@mobilevalidate/mcp

@mobilevalidate/mcp

Model Context Protocol server for MobileValidate. It lets AI agents (Claude Desktop, Claude Code, Cursor, VS Code and any other MCP client) check phone numbers — registration on WhatsApp, Telegram, Viber, Signal and other platforms, carrier and line type, report-based spam reputation — and e-mail addresses (mailbox exists; account on Gmail, Outlook, Apple and others; yes / no / unknown only, never names or profiles). Several services can be checked in one call (checks); list_services shows what the key can use.

It is a thin client of the public MobileValidate API, built on the mobilevalidate SDK. It holds no credentials of its own: your key is only forwarded to api.mobilevalidate.com.

Two ways to use it:

	Hosted (remote)	Local (stdio)
Runs	at https://mcp.mobilevalidate.com/mcp (Streamable HTTP)	on your machine via npx -y @mobilevalidate/mcp
Key	header Authorization: Bearer <key>	env MOBILEVALIDATE_API_KEY
Needs	nothing to install	Node.js ≥ 20 (npx downloads the package)

Both run the same tools with the same spend safeguards. Try either with a test key: the documented test values return fixed answers and nothing is billed.

Keys: an agent key (mv_agent_…, scoped, daily spend cap) or a test key (mv_test_…, free, never billed). Live keys (mv_live_…) are refused by the MCP server on purpose. Get a personal test key at https://mobilevalidate.com/get-test-key; agent keys come with an approved account.

Docs: https://mobilevalidate.com/docs/mcp

Quick start (hosted, test key)

export MOBILEVALIDATE_API_KEY=mv_test_...     # your personal test key
claude mcp add --transport http mobilevalidate https://mcp.mobilevalidate.com/mcp \
  --header "Authorization: Bearer $MOBILEVALIDATE_API_KEY"

Then ask: "Check [phone], [phone] and [phone] on WhatsApp." You get registered, not registered and unknown. Test keys never reach a real network and are never billed: the magic values below give fixed answers, and any other number gets a made-up but stable answer.

Just looking? The public sandbox key mv_test_publicSandboxn9ZgneuhR1B9CRfKG3fulym works too, but it answers only the magic values below (anything else is refused with sandbox_magic_only), has low shared rate limits and cannot run large jobs. Use your own test key for anything more.

Test number	Result
[phone]	registered
[phone]	not registered
[phone]	unknown (registered: null)
[phone]	pending for about 5 s, then registered
[phone]	unsupported_country
[phone]	registered, business account

Test e-mail addresses: [e-mail], not-registered@…, unknown@… (same domain). Full list: https://mobilevalidate.com/docs/test-values.

Client configuration

Claude Code

# hosted
claude mcp add --transport http mobilevalidate https://mcp.mobilevalidate.com/mcp \
  --header "Authorization: Bearer $MOBILEVALIDATE_API_KEY"
# local stdio
claude mcp add mobilevalidate --env MOBILEVALIDATE_API_KEY=mv_agent_... -- npx -y @mobilevalidate/mcp

.mcp.json equivalent of the hosted setup:

{
  "mcpServers": {
    "mobilevalidate": {
      "type": "http",
      "url": "https://mcp.mobilevalidate.com/mcp",
      "headers": { "Authorization": "Bearer ${MOBILEVALIDATE_API_KEY}" }
    }
  }
}

Cursor (~/.cursor/mcp.json or .cursor/mcp.json)

Hosted (live):

{
  "mcpServers": {
    "mobilevalidate": {
      "url": "https://mcp.mobilevalidate.com/mcp",
      "headers": { "Authorization": "Bearer ${env:MOBILEVALIDATE_API_KEY}" }
    }
  }
}

Local stdio: { "mcpServers": { "mobilevalidate": { "command": "npx", "args": ["-y", "@mobilevalidate/mcp"], "env": { "MOBILEVALIDATE_API_KEY": "mv_agent_..." } } } }

VS Code (.vscode/mcp.json)

