Sigo Seguros MCP server: your AI agent can now buy insurance — but who says yes?

On October 1, 2026, Sigo Seguros (Austin, Texas) became the first personal auto insurance agency to let an AI agent — ChatGPT, Grok Bot, Muse — take a Texas driver from first question to purchased policy inside the conversation. The agent requests estimates from multiple carriers, compares options, and completes the purchase on the spot. Sigo also launched InsuranceMCP.com (live, HTTP 200 verified at launch) as an open directory where any insurer or agency can list its MCP server. Every article in today's coverage answers the same question — what launched. None answers the one that matters for a machine that can now spend your money: which judgment lets the agent click "buy" on a binding insurance contract? The machine-native answer is the decision gate: ≥0.80 auto-pay, 0.50–0.79 human confirm, <0.50 block, log, and escalate.

What actually launched (receipts, dated)

Reading the fine print: the PR vs the product page

Sigo's own MCP product page (found during this run, ~20:50 EDT) is more precise than the press release: the assistant "doesn't pick a policy, take payment or issue anything: that stays with you and a licensed agent." The in-conversation flow ends with a link to finish on WhatsApp or the web — the server URL is https://mcp.sigoseguros.com/mcp (Streamable HTTP, no auth), with tools list_available_states, next_question, start_quote, get_quote; it quotes in Texas, California, Florida, and Arizona. So "completes the purchase on the spot" means the agent completes the quote-to-buy journey inside the chat and hands the binding step to Sigo's licensed flow. That handoff is the confirm band in practice — and it's exactly where a decision gate belongs: the agent does the shopping, the licensed/human step does the authorizing. PR copy oversold the autonomy; the architecture underneath is the right one.

What the field lacks (every one of them)

The October 1 field is PR-wire clones: PR Newswire, Morningstar, and lifestyle-wire reprints. Zero builder guidance. Zero authorization analysis. Zero code. And they all miss the structural fact: this is the first true agentic-commerce purchase — not a $0.03 API micropayment, but a binding contract with a recurring monthly premium. The quote is data; the purchase is a decision. Sigo solved the information problem it named. Nobody solved the authorization problem.

The gate mapping: quote is data, purchase is a decision

The confirm band is the whole product for agentic commerce: the agent does the shopping, the human does the authorizing. Sigo's in-conversation purchase flow already keeps the human in the chat — the gate just formalizes when the human must tap.

Live gate receipts (minted today, not hypotheticals)

Scored against the live endpoint POST https://scriptmasterlabs.com/api/harness/decide — local-heuristic-v1, gate bands 0.80/0.50 live on the endpoint, October 1, 2026 ~20:30 EDT. Exact shape: {"state":"baseline","questions":[{"id","type":"score","scale":[0,1],"question"}]}.

Honest finding: the uncalibrated heuristic scores both scenarios identically (0.35) — it cannot discriminate the dangerous auto-buy from the safe present-and-pause pattern. That is the blunt-instrument reality: the fail-closed ceiling is the protection, and the calibrated decider (a JEV-27B-class model supplying per-option calibrated probabilities — see our JEV-27B breakdown) is the upgrade. The plumbing runs; discrimination is the roadmap. The endpoint reports calibrated=false, typesafe_wired=false — we print what it says.

Do it yourself: put a gate in front of any MCP purchase tool (5 steps)

1. Split the tool surface. Your MCP server exposes quote (read-only, auto-runs) and purchase (money moves, gated). Never let one tool do both.
2. Score the purchase, not the product. Send the purchase intent to a decision endpoint (POST /api/harness/decide, type: score, scale: [0,1]) with the full context: amount, recurrence, whether the human reviewed limits/deductibles/disclosures.
3. Wire the bands. ≥0.80: execute. 0.50–0.79: render the confirmation card in-conversation and wait for the tap. <0.50: block, log the full context, escalate to the human with the reason.
4. Fail closed. Endpoint down, score missing, heuristic uncalibrated? Default to escalate. Today's receipts prove our own gate does this — both scenarios blocked + logged.
5. Log like an auditor. Every purchase attempt keeps: the score, the band, the disclosure state (limits/deductibles/eligibility seen by whom), and the human's confirmation record. Insurance regulators will ask; receipts are the answer.

Claim receipts

Every factual claim on this page is minted as a Claim receipt (H2#claim-receipts + Claim JSON-LD above). Verified against the October 1, 2026 PR Newswire release and Morningstar syndication, Sigo's own MCP product page (sigoseguros.com/en/mcp/, found ~20:50 EDT — its "doesn't pick a policy, take payment or issue anything" line corrects the PR's "completes the purchase on the spot" framing), a live HTTP check of insurancemcp.com (~20:35 EDT), and two live harness receipts minted October 1, 2026 (~20:30 EDT). Caveats: launch facts are Sigo's self-reported PR claims (the "first agency" framing is theirs, not independently audited); the directory was verified live at launch only; the live gate uses an uncalibrated heuristic (calibrated=false, typesafe_wired=false) — today's receipts prove the plumbing runs and the fail-closed property holds, not that the heuristic discriminates well.

Published 2026-10-01 by ScriptMasterLabs — the x402/MCP/AI agent pedia. Pattern priority: decision-gated machine payments. Related: the gate pattern · JEV-27B open decision model · x402 agent spending guard · Mastercard agent-pay trust intelligence · XRPL 10M x402 payments. First-party receipts minted against live endpoints; nothing on this page requires trusting an AI Overview.