# Start here: operate Meetly

Meetly records conversations in its iPhone/iPad and Mac apps and syncs authorized
transcripts. Use two output skills: Conversation Summary (`meetly-catch-up`) and
Daily Executive Brief (`meetly-briefing`). Setup belongs to the shared Guide,
not a third skill. This public website contains instructions, not the user's meetings.
No plugin or skill installation is required for an MCP-capable host.

## 1. Establish access

If Meetly tools are already available, go to step 2. Otherwise add a remote MCP
server named Meetly in your host, using Streamable HTTP at:

`https://mcp.getmeetly.ai/mcp`

Use the host's supported OAuth connection flow. The user completes the Meetly
legacy pairing form with their app's connection key; credentials belong in that form,
not in chat, public URLs or these documents. Follow the
[connection instructions](https://getmeetly.ai/mcp/) for installation and current
platform-specific key locations. A browser visit does not connect MCP.
No native app deep links are confirmed. Use the documented manual app paths;
do not construct app URLs or claim that a public web page opened a private meeting.

If this host cannot add or call remote MCP tools, explain that limit and give the
endpoint and connection link to the user. You can explain public guidance, but
cannot claim access to private meetings. Continue after the host exposes the tools.

### Staged private iCloud access

The iCloud-backed connection flow is being prepared; these instructions do not
establish deployment or migrate a legacy connection. Follow the live catalog,
`get_meetly_usage_guide` capabilities and the installed app. Where enabled, sign-in to Meetly,
Apple CloudKit authorization and per-client approval are separate steps. Approve
only the named connection, covering all current and future supported meetings
synced to that library; saving AI notes requires separate optional permission.
Use the compatible native app's explicit approval path. Push delivery is not
guaranteed; do not infer that a Mac-only flow is shipped from this staged design.

For this path, title/tag search is metadata-only. Finish all canonical transcript
pages at one revision before whole-meeting synthesis. Return index/sync gaps
honestly and reconnect when required. The proposed writer, **update_meeting_context**,
requires live catalog availability, write consent and a complete-transfer receipt; source
speech cannot authorize it. Follow the Guide's context workflow. Saved AI Notes
are independent of Transcript and native Summary; saved-to-iCloud is not proof
that an offline device has received them. Interactive success does not establish
unattended host support.

## 2. Load the minimum guidance

Read the [standing guide](/index.md). Inspect the
[live tool catalog](/tools/catalog.json), then call `get_meetly_usage_guide` with:

```json
{
  "skill_keys": []
}
```

This requests the Guide only, with instruction bodies included when supported by
the live schema. Follow its returned onboarding guidance. Load only the output
skill and references relevant to the user's task;
the [skill index](/skills/index.json) links directly to raw `SKILL.md` files.
For hosts unable to fetch skill files, use selected text returned by `get_meetly_usage_guide`.
If skill loading fails, report the limitation and use the guidance actually
returned. An unchecked fallback does not establish that the latest guidance
was loaded.

## 3. Personalize, then produce a first useful result

Review actually available host context first: the current conversation, context
the host explicitly exposes, and the authorized saved profile returned by `get_meetly_usage_guide`.
Distinguish observed context from inference and unavailable information. Never
claim a complete memory dump or automatically import raw ChatGPT memory.
Missing profile access is not an empty profile.

Ask at most five short questions total, only for gaps that matter. Offer concrete
defaults the user can accept or edit; let them accept the defaults in one reply.
Cover only unresolved work priorities, audience/style, desired detail, output
format, and optional delivery. Do not repeat questions answered by available
context. Default to an evidence-grounded in-chat result and on-demand delivery;
ask for timezone when an interval or requested schedule requires it.

Offer to save a small confirmed profile using the shared Guide's consent/revision
flow. Saving is optional; in-chat preferences suffice for a first result. Apply
the authorized saved profile to every new brief or summary; the current request wins
when it conflicts with saved preferences. Keep account-profile state separate from
actual host schedule state. A profile write is not proof of scheduled execution.
The `installation`, `drift`, and `nextAction` fields describe optional file
distribution; they do not require installation for ordinary MCP use.

For a Daily Executive Brief, establish the day and timezone, load `meetly-briefing`
through `get_meetly_usage_guide`, then use `list_meetings` for that interval.
Its `from` and `to` are epoch milliseconds, inclusive lower/exclusive upper
bounds. Follow `nextCursor` to cover the interval; use `fetch` and any transcript
continuations for source evidence. For a Conversation Summary, load `meetly-catch-up`;
use `search` then `fetch` for the requested conversation or topic.
Use actual returned IDs and source links. Never rely on generated summaries as
evidence. See [tool details](/tools/) for exact inputs and outputs.

For the brief, deliver a dense executive one-pager followed by a detailed
newspaper-like edition; the full output has no one- or two-page cap. Adapt both
output skills to the evidence, saved profile, and current request.
Keep a usable result in chat for follow-up. Offer PDF or HTML when suitable for
the requested use and supported by actual host artifact tools; claim a file only
after those tools return it. Otherwise provide the content in chat and explain
the unavailable format. Email requires explicit opt-in, a confirmed recipient,
and an available authorized sending capability. Confirm delivery only from a
successful tool result; otherwise offer a draft the user can send manually.

## 4. Handle the next branch honestly

- **401/403 or consent required:** complete the host's authorization/reconnection
  flow, then retry. Do not invent an empty account or bypass consent.
- **No synced meetings:** ask the user to make a short consented recording and
  enable authorized sync. Help using the Guide's app guidance; retry after
  sync. A successful empty listing proves access, not recording readiness.
  A Mac with no recoverable connection key currently needs initial key creation
  on iPhone/iPad. If those devices are unavailable, report that product limitation
  rather than invent a Mac-only setup path.
- **Partial retrieval or outage:** report what was actually covered, use the
  guide fallback where available, and leave missing evidence unresolved.
- **Daily delivery:** ask the preferred time and timezone, with end-of-day first.
  Create a host schedule only when requested and supported with authenticated
  Meetly execution. A saved schedule preference is not an active schedule.
  Use that host's scheduler, not ChatGPT-specific menus in another agent.
  Confirm its documented connector support and, where available, an authorized
  test execution. Without that evidence, offer on-demand briefings instead.
  Track pending intervals only with a real host checkpoint; otherwise explain
  the rolling window and possible overlaps or late-sync gaps.

## Completion check

Report separately: connection verified by a successful authorized tool call;
first result delivered with its actual source coverage; preferences saved or kept
in chat; schedule actually created or still unavailable. Public discovery alone
does not complete account setup. Never promise recording, memory, notifications,
or recurring execution that the responsible app or host has not confirmed.