M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
H

Hyperresearch MCP Server

Community
A deep research agent for Claude Code that turns every page it reads into a persistent, searchable vault.
How to install ↓ Compare GitHub source repository ↗
Category
Other #83 of 230
Stars
★ 3.9k Very popular
Transport
stdio (local process)
Runtime
Python 3.11–3.13
Credentials
Optional API key
License
MIT
Last commit
Tools
13
59FMRS · C

Hyperresearch is an MIT-licensed individual open-source project that persists research into a searchable markdown vault and ships a stdio MCP server so Claude Desktop, Cursor, and other clients can reuse that vault. Its strength is durability and portability: markdown is the source of truth and SQLite is only a rebuildable cache. Be aware that the project's self-reported benchmark results are pending third-party validation and that the README explicitly does not guarantee factual accuracy, leaving source judgment to the user.

Strongest · Documentation 15/20 Weakest · Reliability 9/20

Reliability
9/20
Security and permissions
12/20
Maintenance
11/20
Documentation
15/20
Setup experience
12/20
Why each score
Reliability 9/20
The README documents a complete MCP path (pip install hyperresearch[mcp], then hyperresearch mcp over stdio with thirteen tools), but no server manifest is cached and no CI workflows, test files, or execution logs are visible in the supplied evidence. Tool behavior (search_notes, read_many, fetch_url, create_note, update_note, etc.) is asserted only by prose; parameter schemas, error-return shapes, retry/timeout semantics and failure modes are not specified. External dependencies (Unpaywall, Europe PMC, Semantic Scholar, OpenAlex, crawl4ai, optional Voyage/OpenAI embeddings) widen the failure surface, though the README does describe degradation paths (provider none needs no API keys; OA recovery never fails a fetch; run resume continues from the failed step; budget cap blocks the run). Under the static calibration rule, the absence of real CI plus committed tests caps reliability at 12; without verifiable execution evidence for the key paths, 9.
Security and permissions 12/20
Positives: web-fetched bodies are served inside an <untrusted-source> fence on both body-serving paths (note show in single/batch/JSON, and search with bodies), forged fence tags are neutralized and left visible, the url attribute is HTML-escaped with control characters stripped, and third-party API-resolved URLs are checked for scheme, embedded credentials and publicly-routable resolution, which is explicit prompt-injection hardening; no real tokens or secrets appear in install examples; CAPTCHAs, 2FA and logins are never solved automatically and are consolidated back to the user. Deductions: authenticated crawling via hyperresearch setup reuses the user's real logged-in browser sessions (LinkedIn, Twitter, Facebook, Instagram, TikTok) and the browser-fetcher agent drives the user's real, logged-in Chrome, a high-privilege, cross-site credential-reuse boundary with no stated credential storage location, encryption, least-privilege scoping, or auditable consent flow; MCP write tools (create_note, update_note, fetch_url) have no documented write-scope restriction or confirmation requirement; the local HTTP server (serve --open on port 8080) mentions no authentication. No red-line item is met (no malware, no credential theft, no unconfirmed irreversible defaults), so this is not scored 0-4, but with permission and confirmation mechanisms incomplete, 12.
Maintenance 11/20
Facts: MIT license is explicit; repository is not archived, 1,906 stars, 27 open issues (a high count with no visible response data); README references PyPI releases (version badge) and git integration commands, implying ongoing work. Gaps: no release cadence, no CHANGELOG, no contributor governance description, no visible security-response channel (no SECURITY.md), and no dependency-update policy; issue triage/closure evidence is missing so maintainer responsiveness cannot be verified. Scores at the "active but governance/versioning gaps" anchor: 11.
Documentation 15/20
Documentation is the strongest aspect. It is layered and covers installation (including the Python 3.11-3.13 window and 3.14 unsupported note, plus the per-project vs --global trade-off), the 16-step pipeline with tier/gear/lever concepts, the subagent roster with default models, the vault data model (markdown is truth, SQLite is a rebuildable cache), the thirteen MCP tool names, CLI examples with -j output, fetching security boundaries, academic-API-first ordering, the four disclosure locations for open-access substitution (body banner, frontmatter fields, the body_is_not_from_source flag in note show JSON, and CLI output), the [scholar] config keys with the Unpaywall contact_email requirement, and an explicit "what it doesn't do" limitations section (does not replace judgment, cannot bypass paywalls, lint only catches structural failures, not factual accuracy). Deductions: no parameter-level schema or examples for the MCP tools, no error codes or rate-limit guidance, no MCP client configuration snippets (Claude Desktop/Cursor), cost only given as time ranges, and no troubleshooting section; marketing claims about leading DeepResearch-Bench with internal benchmarking are footnoted as pending third-party validation but remain external claims needing verification. 15.
Setup experience 12/20
The install path is friendly to mainstream clients: pip install hyperresearch && hyperresearch install and then /hyperresearch inside Claude Code, or pip install hyperresearch[mcp] plus hyperresearch mcp (stdio) for Claude Desktop, Cursor, or any MCP client. The Python version window is explicit (3.11-3.13, 3.14 unsupported) with three concrete environment recipes (pyenv, uv venv, py -3.13), so the path is short and clear. Deductions: no MCP client configuration file example (JSON snippet absent), no Windows/macOS/Linux differences stated, and full functionality depends on Claude Code plus the optional Claude-in-Chrome extension; under static calibration with no CI/test execution evidence setup is capped at 15, and because several manual configuration surfaces remain (browser setup, crawl4ai, embedding provider, Unpaywall contact email), 12.