{
  "servers": {
    "mobilevalidate": {
      "type": "http",
      "url": "https://mcp.mobilevalidate.com/mcp",
      "headers": { "Authorization": "Bearer ${input:mobilevalidate-key}" }
    }
  },
  "inputs": [
    { "id": "mobilevalidate-key", "type": "promptString", "description": "MobileValidate agent or test key", "password": true }
  ]
}

Claude Desktop (claude_desktop_config.json)

Claude Desktop connects to remote servers with a key through a stdio bridge such as the community mcp-remote package (hosted server, live):

{
  "mcpServers": {
    "mobilevalidate": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.mobilevalidate.com/mcp", "--header", "Authorization:Bearer ${MOBILEVALIDATE_API_KEY}"],
      "env": { "MOBILEVALIDATE_API_KEY": "mv_agent_..." }
    }
  }
}

Local stdio:

{
  "mcpServers": {
    "mobilevalidate": {
      "command": "npx",
      "args": ["-y", "@mobilevalidate/mcp"],
      "env": { "MOBILEVALIDATE_API_KEY": "mv_agent_..." }
    }
  }
}

Any other MCP client

• Streamable HTTP: URL https://mcp.mobilevalidate.com/mcp, header Authorization: Bearer mv_agent_…. The server is stateless: POST only, JSON responses, no session id.
• stdio: command npx, args ["-y", "@mobilevalidate/mcp"], env MOBILEVALIDATE_API_KEY=mv_agent_…. To pin a version use @mobilevalidate/mcp@1.2.0. After npm install -g @mobilevalidate/mcp the command is mobilevalidate-mcp.

Protocol versions

Both transports speak MCP 2026-07-28 and the handshake-based revisions 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05 and 2024-10-07 (from 1.1.0; 1.0.x stopped at 2025-11-25). Nothing to configure: the client's opening message picks the revision.

• 2026-07-28: no initialize; every request carries _meta (io.modelcontextprotocol/protocolVersion, client info and capabilities) and, over HTTP, the MCP-Protocol-Version, Mcp-Method and Mcp-Name headers (checked against the body; mismatch → 400, -32020). server/discover returns the supported versions, capabilities, server identity and instructions. tools/list and server/discover carry ttlMs: 3600000, cacheScope: "public" (the list is the same for every key), and tools come back in a fixed order. An unsupported version gets 400 with UnsupportedProtocolVersion (-32022) listing the supported ones. Tool calls return JSON (no SSE).
• 2025-era clients: initialize as before; the hosted server stays stateless (POST only, JSON responses, no session id; GET → 405). On stdio the connection stays on the negotiated revision.
• The server uses no server→client requests (sampling, elicitation, roots) and no protocol logging, so nothing deprecated in 2026-07-28 is involved. Bulk jobs use an explicit job_id handle, the pattern the spec recommends now that sessions are gone.
• The server card's supportedProtocolVersions lists the same versions (newest first).

Tools

Tool	What it does	Spends credits	Annotations
normalize_numbers	Local, free E.164 formatting. Flags ambiguous inputs (no country) and duplicates	no	read-only, idempotent, closed-world
estimate_cost	Free pre-flight (numbers and/or emails): valid, invalid, duplicate and cached counts, plus the maximum cost of a bulk job	no	read-only, idempotent
lookup_numbers	Check up to 100 numbers (optionally with emails, 100 in total) for one or more real-time services (checks; waits up to wait_seconds, default 20)	yes	not read-only, idempotent, open-world
lookup_emails	Check up to 100 e-mail addresses (default check email; e.g. apple.email) in real time	yes	not read-only, idempotent, open-world
check_spam_reputation	Spam reputation of up to 100 numbers (runs number.spam only): per number risk_level, risk_score, reasons, top category, first/last seen, sources, plus counts per level	yes	not read-only, idempotent, open-world
create_lookup_job	Bulk check (up to 50,000 numbers and/or emails) for any active service, including bulk-only ones (e.g. gmail). Returns job_id	yes	not read-only, not idempotent (retries of the same arguments are deduplicated)
get_lookup_job	Job status plus a filtered (registered, service), paginated page of results (one item per number or e-mail)	no	read-only, idempotent
list_services	Services the key can use: input type (phone/email), real time or bulk only, attributes, countries, prices	no	read-only, idempotent
get_account	Balance, reserved credit, today's usage, limits	no	read-only, idempotent

