Skip to content

Guides

Give your AI agent live threat intelligence

A bare LLM answers security questions from training data that is months stale - confidently. On a 150-question benchmark about recent threats, a frontier model scored 0/150 ungrounded and 158/158 grounded (methodology). Grounding an agent takes one of three routes, all free to start.

Route 1: one-click connectors (no code)

Claude: install from the Claude directory - sign in, and Claude gets ten read-only tools over incident clusters, entity profiles, CVEs with KEV/EPSS, and ransomware leak-site victims. ChatGPT, Cursor, VS Code, Windsurf and Zed setups are on the integrations pages.

Route 2: the MCP server (any MCP client or framework)

npx -y threatcluster-mcp     # or: uvx threatcluster-mcp

Set THREATCLUSTER_API_KEY (free keys, 100 credits a day - get one here). Hosted alternative with OAuth, no local process: https://threatcluster.io/mcp. The server is in the official MCP registry as io.github.Jam0k/threatcluster.

A complete working example - a SOC triage agent in ~80 lines that takes an alert and returns a cited verdict - is at github.com/Jam0k/soc-triage-agent.

Route 3: the REST API (any language, no MCP)

curl -s "https://threatcluster.io/api/public/v1/search?q=lockbit&days=7" \
  -H "X-API-Key: tc_live_..."

Every response carries source URLs for citation and the credits it cost; empty results are refunded. The API reference covers all endpoints.

What to let the agent do

The tools are read-only by design. The patterns that work: triage ("is this alert real?" - resolve the indicators, check the CVE, cite the reporting), stack watch ("did anything in our vendors become exploitable this week?"), and context pulls ("what is this actor, who have they hit?"). Let the model decide which tools to call; give it the house rule that threat facts come from tools, never from memory.