An open-source research platform (MIT licensed) published by Jordan Gibbs. Installed into Claude Code, typing /hyperresearch <your question> runs a 16-step pipeline that decomposes the query, gathers information from multiple perspectives, surfaces contradictions, drills into open questions, and puts the draft through review by several critic agents before producing a single report. What sets it apart is how strictly it treats sources: before anything ships, it mechanically checks whether quoted sentences actually appear in the source material and whether any cited paper has been retracted. Every source it reads accumulates in a “vault” of Markdown plus SQLite, so the next session searches what you already have before fetching anything new. It ships as a Python package, and the vault can also be used through a CLI or an MCP server.
Key Features
- A 16-step research pipeline: query decomposition → broad information gathering → contradiction graph → identifying and investigating key points → multiple parallel drafts → synthesis → review by four adversarial critics → patching → citation verification → polish. Each step’s procedure loads into context only when that step actually runs, so instructions are less likely to get dropped partway through a long pipeline
- Three tiers matched to the weight of the question: bounded questions such as fact checks and comparisons are routed automatically to a 5-step fast path (
light), while questions that require argument run the full 16 steps (full). Thedissertationtier, which writes chaptered reports tens of thousands of words long, runs only when you ask for it explicitly - Citation verification and retraction detection: before output, it checks whether each cited sentence is genuinely supported by its source and whether that source has been retracted. If a quoted span does not appear verbatim in a vault note, the report does not ship
- An independence audit so syndication doesn’t count as consensus: five articles reprinting the same press release are clustered into one, counted as a single piece of support rather than five. Every source carries a quality score built from citation counts, source type, retraction flags, and more
- A persistent, reusable vault: fetched pages are stored as Markdown with YAML frontmatter, with SQLite holding the full-text index. The index can be rebuilt from the Markdown, so your research remains readable even without the tool. Full-text search, link graphs, duplicate detection, and export are all available from the CLI
- Automatic open-access recovery: when a paywalled paper yields only an abstract, it asks Unpaywall and Europe PMC for a legal open-access copy and stores the full text instead. Any such substitution is stated both in a banner at the top of the note and in the frontmatter
- Usable outside Claude Code:
hyperresearch mcpstarts an MCP server, letting Claude Desktop, Cursor, or any MCP-capable client search and add to the same vault. A local web UI (hyperresearch serve) is included as well
Pricing
| Plan | Price | What you get |
|---|---|---|
| The software itself (MIT license) | $0 | All features. Install from PyPI with pip install hyperresearch |
| What you need separately | Depends on usage | An Anthropic plan that runs Claude Code (subscription or API usage). Consumption scales with tier, gear, and corpus size |
Pricing is accurate as of August 2026. The software itself is free, but the model usage required to run it is billed separately. Check the official repository for the latest information.
Pros & Cons
✅ Pros
- Citation verification, retraction detection, and the independence audit are built into the pipeline, reducing the manual work of confirming source reliability
- Because the sources you read accumulate locally, researching the same field gets faster and cheaper on the second pass and beyond
- Notes are stored as plain Markdown, so they work directly with existing tools such as Obsidian, and with Git
- An interrupted run can be picked up mid-pipeline with
run resume, so long research sessions are less likely to be wasted - You can set a spending cap, which halts the run before costs balloon unexpectedly
- Fetched page bodies are treated as data rather than instructions, a deliberate guard against prompt injection from hostile pages
⚠️ Cons
- It assumes Claude Code and a Python 3.11–3.13 environment; the setup is a hurdle if you are not comfortable on the command line
- A single run takes a long time — several hours at the
fulltier, and close to half a day fordissertation - Model usage is on you, and costs rise as the number of sources grows
- It is designed around Anthropic models, and no port to other providers is currently available
- The verification gates guarantee structural soundness, not the factual accuracy of the content itself
- Pages behind logins and CAPTCHAs are never bypassed automatically (by design), so some manual work remains
Comparison with Similar Services
| Criteria | Hyperresearch | Gemini Deep Research | ChatGPT Deep Research | GPT Researcher |
|---|---|---|---|---|
| Delivery | Open source (MIT) / runs in your own environment | A feature inside the Gemini app | A feature inside ChatGPT | Open source / runs in your own environment |
| Runtime | Claude Code + Anthropic models | Google’s models | OpenAI’s models | Multiple models selectable |
| Accumulated research assets | Persistent Markdown + SQLite vault | Kept as conversation history | Kept as conversation history | Mostly per-run report output |
| Source verification | Built-in citation verification, retraction detection, independence audit | Source links provided | Source links provided | Source links provided |
| Time per run | 30 minutes to several hours (tier dependent) | Minutes to tens of minutes | Minutes to tens of minutes | Around a few minutes |
| Ease of setup | Requires command-line work | Ready to use | Ready to use | Requires setup |
Who Is It For
- Researchers and analysts working with academic papers and primary sources, where citation accuracy determines the value of the output
- People who research the same field continuously and want the material they read to accumulate in a reusable form
- Anyone spending significant time verifying the sources behind generated reports after the fact
- Developers already using Claude Code who want to hand the research stage to an agent as well
- People who want their data local, or their research assets held in Markdown on their own machine
Summary
Where existing “deep research” features return a report in a few minutes, Hyperresearch goes the other way: it takes the time to verify sources mechanically. Setup requires a command-line environment, and each run costs model usage and waiting time, so it is not the tool for everyday lookups. For research where citation accuracy is under scrutiny, or work that follows one field over a long period, the design that leaves your reading material behind in a searchable vault starts to pay off. A reasonable way in is to run a light question at the light tier first, look at both the report and the notes it saved, and decide from there whether to adopt it seriously.