Skip to content

MeiSIM MCP

An agent that needs a phone number hits the same wall every time: the sign-up form wants a mobile number, and the number it can find is a virtual one that the site already blocks. The other direction is worse. A business wants its agent to receive a text from a real customer, and a shared inbox is not a line.

MeiSIM is a US telecom reseller that put its catalogue behind an MCP server. The catalogue is two things at once: real US mobile lines on a national carrier, and travel eSIM data plans from several vendors. The part worth a page is not the price list. It is that this one states its own gate plainly, in the same response that sells the product: a person opens the account, and the server says an agent cannot open one by itself.

Server type:   remote Streamable HTTP (hosted)
Endpoint:      https://mcp.meisimusa.com/mcp
Server:        meisimusa-mcp 0.3.5, protocol 2025-06-18
Auth:          none for the catalogue; Bearer API key for the account tools
Tools:         16 (12 read only, 4 that spend money or send a text)
Listed on:     mcp.so feed, registry io.github.Shauliko/meisimusa-mcp
Local route:   npx -y meisimusa-mcp with MEISIMUSA_API_KEY

The evidence tier

Keyless and driven, with the boundary measured rather than assumed.

initialize answered 200 with no credential, no session id issued, and serverInfo meisimusa-mcp 0.3.5 on protocol 2025-06-18. tools/list returned sixteen schemas. Twelve tools/call requests were then run: six returned live data without a key, and six returned the server's own refusal text.

That split is the useful finding of the pass. The catalogue is public. The account is not.

The tool surface

Every tool carries a title, an output schema, annotations and an execution.taskSupport field, which is more of the current spec than most entries here declare.

Tool Argument What it returns
get_ai_line_offer none The current offer for lines sold to agents: monthly price for texts, for texts plus voice, volume tiers, minimum term, delivery times
search_us_numbers max_price_usd Prepaid plans that come with a real US number: carrier, voice allowance, SMS allowance, price, term
search_plans country_code, duration_days, data_gb, sort_by Travel eSIM plans for a country: data, days, price, vendor, coverage
get_quote plan_id The plan plus a price breakdown and a short lived quote
purchase plan_id, customer_email, payment_method, billing_country, idempotency_key A hosted Stripe payment link, then the eSIM and its QR emailed to the buyer
get_activation order_id Install QR, activation (LPA) string, status
topup_lookup line_type, phone or carrier, or iccid or email Existing lines and the top-up products available for them
topup_start line_type, product_id, customer_email Order id and a payment URL for a recharge
get_account none Company, status, prepaid wallet balance, how many numbers may be activated today
get_number_pricing none What a new number costs on this account: daily rate, minimum days, activation fee, headroom
buy_numbers count (1 to 50), billing, days or months, voice, line_history, area_code, zip_code Numbers ordered from the prepaid wallet
get_order order Provisioning status and the numbers delivered
list_my_numbers search, limit, offset The numbers on the account with status, voice, billing and paid through date
read_messages number, since, limit Inbound texts, newest first
wait_for_message number, since, timeout_seconds (max 50) Blocks until the next text arrives and returns it
send_message from, to, message (max 800 characters) Sends a text from one of the account's numbers

wait_for_message is the connective tissue an agent actually needs. A verification flow is not a read, it is a wait: send, then hold the line for the code. Every other tool here is a lookup or a purchase.

What the offer is, in numbers

get_ai_line_offer answered without a key and priced the product in one response:

texts only          35 USD per line per month
texts and voice     45 USD per line per month
10 lines            32 or 42
50 lines            29 or 39
200 lines           25 or 35
minimum term        2 months, paid up front, then monthly from the balance

The notes attached to that figure are the ones that matter for an integration:

  • A real US mobile number on a national carrier, not a VoIP number. That distinction is the whole product. A VoIP number fails the verification step this is bought to pass.
  • Inbound texts arrive at a webhook or over the API. Voice runs over SIP with the number as caller ID, one call at a time per line, and outbound calls to US numbers only.
  • From stock, a number is delivered in seconds; a new number takes minutes.
  • An AI agent cannot open the account by itself. The sign-up form carries a robot check, and the ID step needs a government ID plus a phone camera, or a review by a person at the vendor.

That last line is not a defect. It is the compliance posture of a telecom regulator being passed through to the client, and it is stated in the refusal text of every account tool rather than buried in a policy page. A team planning to let an agent buy its own identity should read the sentence before it builds the flow.

The catalogue, measured

search_us_numbers with a price ceiling of 1000 returned 28 plans across five carriers (AT&T Prepaid, LinkUp Mobile, MobileX, Moxee 2, T-Mobile), from 7.50 for a 30 day plan with 100 SMS and no data, through 10.00 for 100 minutes and 100 SMS, to 288. Every row carried includes_us_number: true, which is the field that makes the list relevant here.

