web-researcher-mcp

Purpose

web-researcher-mcp is an MCP (Model Context Protocol) research server that provides AI assistants with real web search, full-page reading, and multi-source research capabilities, while ensuring cited sources are real and verifiable and not fabricated. It is designed to solve the problem of unreliable AI source citations—for example, linking to papers that do not exist, fabricating DOIs, or presenting SEO spam on equal footing with peer-reviewed research. The core idea of the tool is: rather than letting AI search the entire web and hoping the results are reliable, let users explicitly specify which sources AI may search, and these source lists are called "search lenses."

Main capabilities

• Web search: Supports searching the entire web, or limiting to sites the user trusts via search lenses.
• Full-page reading: Analyzes complete source content rather than relying only on summary snippets.
• Multi-source research: Supports multi-step research workflows, obtaining information from user-specified sources.
• Citation verification: Ensures cited sources are real and verifiable, and not fabricated.
• Academic search: Optionally configures academic search providers such as OpenAlex and CrossRef, returning real papers with DOI, authors, citation counts, and open access PDF links.
• Patent search: Optionally configures patent providers such as EPO, USPTO, and The Lens, returning structured patent data, including classification codes, dates, and inventors.
• Search lenses: Includes built-in lenses for multiple domains, and also allows creating custom lenses to limit AI to searching only specified domains.
• Multi-provider failover: Allows configuring multiple search providers, automatically switching to another when one is unavailable.
• Research templates: Provides guided prompt templates to help AI assistants complete research through multi-step workflows.
• Status and diagnostics resources: Exposes real-time status resources (such as stats://tools, stats://sessions, stats://rate-limits, stats://providers), lens catalog (lenses://catalog), diagnostic information (diagnostics://errors/recent, diagnostics://health), and large payload artifact storage (research://artifact/{id}).

Usage

Installation

• Python users (uvx):uvx web-researcher-mcp, no compilation required, available on any operating system. Also supports uv tool install web-researcher-mcp or pip install web-researcher-mcp。
• Homebrew（macOS / Linux）：brew install zoharbabin/tap/web-researcher-mcp.
• One-click install (any OS): Downloads the binary file, verifies the SHA-256 checksum, adds it to PATH, and automatically registers with Claude Code when detected.
• Windows: Supports WinGet, Scoop, Chocolatey.
• macOS Cask: Provides Developer ID signed and notarized binaries.
• Go install:go install installation method.
• Build from source: Supports building from source.
• AUR (Arch Linux): Supports AUR installation.
• NixOS: Provides a Nix flake module.

Connect an AI assistant

The installation script automatically registers with Claude Code. Other applications need to add MCP configuration in the AI's configuration file. Supports Claude, Claude Desktop, Cursor, and any AI assistant that supports tool use.

Search provider configuration

• No API key required: DuckDuckGo is the built-in zero-configuration fallback, ready to use after installation.
• Optional providers: Set SEARCH_PROVIDER=<name> and provide the corresponding key. All providers support search lenses, and can be combined to achieve automatic failover.
• If not set SEARCH_PROVIDER , use Google if a Google key exists, otherwise fall back to DuckDuckGo.
• Academic search providers (OpenAlex, CrossRef) accept a contact email to unlock faster access, with no registration required.
• Patent search providers (EPO, USPTO, The Lens) require API keys.

Search Lenses

• Built-in lenses cover multiple domains.
• Custom lenses can be created: create a directory, add a JSON file for each lens, then point to that directory via configuration. Custom lenses load after built-in lenses, and a custom lens with the same name overrides the built-in lens. Each lens can have up to about 10 domains added.
• Advanced options:cx(Google Programmable Search Engine ID, up to 5,000 domains),routing(force that lens to use a specific search provider).

HTTP Mode

Supports team-shared instances, connecting any AI application to http://localhost:3000/mcp/. Tool behavior is consistent between STDIO and HTTP modes; the differences are authentication (HTTP requires OAuth) and rate limiting (HTTP is limited per tenant; STDIO is limited only by upstream API quotas).

Use Cases

• Academic researchers: need literature reviews with real DOIs, not fabricated citations.
• Business analysts: deliverables need sources that clients can click and verify.
• Lawyers: citing nonexistent cases can lead to hefty fines.
• Journalists: need to cross-check government records and court documents rather than relying on summaries.
• Medical researchers: clinical decisions cannot be based on health blogs.
• Graduate students: avoid spending hours tracking down citations fabricated by AI.
• Enterprise teams: competitive research cannot go through third-party servers.

Limitations

• Some tools require specific providers or configuration to be usable:citation_graph and research_panel require at least one configured fallback provider;filing_search requires EDGAR_CONTACT_EMAIL；local_search requires BRAVE_API_KEY.
• Operators can enable optional tools that require user consent (such as per-user analytics, long-term memory, shared workspaces, saved query monitoring); these tools appear only when the feature is enabled.
• Academic search and patent search remain usable when no provider is configured, but fall back to web search.
• Image search and news search require API keys.
• For heavily JavaScript-based sites, the tool uses a real browser (Chromium) in the background; with a binary installation, it is automatically downloaded on first use (about 200MB).
• The server still starts when credentials are missing (to allow the MCP handshake); API keys must be set in the env block of the MCP client configuration, not in a shell profile file.

Privacy and Security

• Research queries are sent directly from the user's machine to the chosen search provider, without going through third-party servers.
• Provides SSRF protection, OAuth 2.1 (HTTP mode), rate limiting (HTTP mode), and content sanitization.
• See docs/SECURITY.md for the full threat model.