Static review · not runListed 2026-09-10

Read the FMRS scoring method →

Fit and risk

What it can accessReads local filesWrites / deletes local filesUses the networkConnects to a database

Best for

  • Users who already maintain a local Hyperresearch markdown vault and want other MCP clients to reuse the same notes
  • Researchers who want their desktop AI client to read existing local material instead of re-searching the web every time
  • Individual researchers or small teams who care about portability (plain markdown, git-friendly, JSON exportable)

Not for

  • Users who want zero setup and only the built-in Claude Code skill without maintaining a local vault
  • Users who cannot install Python 3.11+ or run local command-line tools
  • Users expecting the server to guarantee factual accuracy — the README states the lint gate catches structural failures only and factual accuracy remains the user's call
  • Users who need paywalled content but have not completed a local login

Required permissions

  • Read and write the vault directory in the working tree (such as research/notes markdown files)
  • Create and rebuild the SQLite index file; hyperresearch sync reconstructs it from markdown
  • Make outbound network requests to fetch external URLs via tools such as fetch_url
  • If embeddings are enabled, reach the configured embedding provider (voyage, openai; the default none needs no API keys)
  • Local file-write permission required to create and update notes

Risks and side effects

  • Fetched web content is untrusted input: the README says fetched bodies are served inside <untrusted-source> fences and treated as data, and clients should keep the same stance to avoid being steered by embedded instructions
  • Write tools (create_note, update_note) modify local vault files directly, so misuse can corrupt the note collection; keeping the vault under git is advisable for rollback
  • Fetching external URLs exposes requests to third-party sites, potentially revealing the research topic or IP
  • Scraping paywalled or login-gated sites is subject to those sites' terms, and login session credentials are the user's own responsibility
  • Claims in the README such as the DeepResearch-Bench leaderboard position and 250+ sources are the project's own statements and a forward-looking projection; third-party validation is pending and should not be treated as proof of quality

Setup

Before you start

Runtime：Python 3.11–3.13

1) Make sure Python 3.11–3.13 is available (3.14 is not yet supported). 2) Run pip install hyperresearch[mcp] to install the MCP-enabled build. 3) Run hyperresearch mcp, which starts the MCP server over stdio. 4) In Claude Desktop, Cursor, or another MCP-speaking client, configure hyperresearch mcp as the server command. The README does not provide a ready-to-copy client config snippet, so follow your client's MCP configuration format.

claude_desktop_config.json 
{
  "mcpServers": {
    "hyperresearch": {
      "command": "hyperresearch",
      "args": ["mcp"]
    }
  }
}

Shown for Claude Desktop. Other clients may use a different file or key (VS Code uses "servers") — the configurator below converts it.

