YaCy peer-to-peer search

yacy-search MCP server

An MCP server for one YaCy peer: web search without a search API or API key, crawling, ranking settings and search quality evaluation. It runs next to the agent (stdio) and talks to the agent's own peer over HTTP. There is no central service: the peer is yours, and it searches the peer-to-peer network it belongs to. Registry name: io.github.pad01g/yacy-search.

It is written for the improved-search fork (results carry author signature verdicts and tags). Search, crawl, status and settings also work with upstream YaCy.

Run

Start a peer and the MCP server on one Docker network:

docker network create yacy 2>/dev/null || true
docker run -d --name yacy --network yacy -p 127.0.0.1:8090:8090 \
  -v yacy_data:/opt/yacy_search_server/DATA ghcr.io/pad01g/yacy-improved-search:latest

# Claude Code
claude mcp add yacy -- docker run -i --rm --network yacy \
  -e YACY_URL=http://yacy:8090 -e YACY_ADMIN_PASSWORD='<your password>' ghcr.io/pad01g/yacy-search-mcp:1.0.0

The fork's image starts with the administrator account admin / yacy: change the password at http://127.0.0.1:8090/ConfigAccounts_p.html before connecting the MCP server, and pass the new one.

Other clients (mcp.json style):

{
  "mcpServers": {
    "yacy": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "--network", "yacy", "-e", "YACY_URL=http://yacy:8090", "-e", "YACY_ADMIN_PASSWORD=<your password>", "ghcr.io/pad01g/yacy-search-mcp:1.0.0"]
    }
  }
}

For a peer outside Docker use -e YACY_URL=http://host.docker.internal:8090 (Linux: add --add-host=host.docker.internal:host-gateway). Without Docker: npm ci && node src/server.ts in this directory (Node 24).

Variable	Default	Meaning
YACY_URL	http://localhost:8090	base URL of the peer
YACY_ADMIN_USER	admin	administrator account (crawl, status counts, settings)
YACY_ADMIN_PASSWORD	unset	administrator password; without it crawl_site, control_crawl, delete_document and set_ranking_setting are refused (even if the peer accepts local requests without a password), and the reading admin tools work only where the peer allows them without one
YACY_CRAWL_ALLOW_PRIVATE	unset	1: crawl_site may reach loopback and private addresses (an intranet). Otherwise only http(s) URLs of public hosts. This checks the start URL only; the links and redirects YaCy follows are limited by YaCy's own network.unit.domain (global on public peers refuses local addresses)
YACY_TIMEOUT_MS	30000	timeout of one request to the peer
YACY_ALLOW_TRUST_SETTINGS	unset	1: set_ranking_setting may also change trust.search.acceptUnverified and trust.policy.excludeTags

On the improved-search fork, the vector search (short questions find pages without the exact words) is used only for searches of the peer's administrator or from its host. When the MCP server runs in another container, set search.vector.public=true on the peer (it then embeds every new query of its search page).

Tools

Tool	What it does
search_web	full-text search; resource: "global" also asks connected peers and waits waitMs before ranking all answers
crawl_site	start a crawl (url, depth, range: domain / subpath / wide, maxPages per host, default 200)
get_index_status	indexed documents, crawl queues, peer type and connections; explains a paused crawler (load average) and a peer without a public address
list_peers	connected peers: signed seed, reach (direct / relay), tags
get_ranking_settings	the ranking and filtering settings with their meaning
set_ranking_setting	change one of them (allow-list only: minimum match, coverage and thin page weights, title phrase boost, wait time; unverified results and excluded tags only with YACY_ALLOW_TRUST_SETTINGS=1)
evaluate_ranking	queries with known relevant URLs: precision@k, recall@k, R-precision, the rank of each relevant URL; finishes within about 50 s (queries that do not fit are reported as skipped)
get_trust_status	the signed delegations and peer lists the peer holds
list_crawls	the crawls started on the peer, with the handle control_crawl needs
control_crawl	pause or resume the local crawler, or stop one crawl (indexed pages stay)
delete_document	remove one page from the peer's own index (stop a running crawl of that site first)

1.0.0 renamed the tools to a consistent verb_noun form (no aliases): search → search_web, crawl → crawl_site, crawls → list_crawls, crawl_control → control_crawl, index_status → get_index_status, peers → list_peers, trust_status → get_trust_status. Clients that call tools by name need the new names.

A typical loop for improving results: evaluate_ranking → set_ranking_setting → evaluate_ranking.

Titles and snippets come from page authors and other peers: the tool descriptions tell the model to treat them as data, not instructions. crawl_site, control_crawl, delete_document and set_ranking_setting are marked destructive so that clients ask before calling them. Values of settings are checked per key, and a value YaCy stores differently is rolled back.

Test

npm ci && npm run typecheck && npm test                      # unit tests
YACY_URL=http://localhost:8090 YACY_ADMIN_PASSWORD=yacy npm test   # also calls every tool on a live peer

License

This directory (the MCP server) is under the MIT license. It talks to YaCy only over HTTP and contains no YaCy code; the rest of this repository, YaCy itself, stays under the GPL.