phishunt-mcp

Model Context Protocol (MCP) server for phishunt.io.

Exposes the public phishing-domains feed as MCP tools so AI agents can look up suspicious domains, list phishings by targeted brand, and sync detection deltas.

• Endpoint: https://mcp.phishunt.io/ (HTTP JSON-RPC 2.0, POST)
• Protocol version: 2025-11-25
• Auth: none (data is CC0)
• License (data): CC0-1.0 · License (code): MIT

Tools

Name	Purpose
check_domain	Is this host (or a list of up to 20) in the active phishunt feed? Exact match; misses are checked against the archive. PREVIOUSLY DETECTED is printed only when the archived row's verdict is medium, high or critical; a low/noise archive row is reported as a low-signal candidate, and a host that is not valid DNS is reported as such instead of an upstream error.
list_brand_phishings	List active phishings targeting a brand (e.g. microsoft).
get_recent_detections	Delta sync: detections since an ISO date.
get_brand_metadata	Curated brand metadata (display name, category, AI characterisation, active count).
get_cert_metadata	Factual metadata for an abused TLS intermediate CA (operator, root, key type, siblings).
search_phishings	Free-text search across active phishing URLs/domains/IPs (min 3 chars).
analyze_url	Passive phishing-signal analysis of any URL/domain - returns a single adjudicated verdict (phishing / likely_phishing / suspicious / no_evidence / not_assessed) plus a calibrated probability block beside it (P of malicious at the API's ~1.3% base rate, relative risk, 80% interval, per-class evidence; URL-only estimates stay low by design, so a low probability is not proof of safety) and the supporting evidence (URL-shape heuristics, stored score/verdict if known, external-feed cross-reference, historical detections). Unknown suspicious domains are auto-queued for full analysis.
analyze_url_deep	ACTIVE deep analysis of a URL (contacts the target: HTTP + TLS cert + RDAP + NS + GeoIP, SOCKS5-isolated) and, when the render lane is free, renders the page in a headless browser through the same proxy; it re-scores with the full 5-layer engine and returns coverage (active_rendered / active_no_render), render.status and a probability block. Slow (typically 15-45s, up to ~50s), token-gated, and rate-limited (shared 50/day budget, single-flight) - requires DEEP_TOKEN configured on this Worker; use only when analyze_url is inconclusive.
get_related_infrastructure	Find infrastructure/content overlap for a known indicator (shared IP, cert, nameservers, favicon, etc.); surfaces the possible campaign / suspected cluster it belongs to.
get_campaigns	List possible campaigns / suspected clusters (shared-infrastructure groupings), optionally filtered by brand or active-only.
get_campaign	Full detail for one possible campaign / suspected cluster: evidence breakdown, every member indicator, export links.

Use with Claude Desktop / Claude.ai / other MCP clients

Add to your MCP client config:

{
  "mcpServers": {
    "phishunt": {
      "url": "https://mcp.phishunt.io/"
    }
  }
}

Quick test

curl -sX POST https://mcp.phishunt.io/ \
  -H 'content-type: application/json' \
  -d '{"jsonrpc":"2.0","method":"tools/list","id":1}' | jq .

Secrets

analyze_url_deep needs a DEEP_TOKEN Worker secret (the backend's X-Phishunt-Deep-Token). Without it configured, the tool still appears in tools/list but fails clean on tools/call — it never reaches the backend.

wrangler secret put DEEP_TOKEN

Develop

npm install
npm run dev          # wrangler dev on http://localhost:8787
MCP_URL=http://localhost:8787 npm test

Deploy

npm run deploy       # wrangler deploy
MCP_URL=https://mcp.phishunt.io npm test
# after the phishunt-backend release that adds `probability` and `history.apex_candidates_seen`:
REQUIRE_PROBABILITY=1 MCP_URL=https://mcp.phishunt.io npm test

Without REQUIRE_PROBABILITY=1 the live probability / history assertions are reported as SKIPPED while the backend field is absent or probability.status is unavailable; with it they fail.