.vscode/mcp.json
{
  "servers": {
    "hyperresearch": {
      "command": "hyperresearch",
      "args": [
        "mcp"
      ]
    }
  }
}

Goes in your project's .vscode/mcp.json (VS Code uses a "servers" key).

Terminal
claude mcp add hyperresearch -- hyperresearch mcp

Run it in a terminal; replace any <…> placeholders with your own values first.

Generate and validate client config

Check that it works

After pip install hyperresearch[mcp] and running hyperresearch mcp, the client should list the 13 tools such as search_notes and fetch_url; calling vault_status should return vault information.

Troubleshooting

  1. Confirm Python is 3.11–3.13; 3.14 is currently unsupported
  2. Confirm pip install hyperresearch[mcp] was run, otherwise the hyperresearch mcp subcommand is unavailable
  3. If the client cannot connect, check that the server is started over stdio and the command is on the client's PATH
  4. If vault search looks wrong, run hyperresearch sync to rebuild the SQLite index from markdown (the index is a rebuildable cache)
  5. Use lint_vault or hyperresearch lint to check broken links, missing tags, and source connectivity issues
  6. If semantic search is unavailable, check the embedding provider configuration; the default none performs no vector search

Things to try

Once connected, you can ask your AI assistant things like:

  • Search the vault for ion-trap gate fidelity and list the most relevant notes
  • Read these notes and summarize the main arguments across them
  • Fetch https://arxiv.org/abs/2310.12345 and save it as a new note in the vault
  • Give me the vault status and run a lint check on it

Tools 13

search_notes read-only
Full-text search across the persistent research vault, optionally including note bodies or semantic search.
read_note read-only
Read the contents of a single note.
read_many read-only
Batch-read multiple notes in one call.
list_notes read-only
List notes held in the vault.
get_backlinks read-only
Return the reverse links pointing at a given note.
get_hubs read-only
Return the most-connected notes in the vault.
vault_status read-only
Report status information about the vault.
lint_vault read-only
Run a vault health check such as broken links and missing tags.
Show 5 more tools
check_source read-only
Check information about a given source.
list_sources read-only
List the sources recorded or cited in the vault.
fetch_url writes
Fetch the content of a given URL.
create_note writes
Create a new note in the vault.
update_note writes
Update an existing note.

Use cases

Let Claude Desktop or Cursor search and read notes from an existing Hyperresearch vault
Pull a batch of previously archived sources and summaries with search_notes and read_many before writing a report
Keep accumulating research notes inside an MCP client via create_note and update_note
Audit sources and vault structure with list_sources, check_source, and lint_vault

Supported clients

Claude Desktop
CursorPartial support

Listed from the project's documentation, not tested by this site.

Overview

Hyperresearch is an agent-driven research knowledge base. According to its README, it turns Claude Code into a deep research agent built around a tier-adaptive 16-step pipeline that produces an adversarially-audited report with full source provenance from a single prompt. Every source it reads lands in a persistent, searchable markdown-plus-SQLite vault, and each new session searches that vault before fetching anything new. The vault is reachable outside Claude Code through a built-in MCP server: install hyperresearch[mcp] and run hyperresearch mcp to speak stdio, exposing thirteen tools (search_notes, read_note, read_many, list_notes, get_backlinks, get_hubs, vault_status, lint_vault, check_source, list_sources, fetch_url, create_note, update_note) so Claude Desktop, Cursor, or any other MCP client can work the same vault. It is authored by jordan-gibbs as an individual open-source project under the MIT license, requires Python 3.11+ (3.11–3.13; 3.14 is not yet supported), and keeps notes as plain markdown with YAML frontmatter so the SQLite index is fully rebuildable and users are not locked into the tool.

Similar servers

BioMCP 75 · B

One binary. One grammar. Evidence from the biomedical sources you already trust.

★ 659 · Tools 11 Compare with this →
Compare with other MCP servers

Source revision 9c16854fed2d Data synced 2026-10-19 Read the FMRS scoring method