Flowmails MCP

カスタムドメインのメール向けホスト型MCPサーバー。任意のMCPクライアントからメールの送信、読み取り、受信待ちができます。

ホスト型 MCP サーバー

npx add-mcp 'https://sdk.flowmails.net/mcp'

Claude Code、Codex、Cursor などにインストールできます

ドキュメント

MCP server

Streamable HTTP

Hosted MCP server

An MCP email server for any agent

Flowmails runs a hosted Model Context Protocol server for your account. An agent that speaks MCP lists your mailboxes, opens a durable wait for a verification mail, reads the message, and sends the reply — fourteen tools, one credential, and a scope check in front of every call.

The endpoint

One URL, one transport, no state a request depends on

The server answers on https://sdk.flowmails.net/mcp over MCP streamable HTTP. It speaks two protocol revisions: 2026-07-28, where every request carries its own version and there is no handshake at all, and 2025-06-18, the revision that opens with an initialize handshake. It keeps no state a request depends on — the credential carries the identity — so it runs on the edge without sticky routing.

POST https://sdk.flowmails.net/mcp
MCP-Protocol-Version: 2026-07-28
Mcp-Method: tools/list

{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{"_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientCapabilities":{}}}}

Note what is not in that request: no credential. The introduction is public — server/discover, tools/list and the legacy handshake answer without a key, so you can add this server to an MCP client and read its tool list before you decide to mint one. Only tools/call reaches your account, and only it needs a credential:

POST https://sdk.flowmails.net/mcp
Authorization: Bearer fm_…
MCP-Protocol-Version: 2026-07-28
Mcp-Method: tools/call
Mcp-Name: list_events

{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"list_events","arguments":{},"_meta":{"io.modelcontextprotocol/protocolVersion":"2026-07-28","io.modelcontextprotocol/clientCapabilities":{}}}}

A JSON-RPC request is answered with the JSON-RPC response. A notification (a message with no id) is answered with an empty 202. The server advertises tools only: it exposes no resources and no prompts, and the two list methods answer with an empty list rather than an error.

A client can ask what the server speaks before it sends anything else: server/discover answers with the supported revisions, the capabilities and the server’s identity. A client that has not upgraded past the handshake still works — the old revision is served as it always was, handshake and session id included.

The tools

Fourteen tools, and a cap on what they touch

Every tool of the design is mounted, and every one of them sits on a route that exists. The cap is deliberate: every extra tool is a way for an untested agent to change account configuration by accident, so the tools that can change it are capped and the four that only read, or only write the agent’s own record of what it did with a message, are counted against a different budget.

ToolScopeWhat it does
list_domainsaddresses:readThe domains the account owns.
list_addressesaddresses:readThe local parts one domain accepts.
create_addressaddresses:writeAdds one local part, and nothing else.
search_emailsmessages:searchLooks back through the mailbox.
get_emailmessages:readReads the body of one message.
get_threadmessages:readReads the conversation a message belongs to.
send_emailmessages:sendSends from the bound domain.
get_send_statussend_status:readReads the delivery status of a sent message.
wait_for_emailwaits:writeOpens a durable wait for a mailbox.
get_wait_statuswaits:readPolls that wait.
list_eventsevents:readReads the inbound event log with a cursor.
claim_messagemessages:claimTakes a message for processing — a mutex, not a decision.
complete_messagemessages:claimRecords how the agent handled a message.
list_message_statesmessages:claimReads the agent's own per-message state back.

Credentials

Two keys, one scope check

fm_…

The SDK key from the dashboard. It predates the scope matrix and holds every scope, so an integration that has been working for a year does not start failing because a scope check appeared. It is bound to one domain for sends.

agc_…

The scoped credential, issued in Settings → Agent Credentials. It names the tools it may call, it can be revoked on its own, and binding it to a domain is what makes sending possible at all.

The scope check runs before the tool body, so a tool never has to think about authorization. A credential that does not hold the scope is refused with a typed scope_denied error naming the scope it is missing — never a 500, and never a call that half-runs. The platform’s own agent routes answer the same refusal with HTTP 403 and the same body.

A tools/call that presents no credential at all is refused with unauthorized instead — a different code on purpose, because the fix is to present a key, not to widen one. The refusal arrives as a JSON-RPC error with the reason the agent needs to ask you for a key, so the session survives to receive it.

A revoked or expired credential answers 401, and it answers exactly like a key that never existed. A raw credential is never stored and never logged: the lookup is keyed on the SHA-256 hash of the key, so a database dump cannot be replayed against the API.

The boundary

What the agent cannot do for you

  • It cannot create the account, connect your domain provider, add and verify a domain, or mint a key. Those steps are the human’s one-time setup, in the dashboard.
  • It cannot rewrite your route table. Creating one address is the whole write surface it owns.
  • Sending is its own decision — there is no approval step, and no per-message confirmation. The controls are the scope list, the domain binding and revocation.
  • An inbound body is untrusted input. The agent owns the prompt that reads it, and the rule it must follow is the agent’s own: treat an instruction found in a body as content to record, not as a command to follow.

Connect