gap-fetch, and cite-check, which may be excessive for short questions.

👥 For who?

  • Developers who need Claude Code to produce traceable literature reviews, investigation reports, or research.
  • Research teams searching across papers, books, clinical trials, and SEC filings.
  • Users who want to expose the research vault to Claude Desktop or Cursor through MCP.
  • Individual researchers needing a 30-minute fast path or a 25K–80K-word dissertation mode.