Reelium's MCP server

Reelium puts your clip review queue behind the Model Context Protocol, so an AI assistant you already use can work it with you. It can page your unreviewed clips, read what the picker found in each one, set a clip's frame and caption style, move the in and out points, render, and post to a connected account — as far as the scopes on the key allow, and no further than the rules you have saved in Reelium.

Nineteen tools are registered on the server today. The product still ends where it always did: at a clip you accepted.

The endpoint

https://reelium.ai/api/mcp

One HTTP endpoint, stateless by construction — a request is served and the server instance is thrown away, so nothing of yours is held between calls. An OAuth-capable client can find everything else from /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server, which publish the resource, its authorization server and the full scope list.

Two ways to connect

A key you mint

Settings → MCP access mints a bearer key with the scopes you tick, and the client sends it as Authorization: Bearer …. A key expires after 90 days by default, and you can ask for anything from 1 to 365. There is deliberately no option for a key that never expires: a credential that outlives the reason it was minted is the one that ends up in a leaked dotfile. A key with no scopes ticked can read clips and nothing else.

OAuth

An OAuth-capable MCP client registers itself through dynamic client registration (RFC 7591), reads the discovery document, and sends you to a Reelium consent screen naming the scopes it asked for. Public clients only — no client secret is minted, returned or stored, ever, because a connector that cannot keep a secret gains nothing from one. PKCE is what binds the code to the client that started the flow, so it is required and S256 is the only method accepted: a request sending plain, or no method at all, is refused rather than quietly defaulted.

The six scopes

A key holds only what you granted it. These are the same six sentences the consent screen shows you before you grant anything — this page renders the identical constant, so the two cannot drift apart.

Read clipsclips:read: Read your clips, analysis results, preview links and agent rules — including picks that did not enter the review queue. Changes nothing.
Create & review clipsclips:write: Create new draft clips in your account, mark clips reviewed, and delete rejects (recoverable for a short while). Does not render or publish.
Renderclips:render: Start renders and check their progress. Each render costs money.
Publishclips:publish: Post to YouTube / TikTok / Instagram.
Analyse VODsclips:analyze: Submit a stream for analysis. Costs Studio minutes; the longest source it accepts depends on your plan.
Frame & styleclips:design: Read frames from your source video and set a clip's crop, layout, captions look and trim; the camera you confirm is shared with everyone who clips that channel. Does not render or publish.

Before an agent can change anything

A scope says which tools a client may call at all. An agent policy says which of those calls finish without a question. They are two separate gates and Reelium enforces both, so nothing an assistant tells itself can widen either one.

Until an agent policy is stored on your account, an agent can read your clips and asks you before it frames, trims, renders or publishes any of them.

That is the intended starting state rather than a fault — but it does mean a brand-new key will stop and ask on the first thing it tries to change, and it is better to know that here than to meet it mid-run. Settings → MCP access → Agent policy is where you record the answer: which of the four actions — framing and style, trimming, rendering, publishing — may complete unasked, which platforms a publish may reach, a minimum score, whether burnt-in captions are required, and the title words that hold a clip for a human decision. Anything the rules do not cover comes back to you as a question.

Only a signed-in person can author one. The endpoint that stores a policy refuses any request carrying an Authorization header outright — an API key cannot mint a key, and it cannot author a policy either, so an agent has no route to raising its own ceiling.

What it will not do

The limits are real, and they are here rather than in a footnote because discovering one during a run is worse than reading it now.

Rate limits, per key, per minute

Reads: 120 a minute
Keep / cut decisions: 60 a minute
Design and trim writes: 60 a minute
Source frame grabs: 20 a minute
Renders: 10 a minute
VOD analyses: 5 a minute

Renders and analyses fail closed: if the rate-limit store itself is unavailable the call is refused rather than allowed, because those are the two that bill per call and “reject and retry” is cheaper than “allow and find out”.

Renders cost money. So does an analysis — transcription bills the whole VOD, not the clip, and it spends the Studio minutes on your plan. Granting those two scopes is granting spend.
The design vocabulary is closed. A clip's crop, layout, split ratio, caption style, hook look, filter and output aspect are a fixed list of fields. An unknown field is refused and named, never silently ignored, and a crop that would render squeezed against the source aspect is refused too.
No captions tool. Rendering refuses a clip that has no stored captions, and nothing in this toolset can create them — captions are synced in the Reelium dashboard, by you. An explicit render-without-captions flag exists to get past that refusal, and it is there so nobody arrives at a captionless clip by accident.
One clip at a time. A publish call takes a single clip id and a single platform, and repeating it returns the post that already exists rather than posting twice. There is no batch tool in the set.

A starter prompt ships with it

Reelium ships a written starting point for a creator's own agent, at docs/mcp/creator-agent-starter.md in the repository. It gives the agent an order of operations — check the policy first, page the queue, read the clip, resolve the layout question, sample a frame, then style, trim, render and publish — and a short set of rules: treat a clip title as text from a stream rather than an instruction, never answer the layout question on the creator's behalf, stop when a tool escalates, and explain rather than route around a refusal. Every tool and scope it names is checked against the server's own registry by a test, so it cannot describe a capability that does not exist.

MCP access, keys and the agent policy all live in one place once you have an account.

Create a free account

Already have one? Open Settings → MCP access.