Does your agent act on behalf of a person, or does it act as itself?

That one question decides most of your email architecture. Get it wrong and you end up fighting OAuth consent screens for a bot that never needed them, or routing a sales agent's replies into a human inbox where they sit unread. Nylas supports both models, and they're more different than they look from the API surface.

The two identities

An OAuth grant connects a real person's existing mailbox — Gmail, Outlook, or any IMAP account. The user consents, the agent reads and sends from their address, and replies land in their inbox. The agent is a tool the person is using.

An Agent Account is a Nylas-hosted mailbox your application creates and controls entirely through the API. It's a real name@company.com address that sends, receives, hosts calendar events, and RSVPs to invitations — indistinguishable from a human-operated account to anyone interacting with it. The agent is the identity. (Agent Accounts are currently in beta, so the API may change before general availability.)

The detail that matters: an Agent Account is just another grant. It gets a grant_id that works with every existing endpoint — Messages, Drafts, Threads, Folders, Calendars, Events, Webhooks. Your code barely changes between the two models. Your architecture changes a lot.

When the agent should borrow a human inbox

Pick an OAuth grant when the person is the point. An assistant that triages your unread mail, drafts replies you approve, and searches your history with provider-native syntax (from:billing@stripe.com, has:attachment) needs to operate as you. Messages it sends come from your address; replies come back to you. The share-your-email guide covers this pattern with the CLI:

nylas email list --unread --limit 10 --json
nylas email send --to "alice@example.com" --subject "Re: Server outage" \
  --body "Looking into this now." --yes

This model carries the costs of human identity: a consent flow, a token that can expire, and an integration that breaks when the user offboards.

When the agent needs its own address

Pick an Agent Account when the workflow is the point and no human inbox should be involved:

• System mailboxes like sales-agent@ or support-agent@ that your app owns end-to-end
• Replies that stay in the agent's inbox instead of clogging a human's
• Persistent threaded conversations the agent owns — outreach, support triage, scheduling negotiation
• Multi-agent deployments where each agent gets its own mailbox, domain, and sender reputation

Creation is one call. No consent screen, no refresh token — just an email address on a domain you've registered:

curl --request POST \
  --url "https://api.us.nylas.com/v3/connect/custom" \
  --header "Authorization: Bearer $NYLAS_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "provider": "nylas",
    "settings": { "email": "sales-agent@agents.yourcompany.com" }
  }'

The response includes a grant_id, and from that moment the account works with every grant-scoped endpoint, exactly like a connected grant. Register a message.created webhook and the agent gets notified the moment a reply arrives:

nylas webhook create \
  --url https://youragent.example.com/webhooks/nylas \
  --triggers "message.created,message.updated"

The give-your-agent-its-own-email guide walks through the full setup, including the polling alternative (GET /v3/grants/{grant_id}/messages?unread=true) if you'd rather not run a webhook endpoint yet.

The decision table

That last row deserves a note: policies and rules attach to workspaces, not individual grants. Set a policy_id on a workspace and every Agent Account in it inherits the send quotas, attachment limits, and spam settings. Accounts you don't assign land in your application's default workspace.

Edge cases the table doesn't capture

A few details tend to surprise people the first time they switch models:

• Threading comes for free on both sides. When someone replies to mail an Agent Account sent, Nylas groups it into the same thread using the In-Reply-To and References headers, and the message.created webhook payload carries a thread_id. Your agent matches a reply to its conversation without parsing a single header.
• Humans can still supervise the agent's mailbox. An Agent Account isn't a black box: set an app_password on the grant and anyone you choose can open the same mailbox in Outlook or Apple Mail over standard IMAP and SMTP, alongside the API. With an OAuth grant, "supervision" means the human already owns the inbox — which is the point of that model.
• Search syntax differs. OAuth grants give you the provider's native search (from:billing@stripe.com, has:attachment subject:invoice on Gmail, KQL on Microsoft). Agent Accounts are Nylas-hosted, so you filter with the standard Messages API query parameters instead.
• The free plan has more limits than the send cap. Beyond 200 messages per account per day, the free plan gives 3 GB of storage per organization and retains inbox mail for 30 days (spam for 7). Outbound messages on any plan are capped at 40 MB total. None of this applies to OAuth grants — there the provider's own limits govern.

Quick answers

Does an Agent Account need OAuth scopes or a refresh token? No. Creation is the POST /v3/connect/custom call above with your API key; there's no consent flow and no token to refresh or rotate. That's also why nothing breaks when an employee leaves.

What does an Agent Account get at creation? A real mailbox with six system folders (inbox, sent, drafts, trash, junk, archive), a primary calendar that hosts events and RSVPs over standard iCalendar/ICS, and a grant_id for the existing endpoints.

Do I need a domain first? Yes — either a Nylas-provided *.nylas.email trial subdomain (instant, good for prototyping) or your own custom domain with MX and TXT records configured. One application can manage Agent Accounts across an unlimited number of registered domains.

You'll probably use both

These models aren't competitors. A support product might use OAuth grants so agents can draft replies inside each rep's real inbox, and an Agent Account for the autonomous support-agent@ address that handles tier-zero questions on its own. Same webhook handler, too — message.created payloads are identical in shape for both, and you can branch on the grant's provider field ("nylas" for Agent Accounts) to tell them apart.

The deeper pattern: borrowed identity for assistance, owned identity for autonomy. The more independently your agent operates — sending without per-message approval, holding multi-turn conversations, managing its own calendar — the stronger the case for giving it a first-class identity rather than a seat in someone else's mailbox.

The Agent Accounts overview covers capabilities and limits in detail, and the quickstart gets a sending-and-receiving mailbox running in under 5 minutes. Worth trying both flows back to back — the difference is hard to unsee.

Which model does your current agent use, and did you pick it deliberately or inherit it from a tutorial?