Say you run a small team, and half of your inbox lives somewhere Claude can't reach. Claude has built-in connectors for Gmail and Microsoft 365. Your own mail is on Yahoo, iCloud, an Exchange server and a plain IMAP box. You want an assistant that can search, read and send, and you want it today.
Nylas published a guide to exactly this gap. It starts by naming it: Claude has connectors for Gmail and Microsoft 365, "but nothing for Yahoo, iCloud, Exchange, or other IMAP mailboxes." The Nylas MCP server covers six provider types through one setup: Gmail, Outlook and Microsoft 365, Yahoo Mail, iCloud Mail, Exchange (EWS) and any IMAP server.
Setup is four steps through the Nylas CLI. You run nylas auth login once per provider, then nylas mcp install --assistant claude-desktop. The same command also takes claude-code, Cursor, Windsurf and VS Code. Once it's connected, Claude can search and read your mail, send it, and read your calendar.
Here is the part you'd want to read twice. The server requires a confirmation call before any send. You might assume that means a person clicks yes. Nylas says otherwise, in a sentence most vendors would leave out: "That confirmation is a call the agent makes, so on its own it is not a human approval, and whether it is shown to you depends on your client."
So the guardrail is weaker than it sounds, and Nylas tells you so. The advice that follows would hold with any vendor. Draft into a review queue instead of sending. Treat every email body as untrusted input, because it can carry a prompt injection. And start read-only on one mailbox before you widen anything.
The post also draws a line you should note if your plans grow. The CLI route is one person and one mailbox. If you're building a product where many users connect their own inboxes, that belongs on the API, with per-user authorization and webhooks.
Now the turn. The APIs.io catalog has its own record of Nylas, and it lines up with that sentence. Nylas is one of the largest records in the catalog, at 43 API pages. The story touches the core of it. Reading and searching mail is the Messages API and the Threads API. The review queue you'd draft into is the Drafts API. Calendar reads are the Events API. Per-user connections are the Manage Grants API, and new-mail delivery is the Webhook Notifications API.
Then the catalog's agentic access profile for Nylas counts 210 operations. Of those, 127 are acting operations, meaning they change something rather than just read. Exactly one is human-in-the-loop. One human-in-the-loop operation across 127 acting ones is the catalog's version of the sentence Nylas wrote itself.
The scores are good, and the post doesn't hide the gaps. Nylas scores 92.6 on the Kin Score, exemplar, with access clarity at 96.8 and developer ergonomics at 94.6. Contract governance at 45.5 is the only facet under 80. Agent Readiness is 59.6, agent-native, with 16 dimensions lit, including consent identity, idempotency, reversibility and an agent card. Dry-run mode is unlit. The post explains why that matters less than it looks: the draft is the dry run.
What you take away is practical. Don't read "requires confirmation" as "asks a human." Check whether your client actually shows you the prompt, because Nylas says that depends on the client. Point Claude at one mailbox, read-only, and let it write drafts that you review before anything goes out. Treat incoming mail as hostile text. And if you ever move from your own inbox to many users' inboxes, plan on the API rather than the CLI.
There's a second thing to take away. A vendor that tells you where its own safeguard stops, and a catalog record that shows the same thin line of human review, are two sources saying one thing. You can plan around that.
Read the original: https://apis.io/2026/11/16/nylas-puts-claude-on-six-mailbox-types-and-says-its-send-check-is-not-a-human/