The hyperresearch CLI runs the open-source pipeline on your machine

The hyperresearch CLI is a harness that turns Claude Code into a deep research agent running the open-source research pipeline the hosted service is built on. It is MIT licensed, installed with pip, and it keeps everything it reads in a local vault of plain Markdown files you own. The repo is at https://github.com/jordan-gibbs/hyperresearch and the package is at https://pypi.org/project/hyperresearch/. It has about 3,600 stars in September 2026, and the current release is 0.12.0.

Install

pip install hyperresearch && hyperresearch install

Then run /hyperresearch <your question> inside Claude Code.

Python 3.11 to 3.14 is required. Claude Code is required, because Claude Code is the runtime: the pipeline is a set of skill files that Claude Code loads one step at a time. A per-project install keeps unrelated Claude Code sessions clean. A global install with hyperresearch install --global makes the command reachable from any session, at the cost of a few lines in every session’s system reminder.

What it does on your machine

The entry skill pins down the canonical research query and then invokes one step skill per phase, so each step’s procedure enters context only when that step actually runs. That is what stops a long pipeline from silently dropping steps as its context fills.

The research stages are the ones described on the how it works page: a width sweep, a contradiction graph, parallel depth investigators that commit to positions, the reconcile and the corpus critic. The CLI then writes the way the pipeline was first built: angle drafts synthesized into one report, adversarial critics, a patcher tool-locked to read and edit so it physically cannot rewrite the report, and a cite-check where retractions and quotes that are not in the source block shipping. The hosted Deep run writes differently: it plans the report section by section, one writer writes the whole report in one pass from all of the evidence, and every cited claim is then checked against its source; the how it works page describes both.

Notes live as Markdown with YAML frontmatter in research/notes/. The SQLite index is a cache: delete it and hyperresearch sync rebuilds it from the Markdown. You can open the vault in any editor, version it in git, and read your own research without the tool installed.

The scholar client: seven sources, one query

hpr scholar search queries OpenAlex, Crossref, CORE, DOAB, ClinicalTrials.gov, SEC EDGAR and FRED through one client layer and returns a single list, deduplicated by DOI first and then by normalised title and year. A work found by two providers keeps both, with the higher citation count and the longer abstract. Every provider shares one cache and one per-host rate limiter.

hpr scholar search "GLP-1 cardiovascular outcomes" --scope papers -j
hpr scholar search "credit default swaps" -s edgar -s fred -j
hpr scholar sources

Books, registered trials and filings come back alongside papers, each tagged so the pipeline knows which is which. CORE and FRED need free API keys. SEC EDGAR rejects requests without a contact address, and setting one also puts you in the OpenAlex and Crossref polite pools. RePEc is listed by hpr scholar sources as a known gap rather than left silent, because its API has no search function.

When a fetch lands a thin page carrying a DOI, the CLI asks Unpaywall, Europe PMC and CORE in that order for a legal open-access copy and stores that full text in the note instead of the abstract. Every substitution is disclosed in the note body, in the frontmatter, and in the CLI output, including which version it got, because an accepted manuscript is not the version of record.

The vault is also reachable outside Claude Code. hyperresearch mcp speaks stdio MCP, so Claude Desktop, Cursor or anything else that speaks MCP can work the same vault through thirteen tools.

How the CLI and the hosted service relate

The same research stages, a different runtime, and on Deep a different writer.

Vaults move in both directions. A hosted export is a tar.gz in the CLI’s own layout, so unpacking it and running sync gives you a working local vault. The reverse import accepts the same archive or a bare folder of Markdown.

Which one to use

Use the CLI when you have Claude Code, you want the pipeline to cost you nothing beyond what you already pay for models, you want your research as files in a git repo on your own disk, and a run taking up to a couple of hours on your laptop is acceptable. It is also the right choice if you want to read the pipeline itself, change a step, or run it against providers of your own.

Use the hosted service when you need a REST call rather than a terminal session, when you want several runs in flight at once into one shared vault, when a teammate who does not use Claude Code needs to read the report and its receipt, or when you want the vault searchable by a hosted MCP server that your whole team’s agents can reach.

They are not a funnel. The CLI is the reference implementation of the research stages and the skill files are their specification: where the hosted service and a skill file disagree about how a stage researches, the skill file is right and the service is the bug. The hosted Deep writer is the one deliberate difference.