# Your rating engine, callable by AI agents
> Rating AI puts Sure's filed raters behind a Model Context Protocol server. Claude, GPT, Gemini, or any MCP-capable client can discover ratable products, fetch the input schema, get a priced quote, and hand the user a real purchase link. The filings stay where they are.
Canonical URL: https://www.sureapp.com/platform/rating-ai
Module slug: rating-ai
Category: Platform
## A rater that agents already know how to call

MCP is the protocol LLM runtimes already speak. Sure's rating server speaks it natively, with auth scoped to your product and environment.
### MCP-native
Mount the Sure MCP server in Claude Desktop, Claude Code, Cursor, or any MCP-compliant runtime. Your ratable products show up as callable tools in one config block.
### Versioned raters
Every price call binds to a released rater version. The rater that priced the quote today will price the same inputs the same way tomorrow, with the version id on the response.
### Schema-first pricing
Agents call get-rating-schema before pricing. They get field names, types, required flags, defaults, and a field_count hint that tells the agent whether the product is chat-friendly or form-grade.
### Scoped authority
Credentials are namespaced per partner and per environment. Tokens come from a secrets vault and cache for thirty minutes, so an agent only sees what its namespace is allowed to see.
### Factor transparency
Quote responses include a full step-by-step breakdown of how the price was calculated, with every rating step, scope, discount, and attribute-level adjustment itemized. The agent returns a price with the math to back it up.
### Every call, logged
Every tool call is logged with the caller, the request, and the response. Trace a quote back to its inputs, spot the calls that misbehaved, and act on what you find.

## Three kinds of agent, one protocol

One MCP server, three audiences. Internal agents, partner agents, and third-party carriers all call the same tools with scoped credentials.
### The agents we ship (Internal agents)
UW triage, renewal, FNOL, and conversational quoting bots all call the same MCP server. They share wire protocol, auth, and logging with anything you build.
- Namespaced to your partner scope
- Use the same rater release train
- Full structured logs per tool call
### Your agents, your LLM (Partner agents)
Bring Claude, GPT, Gemini, Llama, anything with an MCP client. Configure the Sure server, hand your agent a namespace token, and the product catalog becomes a set of tools it can reason over.
- Claude Desktop and Claude Code
- Cursor and the OpenAI Agents SDK
- Any MCP-compliant client
### Carriers and program partners (Third-party agents)
A carrier or reinsurer can call the same rating tools against your released rater, price the same risk on their side, and reconcile the number. Read-only access, same audit log.
- Scoped partner credentials
- Same rater version on both sides
- Cross-check without filing access

## Agents get a pricing tool

The MCP server reads released raters, returns versioned prices, and logs every call. Authoring and binding stay behind human review.
### 01. Read-only on the rater
The MCP tools price against a released rater. Nothing in Rating AI edits a rate table, a ROC, or a filing. Authoring happens elsewhere, behind human review.
### 02. Rater-versioned responses
Every quick-price response carries a rater_version_id. Replay the same inputs against the same version and you get the same number.
### 03. Namespaced auth
Tokens are fetched per environment and namespace from a secrets vault. An agent credentialed for partner-scope cannot call anything outside that scope, and there is no cross-environment leakage.
### 04. Quote sessions only
start-quote-session creates a real quote and returns a redirect url. The human completes purchase in the flow.
### 05. Graceful partial pricing
When required fields are missing, get-quick-price returns a lowest-valid price with as_low_as=true. The agent can say "prices as low as $X" without inventing inputs.
### 06. Structured-logged, generic errors
Tool errors return a generic user-facing message and a structured log line on our side. The agent never leaks stack traces, and the trail is preserved for audit.

## What the protocol guarantees

- Protocol: Model Context Protocol over HTTP streaming, available to partners. Learn more.
- Tools exposed: List products · fetch schema · get a quick price · start a quote session
- Versioning: Every response pins the rater version that priced it
- Auth: Per-namespace bearer tokens from a secrets vault, 30-minute cache
- Environments: Local, test, and production environments available. Selectable per tool call.
- Pricing detail: Total plus a step-by-step breakdown and discounts by attribute
- Clients tested: Claude Desktop, Claude Code, Cursor, OpenAI Agents SDK, custom MCP clients
- Audit: Full call log per agent invocation, retained per workspace

## Customer quote

"Chubb's insurance offering, through Sure's technology platform, further reinforces our commitment to innovative, digital solutions and providing consumers with peace of mind."
— James Walloga, EVP Accident & Health, Chubb North America

## See the platform up close

Get a walkthrough with the team at Sure. We'll dig into the platform and how we impact your roadmap.
Primary action: Book a walkthrough