Every tool declares a title, annotations, inputSchema and outputSchema. Each one returns structuredContent plus a one-line text summary. checks takes service codes or aliases (e.g. whatsapp, telegram, viber, carrier, spam); its description lists the real-time and bulk-only codes known to this build (from the SDK's generated catalog), and list_services gives the live list. Result items keep the v1 top-level fields (WhatsApp if requested, otherwise the first service) and add a checks map per service when more than one service (or a non-WhatsApp service) was requested. Tool descriptions state the anti-enumeration limits (≥ 20 consecutive numbers refused, daily caps).

E-mails. Phone services run on numbers, e-mail services on emails; send at least one check of each kind you send. Result items carry kind (phone/email); e-mail items have email (normalized, null if invalid), email_status (valid, invalid_email, duplicate, suppressed) and number_status: null. Descriptions state the e-mail rules: ≥ 20 addresses on one domain differing only by digits are refused as enumeration, and answers are yes/no/unknown only — no names, photos or profiles. Test keys use [e-mail], not-registered@…, unknown@… (same domain).

Spam reputation. check_spam_reputation (or checks: ["spam"] on the other tools) answers from spam and nuisance-call reports: regulator actions, government complaint data, community reports and recently-unassigned numbers. All countries except sanctioned ones (Cuba, Iran, North Korea, Syria, Russia, Belarus, Venezuela and sanctioned regions → unsupported_country, free). no_reports means no reports are known — not that the number is safe; it is still a conclusive, billed answer. Limits: ≤ 100 numbers per call, the same spend confirmation and anti-enumeration rules as lookup_numbers. Test keys: [phone] high, …002 no_reports, …003 unknown, …004 pending then medium, …005 unsupported_country. Attribute values may be integers (risk_score, sources).

Network checks. checks: ["hlr"] returns live network status (status reachable / unreachable / invalid, ported, roaming, mcc_mnc, country); checks: ["mnp"] returns number portability (porting ported / not_ported, mcc_mnc, country). Unknown answers are free.

Spend safety

Before lookup_numbers, lookup_emails, check_spam_reputation or create_lookup_job spends anything, the server calls the free estimate. Real-time tools quote the maximum at the key's real-time prices (from list_services), create_lookup_job at bulk prices. The call is refused with a confirmation_required error result when either of these is true:

• the maximum cost is above MCP_CONFIRM_ABOVE_USD (default 1.00), or
• a job has more than 100 numbers and e-mails.

Scopes (agent keys): the real-time tools need lookup:write and jobs:write (the estimate is POST /v1/jobs/estimate); estimate_cost and create_lookup_job need jobs:write; get_lookup_job needs jobs:read; get_account needs account:read. list_services needs none.

The error text states the amount and tells the agent to ask the user. The agent then re-invokes the tool with confirm_max_cost set to the amount shown. The confirmed (or estimated) amount is always sent to the API as max_cost, so the API refuses any request that would cost more.

Limitation: this confirmation is agent-mediated. The server cannot prove that a human approved the amount. The hard limits are the key's own spend cap and the max_cost guard. A human confirmation URL is planned.

Other rules:

• Only mv_agent_… (scoped, spend-capped) and mv_test_… keys are accepted. mv_live_… keys are rejected with a clear message.
• There is no webhook_url parameter, and request metadata is never echoed into tool output.
• Logs go to stderr and contain counts and ids only. Anything that looks like a phone number or an e-mail address is masked.

