RegBench MCP server

A stateless Model Context Protocol endpoint over Streamable HTTP at https://demo.regbench.ai/mcp. Deterministic read tools first; two write tools append to the ledger and say so in their annotations. No tool description instructs the model.

Connect

Claude Code

claude mcp add --transport http regbench https://demo.regbench.ai/mcp --header "X-RegBench-Key: rb_demo_bluebonnet"

Claude Desktop / claude.ai: Settings → Connectors → Add custom connector → URL https://demo.regbench.ai/mcp (needs a public URL, for example the Railway deployment).

curl

curl -s https://demo.regbench.ai/mcp -H "Content-Type: application/json" -H "X-RegBench-Key: rb_demo_bluebonnet" -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Credentials (prototype): one API key per demo carrier: rb_demo_bluebonnet, rb_demo_summitrange, rb_demo_harborline. Without one, calls run as the Bluebonnet demo tenant. Production: OAuth 2.1 resource server with resource indicators and enterprise-managed authorization, so the carrier's identity provider decides which agents may call which tools.

Tools

Tool	Access	What it does
regbench_check_transaction	read	Check a proposed AI-influenced decision (no record kept). Evaluates a proposed insurance decision against the rules in force in its state on its date and returns PASS, FLAG or HOLD with cited findings. Nothing is recorded; use it to test an action before taking it.
regbench_assess_transaction	write	Assess and record an AI-influenced decision. Evaluates a recorded insurance decision, stores the Regulatory Assessment, appends it to the carrier's hash-chained ledger and opens a review item when the disposition is FLAG or HOLD. Returns the assessment with its ledger position.
regbench_applicable_rules	read	List rules in force for a state, line and stage. Returns the deterministic rules that apply to a state, line of business and lifecycle stage on a date, with their authority layer, predicate, citations and effective dates.
regbench_get_provision	read	Get a provision from the corpus. Returns one legal provision by identifier: label, jurisdiction, effective window, excerpt, source URL, legal basis and verification status.
regbench_get_case_file	read	Get the case file for a decision. Returns the Exhibit C style case file for a transaction: recorded facts, the latest assessment with cited findings, point-in-time stamps, review status and ledger position.
regbench_build_exam_pack	write	Build an exam pack for a period. Assembles an exam pack aligned to the NAIC AI Risk Evaluation Supplement exhibits for the caller's carrier over a period, records its hash on the ledger and returns a summary with the pack identifier.
regbench_verify_ledger	read	Verify the carrier's ledger. Recomputes the carrier's hash chain and reports whether every link holds, the number of entries and the head hash.
regbench_benchmark	read	Cross-carrier benchmark. Returns hold and flag rates per thousand assessments by state and line across all carriers, publishing only cells with at least three contributing carriers and no dominant contributor.
regbench_resolve_insurer	read external	Look up a licensed insurer. Searches the open insurers.global register for a licensed insurer by name and returns matching records with NAIC code, domicile and class, or an explanation if the register is unreachable.

Example messages

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "initialize",
  "params": {
    "protocolVersion": "2025-06-18",
    "capabilities": {},
    "clientInfo": {
      "name": "claims-agent",
      "version": "1.0"
    }
  }
}

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/list"
}

{
  "jsonrpc": "2.0",
  "id": 3,
  "method": "tools/call",
  "params": {
    "name": "regbench_applicable_rules",
    "arguments": {
      "state": "TX",
      "line_of_business": "auto",
      "stage": "claims",
      "as_of": "2026-10-30"
    }
  }
}

{
  "jsonrpc": "2.0",
  "id": 4,
  "method": "tools/call",
  "params": {
    "name": "regbench_check_transaction",
    "arguments": {
      "transaction": {
        "txn_id": "AGENT-CLM-0001",
        "tenant_id": "bluebonnet",
        "occurred_at": "2026-10-30T16:28:19.135938+00:00",
        "stage": "claims",
        "line_of_business": "auto",
        "jurisdiction": {
          "state": "TX"
        },
        "ai_system": {
          "system_id": "fnol-fraud",
          "model_version": "3.1",
          "vendor": "third_party",
          "vendor_name": "Sentinel Analytics (fictional)",
          "high_risk_class": "fraud"
        },
        "inputs": {
          "feature_classes": [
            "claims_history",
            "geography"
          ],
          "pii_token": "tok_agent_0001",
          "data_source_refs": [
            "vendor://sentinel/fnol"
          ]
        },
        "output": {
          "decision": "deny",
          "score": 0.91
        },
        "human_involvement": {
          "reviewed": false
        },
        "adverse_action": {
          "is_adverse": true,
          "reasons_provided": false
        }
      }
    }
  }
}

Why this is safe for a model-risk team

Deterministic first. Findings come from versioned rules over recorded facts; no language model is in the path. The rules version and corpus snapshot are stamped on every result, so any result can be reproduced.
Fails closed. A citation that does not resolve to a provision in force in the transaction's state on its date withholds the assessment rather than guessing.
Reads and writes are separate tools, with annotations a gateway can use to auto-allow reads and require approval for writes.
Evidence by construction. Every recorded assessment is a ledger entry an examiner can verify without trusting RegBench.
Planned. Signed tool manifests (no silent description changes), OAuth 2.1 with enterprise-managed authorization, OpenTelemetry trace passthrough, and listing in the Anthropic, Microsoft, Salesforce and AWS catalogs.