@realurls/mcp
MCP server that tells an AI agent which domain officially belongs to which software product or company — so it stopshanding users SEO-poisoned lookalike download sites.
Ownership only. Never a safety judgement. Read TRUST.md for exactly what we do and do not claim.
Tools
| tool | when the agent calls it | returns |
|---|---|---|
get_official_url(name) |
before giving any download / login / official-site link | official + verified URLs, or insufficient_evidence / unknown |
verify_url(url) |
user pastes a link and asks if it's real | official / not_official (+ the real domain) / insufficient_evidence / unknown |
The server ships instructions telling the host to call get_official_url even when it thinks it already knows the URL.In our tests that single line is what makes the difference: without it, agents answer well-known tools from memory and never verify.
Every result also carries say_to_user: the sentence the agent should give for that verdict ("I could not confirm the official site", "X is not a known domain of Y; Y's verified site is …"). Only official results contain URLs the agent may hand out.
Install
Claude Desktop — claude_desktop_config.json:
{ "mcpServers": { "realurls": { "command": "npx", "args": ["-y", "@realurls/mcp"] } } }
Claude Code:
claude mcp add realurls -- npx -y @realurls/mcp
Point at a local API during development: REALURLS_API=http://127.0.0.1:8787.
What "official" means
official = the domain has ≥1 anchor (GitHub-verified org, DNS self-attestation, restricted government TLD, …) and ≥2 independentcorroborations, all machine-reproducible. Anything less is reported as insufficient evidence — we would rather say "don't know" than be wrong.
