Reviewing your clip queue from a chat window

Reelium now speaks the Model Context Protocol, which is a plain way of saying that Claude can talk to your Reelium account directly. You add Reelium once, sign in through your browser, and then the review queue you would normally open in the dashboard is something you can work through in a conversation instead.

I built the server months ago and sat on it. This post is the honest version of what it does today — including the three places it will stop and refuse, which are the parts I would want explained before I trusted something with my own channel.

What MCP is, in one paragraph

The Model Context Protocol is a standard way for an assistant like Claude to call a real service rather than guess at it. Reelium exposes a small set of named actions — list the unreviewed clips, read one, restyle it, trim it, render it, post it — and Claude calls those actions. There is no scraping and no simulated browser. If Claude tells you a clip rendered, it is because the render tool returned, not because it wrote a plausible sentence.

Two ways in, and neither needs a key

In Claude or claude.ai you add Reelium as a custom connector. In Claude Code it is one command. Either way the next thing that happens is a browser window with a Reelium consent screen in it, and you sign in the way you always do. There is no API key to generate, copy, and leave sitting in a config file in plain text.

The consent screen lists six permissions — reading your clips, reading what the picker found, changing framing and design, trimming, rendering, publishing — and you approve them one at a time. Declining one is a supported answer, not a failure. The step-by-step for both routes has the exact sequence, and the server reference has the endpoint, the rate limits and the full tool list.

What it is actually good at

Nineteen tools are registered today, and the useful shape of them is a review pass. You ask what is unreviewed. You get the queue back with each clip's score and the evidence behind it — the moments a viewer already clipped, the bursts in chat — and you keep or cut them by talking through the list rather than clicking down it. For the ones you keep, you can set a caption style, move the in and out points a few seconds either way, render, and post to an account you have already connected.

The part I did not expect to like: asking why. Every pick carries its measured signals, and the tools hand those over, so "why did this one score higher than that one" is a question with a real answer rather than a guess.

Three things it will refuse, and why

• It will not render a clip that has no captions yet. Captions are made in Reelium — syncing subtitles is a button on the edit step — and nothing in the tool set can make them from a conversation. So the first render attempt on a fresh clip comes back as a refusal naming the missing piece. Annoying the first time, correct every time: a render is the expensive step, and a caption-less render is a wasted one.
• It cannot change, render or publish anything until you have written an agent policy. A new connection can read. Restyling, trimming, rendering and publishing are each gated, and by default every one of them comes back to you as a question instead of happening. You lift that by writing the rules in Settings, in your own words, naming what is allowed. Saying "yes, go ahead" in the chat does not lift it — the policy is re-read from your account on every attempt, so the only thing that counts is what you wrote down.
• It publishes only to accounts you have already connected. There is no path from a conversation to a new social account. If TikTok is not linked in the dashboard, posting to TikTok is not an option that exists.

All three are the same idea from different angles. The interesting failure mode with an assistant is not that it does nothing — it is that it does something confidently on a channel that has your name on it. The refusals are where that gets stopped.

One upgrade note

If you connected before 8 September, reconnect once. The framing and design permissions were added after older grants were issued, and a grant does not gain permissions after the fact — so an older connection can read and render but will refuse to restyle. Disconnecting and reconnecting issues a fresh grant with all six.

What this does not replace

The dashboard is still where the visual work happens. Choosing a cover frame, dragging a crop, judging whether a caption sits over someone's face — those are things you look at, not things you describe. The queue is what moves well into a conversation: the deciding, the asking why, the batch of small yes-and-no calls that is most of the work and none of the fun.

Nothing here posts anything you did not agree to. It is the same product with a second door on it.

Set it up, or start with a VOD if you have not tried Reelium yet.