The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Verifa listing page.
Run identity verification from any MCP client — Claude Desktop, Cursor, Claude Code, or your own agent. Create verification sessions, hand the person a capture link, screen against sanctions and PEP lists, read results, and work the review queue, all as tools an agent can call.
That is the whole install. Get a sandbox key at app.withverifa.com → Developers → API keys
(free, no card, no sales call). Sandbox keys never touch real data and can simulate every verification outcome,
so you can see the full lifecycle in a few minutes — examples/id-check does exactly that.
This package is deliberately small. It opens one authenticated connection to Verifa's hosted MCP endpoint
(https://api.withverifa.com/mcp, streamable HTTP) and forwards tools/list, tools/call, prompts/list
and prompts/get over stdio. Nothing is reimplemented here: tool schemas, PII scrubbing, rate limits and
the audit trail all live server-side, so this wrapper never drifts from the product.
If your client already speaks streamable HTTP with OAuth 2.1, you can skip this package and connect to the hosted endpoint directly — same tools, sign-in instead of an API key.
The server exposes 45 tools, grouped into toolsets. By default you get everything except destructive.
| Toolset | What an agent can do |
|---|---|
sessions | create_session, get_session, list_sessions, list_session_events, simulate_session (sandbox), reprocess_session |
checks | list_checks, get_check, list_check_hits, rerun_check — sanctions, PEP, adverse media, watchlists |
identities | search_identities, list_identities, get_identity, tag / untag |
cases | list_cases, get_case, claim_case, assign_case, add_case_comment, approve_case, reject_case, escalate_case … |
findings | list_findings, get_finding, acknowledge_finding, dismiss_finding |
lists | blocklist and custom list reads and writes |
workflows | list_workflows, get_workflow, trigger_workflow, list_verification_policies |
org | whoami, get_usage_stats, list_org_users, list_api_keys |
search | search / fetch for ChatGPT-style connectors |
destructive | redact_session, redact_identity, bulk_redact_sessions, … — off unless you opt in |
Every tool is scoped: the API key's scopes cap what the connection can do, whatever toolsets are enabled. Every call — read or write — writes an audit row that records it was made by an agent and by which key or OAuth client, so an agent's decision is never mistaken for a human's.
Fewer tools means less context spent per request and better tool selection. Pick what the job needs:
| Variable | Effect |
|---|---|
VERIFA_API_KEY | Required. vk_sandbox_… or vk_live_…. Publishable keys (vk_pub_…) are rejected. |
VERIFA_ENV | sandbox or live. Inferred from the key if omitted; must match it if set. |
VERIFA_MCP_TOOLSETS | Comma-separated toolsets to expose. Default: all except destructive. |
VERIFA_MCP_READ_ONLY | 1 hides every tool that writes. |
VERIFA_MCP_ALLOW_DESTRUCTIVE | 1 adds the redaction tools. Each call also needs the redact:write scope, a written reason, and is capped at 5 per hour per key. Redaction is irreversible by design. |
VERIFA_API_URL | Override the API host (e.g. a staging environment). |
Check what a given configuration exposes without opening a client:
The demo creates a session, generates the one-time capture link and QR code an integrator would show,
simulates an outcome (no passport needed — sandbox only), and reads the result. It prints the REST calls
in one column and the identical flow driven through this MCP server in the other. Add --webhooks with a
public URL (cloudflared tunnel --url http://localhost:8787 works) to receive the signed webhook deliveries
and see signature verification — the part most integrations get wrong.
Node 20+. Logs go to stderr; stdout is the MCP transport.
VERIFA_API_URL.MIT — see LICENSE.