An agent that gets invited to meetings should respond like a real attendee. Not with a polite "I'm an AI assistant, please contact my human" auto-reply. Not by silently dropping the invite on the floor. It should look at the meeting, look at its own calendar, and click Yes, No, or Maybe — the same three buttons every other person on the invite gets.
Most "AI calendar" demos run the other direction. They point a model at a human's Google Calendar and let it organize: propose times, send invites, chase RSVPs. That's the organizer flow, and it's useful. But the moment your agent has its own email address and people start inviting it to things, you need the mirror image. The agent is now an invitee, and invitees don't organize — they respond.
This post is about that response. An invite lands in the agent's mailbox, Nylas turns it into an event on the agent's calendar, your code decides yes/no/maybe based on whether the agent is actually free, and you fire a single send-rsvp
call that updates the organizer's calendar the way any human's RSVP would. I work on the Nylas CLI, so the terminal commands below are the exact ones I reach for, and I'll pair every one with the raw HTTP call so you can wire it into a backend with no SDK.
Before the calendar mechanics, the one abstraction worth internalizing: an Agent Account is just a grant. It has a grant_id
, and that grant_id
addresses the same /v3/grants/{grant_id}/*
endpoints every other Nylas grant does — Messages, Threads, Events, Calendars, free/busy, all of it. There's no special "agent API." If you've ever listed events or sent an RSVP for a connected Google or Microsoft account, you already know the surface area here. Nothing new to learn on the data plane.
What makes an Agent Account different is the identity: it owns a real mailbox and a real calendar on Nylas-hosted infrastructure, so it can be addressed, invited, and RSVP'd-against like a person. The endpoints are boring on purpose. The interesting part is that a piece of software is now a first-class participant on the invite.
If you're new to Agent Accounts, the Agent Accounts overview and the calendar mechanics page are the two pages to read alongside this one.
Here's the lifecycle, because it's the part people get wrong. The agent doesn't parse ICS attachments by hand, and it doesn't reply to the invite email to accept. Nylas does the ICS plumbing for you:
assistant@yourcompany.com
) as a participant.event.created
webhook fires.participants[]
with status: "noreply"
. The organizer is whoever sent the invite — That third point is the conceptual pivot. The agent is a participant, not the organizer. It can't update or cancel the meeting; it can only respond. And because the event object already carries the organizer, the participants, the time window, and the description, you can drive your whole decision off event.created
without ever opening the invite email.
The agent also gets a message.created
webhook for the invitation email itself, because an Agent Account is always a real mailbox too. You get to pick which signal drives your logic. My advice: drive the RSVP off event.created
(it's already structured data) and treat the message.created
as optional context you fetch only if you want the human-readable note the organizer wrote.
One honest caveat on the webhook body, because the Nylas docs hedge on it: don't rely on the webhook payload to hand you the full message body. Fetch the full message with GET /v3/grants/{grant_id}/messages/{message_id}
when you need it, and branch on the message.created.truncated
trigger name (it appears when a body exceeds ~1 MB and the body is omitted). For RSVP decisions you usually don't need the email at all — the event has everything.
You can react to event.created
from a webhook, or you can poll. Webhooks are application-scoped — you subscribe once at the app level with POST /v3/webhooks
, every grant's events land at your endpoint, and each payload carries a grant_id
you filter on. That's the right call for near-real-time RSVPs. But for a tour, polling the events list is the clearest way to see the invite that just arrived.
List the agent's upcoming events with curl:
curl --request GET \
--url "https://api.us.nylas.com/v3/grants/<GRANT_ID>/events?calendar_id=primary" \
--header "Authorization: Bearer <NYLAS_API_KEY>" \
--header "Accept: application/json"
And the CLI equivalent — same grant, same primary calendar:
nylas calendar events list --calendar primary
The invite shows up as an event where the agent's own address sits in participants[]
with status: "noreply"
and the organizer
is someone else. That's your signal: this is an invitation the agent hasn't answered yet. Grab the id
off that event — you'll need it to RSVP. Pull a single one for detail:
nylas calendar events show <event-id>
If you want the organizer's actual words — the "hey, can you join our planning sync?" note — fetch the invite email directly. The event.created
and message.created
webhooks both hand you a message id, so read that one message by id instead of listing the inbox:
curl --request GET \
--url "https://api.us.nylas.com/v3/grants/<GRANT_ID>/messages/<MESSAGE_ID>" \
--header "Authorization: Bearer <NYLAS_API_KEY>" \
--header "Accept: application/json"
nylas email read <message-id>
If the invite came in on an existing thread, email threads show <thread-id>
gives you the whole conversation so the agent has context before it commits to a time. None of this is required to RSVP — it's the layer you add when "should I accept?" depends on what the organizer wrote, not just on whether the slot is free.
