TinySearch MCP
Local web search for agents: searches, reads pages and returns compact passages with sources, saving context tokens
- 236 ★ on the repository
- Last commit
Extended access
Why this level
- The server goes online on your behalf, opens pages and runs browser sessions where entered data may end up
- The full Docker setup starts SearXNG and a network port for MCP
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/tinysearchDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The server goes online on your behalf, opens pages and runs browser sessions where entered data may end up
- The full Docker setup starts SearXNG and a network port for MCP
This is third-party code. Review the repository files before installing.
What it does
TinySearch does the work with web pages before the model. It searches the web, opens the best pages, strips navigation and boilerplate, reranks passages locally and returns only evidence with source URLs to the agent. The model spends tokens on the useful paragraphs rather than on whole pages. TinySearch is part of TinySuite, whose aim is to make agent work cheaper. MCP tools: `search(items)` quickly finds sources without crawling or reranking, `scrape_urls(items)` reads one to five known pages with a token budget, `browser_navigate` and `browser_act` handle pages that need interaction, and `get_current_datetime()` answers questions about time. For each connection the interactive tools get their own browser and page, so clients do not see each other's tabs or cookies. Three ways to run: a Python library, a one-line `uvx` for an MCP client and Docker with SearXNG and HTTP.
Who it is for. Agent developers who pay for tokens and want to search the web without paid search APIs and without dumping whole pages into context.
Good fit when
- A research agent gathers facts on a topic and has to fit a small context token budget
- You need search without a paid provider: by default it works through DDGS and the full setup through SearXNG
- A page shows data after a click and the agent goes through it with `browser_act`
- Several clients share one server and their browsers are isolated from one another
Not a fit when
- You need your own permanent web index: TinySearch stores nothing as a search engine
- You need an agent or chat: this is a retrieval layer, not a separate agent or interface
- The model's built-in web search is enough: a separate server is extra for rare queries
Example request
Find three independent sources on the new library release, read them and give a short summary with links to eachMCP
- Transport
- stdio, http
- Authentication
- not required
Limitations
TinySearch stays a retrieval layer and does not replace an agent, a chat interface or a permanent index. Tiers 1 and 2 use DDGS and need no search service, while the full local setup needs tier 3 with Docker, SearXNG and persistent model storage. No more than four interactive browser sessions are open at once (`browser_max_sessions`) and a new caller waits up to 30 seconds. Pages are read without any guarantee of access to closed sites. The MCP registry version is 0.6.8 and the license is MIT.
How to disable. Remove the tinysearch block from the client config, stop the container with `docker compose down` if you ran the full setup and uninstall the `tinysuite-search` package.
FAQ
How do I run TinySearch for an MCP client?
With `uvx --from "tinysuite-search[server]" tinysearch`: the client launches the server itself. Search goes through DDGS and needs no separate service.
How do the three ways to run differ?
A Python library (`pip install tinysuite-search`) for embedding, one `uvx` command for an MCP client and Docker with SearXNG, persistent model storage and network MCP over HTTP.
What tools does the server have?
`search`, `scrape_urls`, `browser_navigate`, `browser_act` and `get_current_datetime`. Full parameters are in the reference at tinysuite.dev.
How does TinySearch save tokens?
It returns only the best passages within an evidence budget rather than whole pages and reranks them locally. Free search through DDGS removes the fee for a metered API.
Are browser sessions isolated?
Yes, each MCP connection gets its own browser and page so clients do not see each other's tabs or cookies. No more than four sessions are open in parallel.
Discussion
No questions or reviews yet.
Related
MCP serversA local MCP server with no API keys: multi-engine web search, page fetching and document reading tuned for LLMs.
A unified search MCP server: one interface to Tavily, Brave, Kagi, Exa, Linkup, Firecrawl and GitHub search
MCP server for real-time web search: general and vertical search, batch queries and page content extraction
MCP server with DuckDuckGo search and page text extraction: several URLs can be passed in one call
SearXNG AI Kit gives agents web search across 180+ engines through a CLI and an MCP server, with no search API key
Metasearch for agents: one MCP server and HTTP API merge results from many search providers into unified JSON with deduplication.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| TinySearch (this card) | no | not required | 236 | |
| Context7 | yes | API key | 64,656 | library documentation on demand rather than search across the open web and page reading |
| MCP Local RAG | no | not required | 430 | search over your local documents rather than the web |
README in short
The text promises to spend tokens on answers rather than webpages and shows the pipeline: search and crawl, local reranking, compact evidence. Then come three installation tiers, an explanation of how TinySearch cuts cost and a table of MCP tools.