plumb-search

Plumb Search

A free, open-source web search engine that anyone can run. Plumb answers from its own index, not by scraping Google or Bing, so it needs no API key and has no quota. It speaks MCP and SearXNG's JSON API, so AI apps and local models can search with it too. Nodes share their crawling over a peer-to-peer network, so no one has to crawl the whole web alone.

Search at plumbsearch.org · Run your own node · Documentation · How search is doing

Use it

From	How
A browser	Search at plumbsearch.org, or add https://plumbsearch.org/search?q=%s as your browser's search engine
Claude Code and other MCP apps	claude mcp add --transport http plumb https://plumbsearch.org/mcp (MCP tools)
Apps that take a SearXNG address	https://plumbsearch.org, which answers /search?q=...&format=json (setup for each app)
Zotero	The Zotero plugin adds the papers, books and pages you find to your library
Your own code	/api/search?q=... on any node returns JSON (the API)

plumbsearch.org keeps no access log, and its results pages load nothing from other servers: Plumb serves site icons itself and draws maps from coordinates.

Run your own node

Your own node keeps your searches, history and settings on your machine. A new node copies the best sites of a node it trusts (plumbsearch.org by default), is searchable within minutes, and then crawls its share of the web. No port forwarding is needed.

On	Do this
A server or homelab	docker compose up -d in a clone of this repository, then open port 8080 (Docker)
A desktop or laptop	Install the app from the releases page; search opens at http://127.0.0.1:7586 (desktop app)
Anything with Rust	cargo build --release -p plumb-node, then ./target/release/plumb run --data plumb-data --network and open http://127.0.0.1:8080 (running a node)

What it finds

• The official site first. Names count for more than text, so look-alike sites stuffed with a brand's words stay below the real one.
• Pages from open page sets next to sites: Wikipedia, GitHub, Stack Overflow and other Stack Exchange sites, software packages, books, papers and podcasts (page sets).
• Places from OpenStreetMap for searches like "pizza in denver", with a small map (places).
• Answers and context: sums, conversions and the time somewhere, an info box about what the search names, links to a site's key pages, and recent headlines from its feed (news).
• Search by meaning, so "electric car maker" finds sites that never use those words (on in the Docker image and the desktop app).
• Private search at /private, where the browser ranks results itself and the node never sees the query (private search).
• Plugins that add results from other sources, run in a WebAssembly sandbox (plugins).

Plumb indexes sites, not every page: for each site it keeps names, the homepage's title, description and key terms, the words other sites link to it with, and Wikidata's facts. That keeps a node small enough for an ordinary computer, and it is why a blog post or forum thread deep inside a site is harder to find than with a full-text engine.

How it works

flowchart LR
    seeds["Seed lists<br/>Tranco, Common Crawl, Wikidata"] --> records
    homepages["Homepages<br/>its daily share"] -->|"crawl"| records
    peers["Other nodes"] <-->|"signed crawl batches"| records
    records[("Site records")] --> index["Index<br/>words and meaning"]
    sets["Page sets<br/>Wikipedia, packages,<br/>papers, places"] --> index
    index --> serve["Search page<br/>JSON API, MCP"]

Each day every node is given a random share of all sites to crawl. It signs what it saw and passes it on, and other nodes take a crawl in only when they can check it. A search is answered from the node's own index: candidates matched by words and by meaning, put in order by a small ranking model, with answers, page-set results and places alongside.

The architecture shows the crates and what runs inside a node, how search works covers matching and ranking, and the network covers sharing, trust and network search.

How search is doing

Measured on 1,000 real searches graded 0 to 3, in run 7 on 2026-10-25 (build 1584e58):

Mean grade	2.08 of 3
Known right answer at #1	75% of 464
Known right answer in the top 3	84%
Median time per search	0.72 s

Plumb is good at finding a thing by its name and weaker when a search describes it in other words. How search is doing has every run, each kind of search and where it falls short.

Documentation

For	Read
Searching	Using search · Private search · Places · Headlines
AI apps	MCP server · Local models and AI apps
Running a node	Running a node · Docker · Desktop app · Remote control · Plugins
How it works	Architecture · Search and ranking · The network · Page sets
The code	A README in each crate · Evals · How search is doing

docs/README.md lists every page.

Contributing

Issues and pull requests are welcome: CONTRIBUTING.md has the details, and SECURITY.md says how to report a security problem. Before opening a pull request, run:

cargo fmt --all --check
cargo clippy --workspace --exclude plumb-desktop --all-targets -- -D warnings
cargo test --workspace --exclude plumb-desktop

Data sources

Plumb is built from open data. Nodes keep titles, short descriptions and counts, not whole pages.

Source	What Plumb takes from it	License
Wikipedia (English)	Article titles and short descriptions, from Wikimedia's dumps; page views	Text CC BY-SA 4.0; page views CC0
Wikidata	Official websites, names, countries, kinds of site, descriptions and official profiles	CC0
Stack Overflow and other Stack Exchange sites	Question titles, tags and views, from Stack Exchange's data dump	CC BY-SA 4.0
ecosyste.ms	Package names, versions, licenses and links for eight registries	CC BY-SA 4.0
OpenStreetMap	Places; © OpenStreetMap contributors	ODbL
OpenAlex	The most cited papers, and where each can be read free (from Unpaywall's data and arXiv)	CC0
CORE	Where papers OpenAlex knows no free copy of can be read free, from university repositories (optional, with a CORE API key)	CORE's terms
Open Library	The most read books	CC0
Podcast Index	Titles, authors, categories and popularity of the most popular podcasts	Free for any use
GitHub	Names, descriptions and stars of public repositories, from its API	GitHub's terms
Tranco	The top million sites, for first ranks	See its site
Common Crawl	Web graph domain ranks; optionally titles and link text from WAT files	Terms of use
The Block List Project	The adult sites list for safe search	Public domain (Unlicense)
European Central Bank	Daily euro reference rates for currency conversions	Free to reuse with the source named
Open-Meteo	Weather forecasts for weather answers	CC BY 4.0, free without a key for non-commercial use
BAAI/bge-small-en-v1.5	The embedding model for search by meaning	MIT
EmbeddingGemma 2	The many-language model search by meaning can run instead (--meaning-model gemma)	See its model card

Site titles, descriptions and icons come from the sites' own homepages. The info box credits Wikipedia under each description it takes from an article, and lists of places credit OpenStreetMap.

License

Licensed under either of

• Apache License, Version 2.0 (LICENSE-APACHE)
• MIT license (LICENSE-MIT)

at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.