I'm trying this because a growing share of our prospects start by asking their AI of choice what coverage to buy and who to get it from. Assistants can read my site, so like everyone else I'm trying to get recommendations via content. But the next natural step is agents actually taking actions on sites rather than just reading them, so I created seven tools, six read-only (a verified-facts knowledge base, carrier comparison, specialty and education guides, benefit-cap math, rider definitions) and one action, "quote request," which writes a lead straight into our CRM through the same Salesforce web-to-lead path that our web form uses. Then it switches back to the "normal" process and a licensed broker follows up through traditional means. This is the best use case I could figure out at this point, but expect to find others as the space matures. To my knowledge it is the first agent-callable quote action in my corner of insurance.
I had originally built it with an OAuth flow, but there was friction that kept killing the test interactions so I made it open, no auth. Unless there is something I'm missing the worst case scenario is junk leads and not a data leak since the read tools only contain published information and the quote request collects normal lead stuff like name, DOB, etc. According to Claude, spam "gets handled in layers with strict input validation including a DNS-over-HTTPS check that the email domain can receive mail, per-IP rate limiting in Workers KV, and short-window duplicate suppression." Either way every agent-originated lead is tagged in my CRM so junk is easily filterable. The rate limit should in theory block you if you hammer it. But this is the sort of stuff I wouldn't mind some people playing around with since I'm learning on the fly and an experienced developer could probably find some holes.
My setup is a single Cloudflare Worker (Streamable HTTP transport), KV for rate limiting and knowledge, knowledge compiled at deploy from the same source of truth that feeds our site chatbot, so the MCP and the site regenerate from the same file.
Endpoint: https://mcp.seaworthy.io/mcp Human-readable reference: https://seaworthy.io/for-ai-agents/
The honest open question is whether or not this will ever accomplish anything. I have not yet received a lead through these means, and do not expect to for some time yet, but I do believe that rather than people asking agents where to get quotes, they will eventually just ask the agent to get the quotes on their behalf (and then eventually to just fully set up the coverage), and I want to be the first one in my niche who has the capability. Happy to answer questions, and would be open to input if anyone has anything to share.