Knocklog

Who's knocking?

AI agents read your website every day: crawlers training models, AI search engines, and assistants fetching pages for someone who asked about you. Some of them are not who they say they are.

Free, no sign-up. We read your robots.txt, homepage and llms.txt the way an agent would.

What Knocklog sees at the door (example)
  1. GET /pricing GPTBot/1.4 20.171.206.12 verified IP in OpenAI's published ranges
  2. GET /docs/api ChatGPT agent signed verified Web Bot Auth signature from chatgpt.com
  3. GET /wp-login.php GPTBot/1.4 203.0.113.54 spoofed Claims GPTBot, IP is not OpenAI's
  4. GET /blog/agents Googlebot/2.1 66.249.66.1 verified Address resolves to googlebot.com and back
  5. GET / python-requests/2.32 198.51.100.7 unverified Script with no published identity
  6. GET /pricing Chrome 141 referer chatgpt.com person Clicked a link in a ChatGPT answer

Measure real agent visits

Create a free project, send us your traffic, and get a report of every agent that visited: who it was, whether it was verified, which pages it read, and how many people arrived from AI assistants.

How we tell a real agent from an impostor

Anyone can write "GPTBot" in a request. We only call an agent verified when we can prove it, and only call it spoofed when we can prove the opposite. Strongest proof first:

  1. Cryptographic signature

    Agents such as ChatGPT agent sign their requests (Web Bot Auth). We check the signature against the operator's own published keys.

  2. Published IP ranges

    OpenAI, Anthropic, Perplexity, Google, Apple and Microsoft publish the addresses their agents use. A request claiming their name from elsewhere is spoofed.

  3. Reverse DNS

    For search crawlers that document it, the address must resolve to the operator's domain and back.

verified
Proven to be the agent it claims.
unverified
Claims an identity we cannot check either way.
spoofed
Proven not to be the agent it claims.

For AI agents

Knocklog is an MCP server. An agent can check a site, create a project for its user, import logs and read the report without a human filling in forms.

MCP endpoint https://api.knocklog.ai/mcp

Add it to Claude Code:

claude mcp add --transport http knocklog https://api.knocklog.ai/mcp

Tools: scan_site, classify_request, list_known_agents, create_project, get_install_plan, import_logs, get_agent_traffic.

Plain-text summary for agents: /llms.txt