Environment variables

Env	Default	Used by
MOBILEVALIDATE_API_KEY	none (required)	stdio: the mv_agent_/mv_test_ key
MOBILEVALIDATE_BASE_URL	https://api.mobilevalidate.com	both.
MCP_CONFIRM_ABOVE_USD	1.00	both
MCP_HOST / MCP_PORT	127.0.0.1 / 3300	HTTP
MCP_ALLOWED_HOSTS	none	HTTP: extra Host values allowed (comma-separated), e.g. behind a reverse proxy
MCP_LOG	on	set off to silence stderr logs
MV_MCP_FORWARD_SECRET	none	HTTP, hosted deployment only: secret shared with the API so sandbox-key calls are rate limited per end client, not per server
MCP_CLIENT_IP_HEADER	cf-connecting-ip	HTTP: header with the end client's IP, trusted only from a loopback peer (tunnel/proxy)

Self-hosting the HTTP transport

MCP_HOST=127.0.0.1 MCP_PORT=3300 npx -y -p @mobilevalidate/mcp mobilevalidate-mcp-http   # once published

POST /mcp only (stateless, JSON responses, 2026-07-28 and 2025-era clients on the same endpoint) and GET /healthz. Each request's Authorization: Bearer key is forwarded to the API for that request only and never stored. Host must be 127.0.0.1:<port> or localhost:<port>, plus any value in MCP_ALLOWED_HOSTS (DNS-rebinding protection) — set it when you put the server behind a reverse proxy, and terminate TLS in front of it. Bodies are limited to 2 MB.

Security notes

• Use an agent key (mv_agent_…) with a spend cap you are comfortable with; the key's cap and the max_cost guard are the hard limits. Test keys are free.
• Put the key in the client's env / header configuration, not in prompts. The server never logs keys; its stderr logs contain counts and ids only, and anything that looks like a phone number or e-mail address is masked.
• Spending above MCP_CONFIRM_ABOVE_USD requires the agent to ask the user first (see Spend safety); this is agent-mediated, so keep the key's spend cap as the real limit.
• Only check numbers and addresses the user has a legitimate relationship with. Enumeration patterns are refused and daily caps apply.

Maintainers

Source is TypeScript in src/ (run directly with tsx inside the monorepo: pnpm --filter @mobilevalidate/mcp start for HTTP on 127.0.0.1:3300, start:stdio for stdio). The published package contains only the compiled dist/, README.md, LICENSE and server.json; publishConfig swaps main / exports / bin from src/ to dist/ and pnpm rewrites the workspace:* SDK dependency to the exact SDK version — so always pack with pnpm (ops/publish-packages.sh), never plain npm pack / npm publish from this directory. The build compiles against the SDK's dist/ types, so build the SDK first.

MCP SDK: the server uses the official v2 TypeScript SDK (@modelcontextprotocol/server, which implements 2026-07-28 with dual-era serving: createMcpHandler + isLegacyRequest for HTTP, serveStdio for stdio). The v1 package @modelcontextprotocol/sdk and @modelcontextprotocol/client are dev dependencies only: the tests use them as a real 2025-11-25 client and a real 2026-07-28 client (test/protocol.test.ts).

pnpm --filter @mobilevalidate/mcp run typecheck
pnpm --filter @mobilevalidate/mcp test
pnpm --filter mobilevalidate run build && pnpm --filter @mobilevalidate/mcp run build   # → dist/
ops/publish-packages.sh                     # build, test, pack, leak-scan, npm publish --dry-run (SDK, then MCP)
NPM_TOKEN=… ops/publish-packages.sh --yes   # real publish

Release checklist: bump version in package.json, SERVER_VERSION in src/tools.ts and both version fields in server.json (a test enforces they match). The SDK must be published first, because this package pins its exact version.

MCP Registry. Listed as com.mobilevalidate/mcp (domain-verified namespace). server.json follows the registry schema 2025-12-11; mcpName in package.json must equal its name (the registry checks it on npm).

License

MIT © 2026 BroadNet Technologies Inc. See LICENSE.