CoverBureau — signed COI compliance checks for AI agents

CoverBureau reads an ACORD 25 certificate, checks it against the requirements you supply, and returns the result inside a cryptographically signed envelope. It is a compliance-checking service: it reports what a certificate says and whether it meets your requirements. It does not make, file, advise on, or publish insurance rates.

Calling it

The service is a remote MCP server. Send JSON-RPC 2.0 over HTTP POST to https://coverbureau.com/mcp. An unpaid call returns 402 with a payment challenge; pay it and retry the same request. This is the standard x402 flow, so any x402-capable client handles it for you.

POST https://coverbureau.com/mcp

{"jsonrpc":"2.0","id":1,"method":"tools/call",
 "params":{
   "name":"check_carrier",
   "arguments":{
     "carrierName":"Federal Insurance Company",
     "state":"NY"
   }
 }}

Without payment that returns 402 and a challenge naming the price, the asset, and the recipient:

{"x402Version":2,
 "accepts":[{
   "scheme":"exact",
   "network":"eip155:8453",
   "amount":"10000",
   "asset":"0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913",
   "extra":{"name":"USD Coin","version":"2"}
 }]}

Settlement is USDC on Base mainnet. There is no signup, no API key, and no minimum: an agent that can pay can call.

Tools

Which one to call. extract_coi gives you the fields and no judgment — use it when you will apply your own rules. validate_coi reads the certificate and judges it against requirements you supply; it extracts as part of its work, so calling both on one document pays for the same read twice. check_carrier needs no document at all — use it when you have a carrier but no certificate, since validate_coi already checks the carrier when you do. A fourth tool, screen_coi_plausibility, is specced but not built: validate_coi judges a certificate against your requirements, while that one would screen it against market reality — neither subsumes the other.

extract_coi $0.02 / call

Reads a certificate into structured data. Extraction only, no validation — for callers that just want the fields: insured, producer, carriers, policy numbers, coverage limits, effective and expiration dates, and the description of operations. Use it when you will apply your own rules; for a verdict against requirements you supply, use validate_coi instead.

validate_coi $0.075 / call

Reads a certificate and judges it against your requirement profile, check by check. Every failure states what was required, what the certificate showed, and why it fails: enough to write the request for a corrected certificate without opening the PDF.

Every required coverage present and in force on the date you specify (default today), with policies expiring inside your notice window flagged before they lapse
Limits at or above your minimums: each occurrence, general aggregate, products and completed operations, combined single limit, employers liability
General liability written on an occurrence basis rather than claims-made
Additional insured and waiver of subrogation, with the evidence source recorded: the ADDL INSR / SUBR WVD column, or endorsement wording in the Description of Operations (CG 20 10, CG 20 37, CG 24 04)
Named insured matches the entity you contracted with; certificate holder matches yours
Each carrier looked up by NAIC number or name against a curated dataset, with its financial-strength rating compared to your threshold. Carriers outside the dataset come back as review rather than a guessed pass, and every response reports the dataset's status.
ACORD edition checked, older editions flagged

Anything the service cannot read cleanly comes back as review rather than a guessed pass: a blank limit, an illegible scan, a coverage row pointing at an insurer letter the certificate never lists. Review is a verdict, not an error. The overall result carries a risk score computed from fixed weights, not a model's opinion, and the whole verdict is signed. The verdict also records the evaluation date it used, so an as-of result cannot be mistaken for “valid today”.

Use extract_coi instead if you only want the fields without a verdict. Use check_carrier only when you have no certificate — as the carrier bullet above shows, this tool already checks the carrier when your profile sets a carrier requirement.

check_carrier $0.01 / call

Given an NAIC number or carrier name plus a US state, returns recognition status, financial-strength rating, and whether the carrier meets a minimum rating threshold. It needs no document. Use it when you have a carrier but no certificate — vetting an insurer named in an email, or before requesting a COI. If you do have a certificate, validate_coi already checks the carrier, so calling both pays twice for the same lookup.

Verifying a response signature

Every response comes back inside a signed envelope: the payload, an Ed25519 signature over the canonical JSON of that payload, the key id, and an issued-at timestamp. Fetch the public key and check the signature yourself — verification needs neither trust in CoverBureau nor a second call.

Public key: https://coverbureau.com/.well-known/jwks.json
Canonicalization: RFC 8785 (JCS)

A signed verdict is a durable artifact. If a compliance decision is ever questioned, the envelope shows exactly what was returned, for which document, and when — verifiable by anyone holding the public key, including someone who does not trust us.

What to know before relying on it

Decisions are made by code, not by a model. A vision model reads the document; a deterministic rules engine makes the pass/fail determination and computes the risk score. The same inputs give the same verdict, and a document cannot talk the system into a different answer.
Low confidence becomes review, never a guess. When a reading is uncertain or a carrier is unknown, the verdict routes to human review rather than inventing an answer.
Carrier data is a curated seed dataset. NAIC numbers, ratings, and licensing are best-effort and may be stale or wrong. Every check_carrier response reports its datasetStatus, so the status of the data travels with the answer — and an adverse conclusion drawn from unverified data routes to review rather than returning a hard failure.
CoverBureau checks what a certificate says and whether it meets your requirements. It does not assess whether the document is genuine. That is a separate discipline, document forensics, and a well-made fake that satisfies your requirements will pass here. If artifact authenticity matters for your use case, run a document-authenticity tool alongside it.
Accuracy is gated, not advertised as a headline. A release cannot ship unless every material extraction field scores at least 95% against our own hand-labelled golden set, and unless, on severely degraded documents, every material field is either correct or carries low self-reported confidence — a misread on an illegible document must be flagged, never silent. That corpus is internal, not an independent benchmark.
Scope is US ACORD 25. Other ACORD forms and non-US jurisdictions are out of scope.

For agents

A machine-readable summary of this service is published at /llms.txt.