# Search that shows its work

> Search that shows its work is an open-source MCP server for web search and scraping. Self-hosted MCP web search that flags unreliable engines and results for AI clients. It is available on API, CLI and self-hosted.

- **Kind:** MCP server
- **Pricing:** Open source
- **Platforms:** API, CLI, self-hosted
- **Repository:** https://github.com/ag-bureau/mcp-search
- **MCP Registry name:** com.ag-bureau/search
- **Categories:** [web search and scraping](https://ai-powered.com/categories/web-search-and-scraping)
- **Listing updated:** 2026-11-18

## About

The README says this server runs on your own machine behind a self-hosted metasearch instance, so queries go out from your network rather than a hosted search API. Its distinguishing idea is that failures which usually look like success, such as an engine answering with a different subject, a page returning an anti-bot shield, or duplicate sources counted several times, are reported in explicit fields that a calling model can branch on.

The server exposes five tools: a combined search-and-read call, page reading with text, PDF and optional vision-model recognition of scans, image search that separates file and page addresses, screenshots returned alongside the page text from the same visit, and a multi-step deep search that reports what it could not confirm. A model key is optional; plain search, reading, image search and screenshots make HTTP requests without spending model tokens.

The project is aimed at developers and operators who connect AI clients to web research and want to control cost, access and robots.txt handling themselves. It ships with Docker Compose setup, a token-based access door with tiers for callers, and an AGPL-3.0 license that requires published changes when the service is offered over a network.

## Features

- **Explicit failure fields:** Engine refusals, silent empty results and irrelevant answers are reported in a trouble object instead of passing as success.
- **Search with optional page reading:** One call returns links and can also read the top results, with a flag to skip reading when only links are needed.
- **Page reading with browser mode:** Reads pages as text, PDFs or scanned documents, and can use a sandboxed browser for JavaScript-rendered pages.
- **Screenshots with matching text:** Returns a PNG of a page together with the text from the same visit so the two can be compared.
- **Caller access tokens:** Optional token list with per-tool scopes controls which clients may call which tools.
- **Robots.txt reporting:** Reports whether a site's rules allow a path, disallow it, or could not be checked, without enforcing them.

## Good for

- Giving an AI assistant web search that reports when results are unreliable
- Reading and summarising specific web pages or PDFs from an MCP client
- Running a private search backend without sending queries to a hosted search API
- Checking a page's visible text against its screenshot
- Running multi-source research with a record of what could not be confirmed

## Tools

- `web_search`: Finds pages and reads the top ones in one call.
- `web_read`: Reads pages by address, including text, PDF and scanned documents.
- `web_image_search`: Finds images and the pages they appear on.
- `web_screenshot`: Returns a PNG of a page plus its text from the same visit.
- `web_deep_search`: Composes its own queries, reads in waves and answers from several sources.

## FAQ

### Does it need a language model key?

No. The README says search, reading, image search and screenshots spend no model tokens. A model is only called for deep search and for recognising scanned PDFs.

### Can it be run over stdio as well as HTTP?

Yes. The README describes an HTTP mode for a running server and a stdio mode where the client starts the server as a process, selected explicitly.

### Does it obey robots.txt?

It reports each site's robots.txt status but does not enforce it. The README says the decision to add a gate is left to the operator.

### What license is it under?

GNU Affero General Public License v3.0, with a commercial licence mentioned as available by enquiry.

### Is Search that shows its work free?

Yes. Search that shows its work is open source, so you can use and self-host it for free.

### What are the best alternatives to Search that shows its work?

The closest alternatives to Search that shows its work are ShadowCrawl, Open Agent Search, searchio, Parallel Search MCP and WebSearch MCP Server.

## Where this comes from

- Description: written from the project's README and its MCP Registry entry.

## Alternatives

- [ShadowCrawl](https://ai-powered.com/mcp/shadowcrawl): ShadowCrawl is an open-source MCP server for web search and scraping.
- [Open Agent Search](https://ai-powered.com/mcp/open-agent-search): Open Agent Search is an open-source MCP server for web search and scraping.
- [searchio](https://ai-powered.com/mcp/searchio): searchio is an open-source MCP server for web search and scraping.
- [Parallel Search MCP](https://ai-powered.com/mcp/parallel-search-mcp): Parallel Search MCP is a freemium MCP server for web search and scraping.
- [WebSearch MCP Server](https://ai-powered.com/mcp/websearch-mcp-server): WebSearch MCP Server is a free MCP server for web search and scraping.
- [Visdenda Free MCP](https://ai-powered.com/mcp/visdenda-free-mcp): Visdenda Free MCP is a free MCP server for web search and scraping.
- [MCP Server Search](https://ai-powered.com/mcp/mcp-server-search): MCP Server Search is an open-source MCP server for web search and scraping.
- [Librecrawl — Technical SEO Audit MCP Server](https://ai-powered.com/mcp/librecrawl-technical-seo-audit-mcp-server): Librecrawl — Technical SEO Audit MCP Server is an open-source MCP server for marketing and seo.

---
Source: https://ai-powered.com/mcp/search-that-shows-its-work
