MCP Server

Connect an AI assistant to Certivant over the Model Context Protocol — triage the verification queue, send and track forms, chase stalled applicants, explain decisions, and report, all through your existing API key.

The Certivant MCP server exposes the verification platform to AI assistants over the Model Context Protocol. Point a supported host — Claude Desktop, Claude Code, Cursor, VS Code, OpenAI Codex, Gemini, or Mistral — at it, and the assistant can work your queue: see what needs attention this morning, send a verification form to a named customer and track whether they responded, chase the people who have gone quiet, explain why a case was rejected, and pull the month's compliance numbers.

The MCP server is a client of the same /api/v1/kyc REST API you already use. It holds no database connection and invents no second permission model: your API key's organization, scopes, and rate limit are exactly what the assistant gets.

At a glance


Transport	Streamable HTTP — `POST /mcp`
Default port	`3007` (`MCP_PORT`)
Upstream API	`CERTIVANT_API_BASE_URL`, default `http://localhost:3005`
Authentication	`Authorization: Bearer kyk_…` — an existing Certivant API key
Health check	`GET /health` (unauthenticated, reports this process only)
Catalogue	26 tools · 4 resources · 5 prompts

The jobs it does

Get through the day — one call returns the whole queue snapshot: what is escalated, what is stalled, which sanctions checks failed, which escalations have breached the review SLA.

Move work forward — start a verification (single or up to 50 at once), send a form to a customer by email, remind someone who stalled, renew an expiring check.

Understand a decision — read the scorer's own factor verdicts, disqualifiers, and review triggers rather than guessing at a score.

Run KYB research — read the declared business record and diff the declared directors against the public registry.

Report — approval rate, the full funnel, volume by document type, per-check failure rates, and a 30-day trend, with no row data and no personal data.

The safety model in one paragraph

Responses are verdict-shaped by default: a status, a decision, a score — never a date of birth, a document number, or raw OCR text. Anything an applicant typed or that was lifted off a document they uploaded arrives inside a labelled untrusted envelope, so the model reports on it instead of obeying it. Tools that email real customers are annotated as non-read-only so your host can confirm before firing, and the one genuinely destructive tool says so. Read the details on Safety & limits.

What is not here yet

Authentication. Every install path on this site uses a bearer API key. An OAuth 2.1 sign-in flow (Dynamic Client Registration, consent, tokens scoped to one organization) is also implemented for connector directories; the tokens it issues behave exactly like API keys.

One limit worth knowing before you start, because you will meet it:

explain_verification_decision cannot explain most older verifications. Structured scorer reasoning is only present on requests scored after it was recorded; the tool reports reasoningAvailable: false for the rest.

Every tool call is recorded in the same tamper-evident KYC audit log (kyc_audit_log) as human actions, through the REST API with the caller's own key.

Install — per-host configuration JSON, keyed to your existing API key.

Tool reference — all 26 tools grouped by job, with what each returns and what it does not.

Resources & prompts — the context the model can pull for free, and the five packaged workflows.

Safety & limits — PII tiering, the untrusted envelope, throttles and rate limits.

MCP Server

At a glance

The jobs it does

The safety model in one paragraph

What is not here yet

Next

On this page