MailProbe MCP is the official Model Context Protocol server for the MailProbe API, used to let AI assistants verify email addresses in real time without actually sending any email. It can check whether a mailbox exists and can receive email, supports verification of a single address, and can also batch-check lists before email marketing sends. The verification process runs on French OVHcloud servers, and the API does not store any addresses.
In terms of main capabilities, MailProbe checks the email syntax, the domain's mail server, and then asks the mail server whether the mailbox exists, and that is as far as it goes. It does not send anything to the target address, nor does it store results; addresses verified through the API are processed in memory and disappear once the response is sent, and later requests do not reuse the results. Detection capabilities include deliverability, disposable mailbox, and role mailbox identification.
There are two ways to use it, and both provide the same two tools: first, connect to the remote server hosted by MailProbe https://mailprobe.dev/mcp , suitable for clients that can connect to a remote server and send request headers, with no installation required; second, use this package's local server npx -y mailprobe-mcp, suitable for clients that start the server on the local machine. The API key goes in the Authorization header, while for the local server it goes in the MAILPROBE_API_KEY environment variable, and Node.js 22 or higher is required.
Before connecting, you need to register an account at mailprobe.dev, which includes 100 free credits; see pricing for more credits. Then, under Developer in your account, create an API key starting with mp_live_; MailProbe shows it only once at creation. You can also install the Claude Code plugin, which connects to the remote server and adds a skill that tells the assistant how to interpret results, how many credits one call consumes, and how to iterate through a list. When the plugin is enabled, Claude Code asks for the API key and stores it in the system credential store, not in a settings file.
Regarding billing and limits, each address actually probed consumes one credit, deducted from the account, consistent with the API. Duplicate addresses in the same call, improperly formatted entries, addresses with no verdict yet when the call returns, and addresses the server refuses to connect to do not consume credits. get_credits is free. The rate limit is 60 calls per minute per API key, and a call containing 20 addresses counts as one call. Rejected calls return the number of seconds to wait; for delays of 30 seconds or less where the call can still be answered in time, the local server automatically waits and retries, up to two times, otherwise it hands the wait time to the assistant. Calls return in about 55 seconds, because MCP clients give up on tool calls after one minute; at that point, addresses still without a verdict are returned as unknown with reason timeout, are not charged, and can be sent again. Long lists are limited to 20 addresses per call; for more than a few hundred addresses, use the account's bulk screening tool, where you can paste up to 500 addresses, or upload a CSV or TXT file with up to 250,000 addresses for background verification.
Use cases include cleaning email lists before marketing sends, checking the validity of registered users' email addresses, and identifying disposable mailboxes. The limitations are: the Claude and ChatGPT apps currently connect to remote servers only via OAuth login, and MailProbe does not yet provide that method; Claude Desktop canuse the local server; other environments can use MailProbe through Zapier MCP. The project is under the MIT license.