This is the part that makes the RSVP mean something. Before the agent says yes, ask its calendar whether it's busy during the meeting window. Nylas exposes free/busy at the grant level, and the agent can query its own address.
With curl, you pass the meeting's start and end as epoch seconds and the agent's own email:
curl --request POST \
--url "https://api.us.nylas.com/v3/grants/<GRANT_ID>/calendars/free-busy" \
--header "Authorization: Bearer <NYLAS_API_KEY>" \
--header "Content-Type: application/json" \
--data '{
"start_time": 1744387200,
"end_time": 1744390800,
"emails": ["assistant@yourcompany.com"]
}'
The response comes back with the busy blocks in that window. If the meeting's window overlaps a busy block, the agent has a conflict. If it's clear, the agent is free.
The CLI wraps the same endpoint with human-friendly time parsing, which is genuinely nice when you're testing by hand:
nylas calendar availability check \
--emails assistant@yourcompany.com \
--start "tomorrow 2pm" \
--end "tomorrow 3pm"
You can also pass --duration 1h
instead of an explicit --end
. Either way you get the agent's busy slots for the window, and that's the raw signal your decision logic consumes.
Here's the boundary worth being explicit about, because it's where people expect Nylas to do more than it does. Nylas tells you when the agent is busy. It does not decide whether to accept. That judgment — accept if free, decline if double-booked, maybe if it's a low-priority meeting during focus time — is your application logic. A reasonable first cut:
yes
no
maybe
Two constraints to design around. First, store the decision and its reasoning in your own database. Agent Accounts don't support custom metadata
on events yet, so you can't stamp "declined because double-booked at 14:00" onto the Nylas event and read it back later. Keep that state — the event id, the decision, the rule that produced it — in your own store. Second, this is genuinely your logic to own. The model (or your heuristics) decides; Nylas just executes the RSVP. If you want the agent's "reasoning" to be auditable, that audit trail lives in your DB, not in Nylas.
If your real need is round-trip negotiation — the agent proposing alternative slots, collecting picks, booking the winner — note that Scheduler isn't available for Agent Accounts today. Counter-proposing a time isn't a first-class endpoint either. The honest pattern for "that time doesn't work" is to RSVP no
or maybe
and reply to the invite email with a suggested alternative, then let the organizer create a new event.
Once your code has picked a status, one call does it. Use the send-rsvp endpoint — POST /v3/grants/{grant_id}/events/{event_id}/send-rsvp
— with status
set to yes
, no
, or maybe
, and the calendar_id
the event lives on:
curl --request POST \
--url "https://api.us.nylas.com/v3/grants/<GRANT_ID>/events/<EVENT_ID>/send-rsvp?calendar_id=primary" \
--header "Authorization: Bearer <NYLAS_API_KEY>" \
--header "Content-Type: application/json" \
--data '{ "status": "yes" }'
The CLI command is nylas calendar events rsvp
, which takes the event id and the status as positional arguments:
nylas calendar events rsvp <event-id> yes
nylas calendar events rsvp <event-id> no --comment "I have a conflict at that time"
nylas calendar events rsvp <event-id> maybe
--comment
attaches a short note to the reply, and --calendar
(defaults to primary
) targets a non-primary calendar if the invite landed on one.
Behind the scenes, send-rsvp
sends an ICS REPLY
to every participant. From the organizer's side, the agent now shows as "accepted" (or "declined", or "tentative") next to every other attendee — exactly like a human clicking the button in their mail client. Other attendees see it too, because their calendars get the update automatically. This is why you must use send-rsvp
and not a plain reply email: a reply email won't move the organizer's calendar status. The RSVP endpoint is the only thing that does.
After the call, an event.updated
webhook fires on the agent's own calendar, so the agent observes its own state change land. If you're keeping a mirror of RSVP state in your DB, that's a clean confirmation signal to reconcile against.
A few things I'd bake in before pointing this at production traffic:
id
, which stays constant across retries of the same event. You can additionally guard on the inner event id so the agent never RSVPs twice to one invite.PUT
on the event.noreply
is your "needs action" filter.status
is still noreply
. Once you've RSVP'd, it flips, which keeps you from reprocessing.The mechanics are small once the model clicks: an invite becomes an event, free/busy becomes a decision, send-rsvp
becomes a reply on the organizer's calendar. Everything rides the same grant you'd use for a connected account, so there's no second integration to build.
From here:
nylas calendar
and nylas email
command used above.If you've already built the organizer side — the agent that sends invites and chases RSVPs — this is the missing half. Wire both and your agent isn't watching a calendar from the outside. It's on the invite.