The price ceiling works, which is not universal in this category:

max_price_usd 10    2 plans, highest price 10.00
max_price_usd 25    9 plans, highest price 21.00
max_price_usd 1000  28 plans

No duplicate plan_id values, no duplicate carrier and title pairs. The filter is ascending by price and it holds.

get_quote on the cheapest plan returned its own arithmetic, clean and addable:

subtotal 7.50   taxes 0   fees 0   final 7.50 USD
quote_expires_at  2026-11-17T18:17:36.155Z

The quote window is 15 minutes, measured twice. A plan bought from a stale quote is the failure this field prevents, so a purchase flow should re-quote rather than reuse an id from a previous turn.

The eSIM side is an aggregator

search_plans for the same country returns plans from several suppliers, not only the house brand. Sampling Japan, France and the United States at 30 days and 5 GB gave 58, 80 and 86 plans, with vendors including Airalo, Sparks, 3, Bouygues, NextLink, RoamVault, IIJ and MeiSIM itself. Cheapest Japan plan 5.99, cheapest France 4.70, cheapest US 5.99.

A country with no plans answers an empty list and no error: a request for AQ (Antarctica) returned zero plans, which is a coverage statement rather than a failure. That is the right shape, and it is worth noting because a silent empty is often indistinguishable from a broken call.

What it refuses, and how the refusal reads

Six account tools answered without a key, all at HTTP 200 with isError: true and the same readable body:

{"error": "unauthorized",
 "message": "This tool needs a MeiSIM API key. Remote: send it as an Authorization: Bearer header on the connection. Local: set MEISIMUSA_API_KEY. No key yet? Sign up at ...",
 "status": 401}

Three properties are worth copying here. The message names both routes in, remote header and local environment variable. It states that a person must open the account. And it masks the credential in its own example rather than printing a token shape an agent might echo.

topup_lookup is the exception: it answered without a key, returned an empty lines array with a hint naming the three fields that would find an order. That inconsistency is a small design choice, harmless in practice, and welcome because it lets a flow check a number before asking for a credential.

Three defects, recorded rather than smoothed

1. A required field the description says is unused. purchase declares payment_method as a required object with a type enum (stored_card, stripe_payment_method) and a token, while its own description reads Not used: payment happens on the hosted Stripe page via payment_url. Reserved for a future direct-charge path.

A caller reading the description will omit it and be refused. A caller reading the schema will send a fake token to satisfy a field nobody reads. Measured: purchase with payment_method as the string "card" and a non-GUID idempotency_key returned

MCP error -32602: Input validation error: Invalid arguments for tool purchase:
Invalid input: expected object, received string at payment_method Invalid GUID at idempotency_key

Two failures concatenated in one sentence, the second with no at separator. Both were real, both were ours, and one message per failure would have made the retry a single edit instead of two.

2. A search that requires a filter it presents as optional. search_plans lists data_gb with the note Omit to return all sizes, and duration_days as a plain integer. Calling it with a country code alone returns

MCP error -32602: Input validation error: Invalid arguments for tool search_plans:
Invalid input: expected number, received undefined at duration_days

So the one filter a caller is invited to omit is not data_gb, it is the term. The guide records it because the first call of a session will hit it.

3. A zero ceiling is a validation error, not a query. max_price_usd must be greater than zero, so show me the cheapest has to be asked with a number the caller guesses. The refusal names the bound, which is the difference between a one minute fix and a support thread.

What this replaces in an operator's stack

The class exists without MCP: Twilio provisions numbers, a carrier API sends texts, and a spreadsheet tracks renewals. What changes with this surface is that an agent can hold the whole loop in one place, because the three operations that normally live in a dashboard are tool calls: read inbound, wait for the next one, send a reply.

Two limits are stated by the server and should be quoted to anyone planning on it. A number is billed in a two month minimum term paid up front, and there are no refunds once a number is delivered. An agent with purchase authority over this tool can spend real money on a line it will keep for sixty days.

Why it earns a page

This catalogue carries several ways to send a message and no way to own the number the message arrives on. Mobile Text Alerts covers broadcast texting to a subscriber list; WhatsApp Business covers the Meta messaging platform on a business number that Meta verifies. Neither gives an agent a US mobile line with inbound texts and a SIP voice channel on a per-line price, and neither sells the number at all.

The find is not that the telecom API is now an MCP server. It is that the vendor wrote its own gate into the refusal text: the agent buys, provisions, reads and replies, and a human opens the account with an identity document. That is a workable shape for giving an autonomous process a real phone line, and it is the first entry here that draws that line explicitly.

Integration with CorpusIQ

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

AI answers on your data. See where every number comes from.

Connect your tools and ask in ChatGPT, Claude, or Perplexity. Enterprise-grade answers in minutes. From $29.95/month.

Start free, no credit card
Contact Terms Privacy Pricing