Updated 10 November 2026 · Markdown

Assistants (MCP)

KitCut is an MCP server at https://kitcut.ai/mcp, so an AI assistant can make films from a chat. The person signs in to KitCut once. The assistant can then:

• start films, follow them, and make them again; line up several, which start one after another by themselves;
• pass pictures, text files and voice notes;
• bring back characters, and put a project's cast in an episode by @name;
• run projects: start one, write its brief, give it pictures, keep drafts and voice lines, and make its episodes;
• write a finished film's YouTube title, description and thumbnails, publish it to a connected channel, and follow it until it is live;
• work across every connected channel at once: publish or schedule a film on several, and read their calendars and numbers side by side;
• read YouTube's own numbers for a channel or a video, read one video in full, and change a published video's words, thumbnail or who can watch it;
• schedule films, even while they are being made, one at a time or every finished episode of a project by itself;
• say how the films are doing, and gather everything to plan the next videos from;
• report the credit balance.

Films made this way are the same films, with the same credits and limits, as on the website (Making a film).

An assistant does only what the person can do with their own account on the website. No tool reaches another account, and no tool can do something the site does not let the person do.

Connecting

Using it to run a faceless YouTube channel, with episodes researched, made and scheduled on a timer: A faceless YouTube channel, run from your AI assistant.

Sign-in and connection

• Transport: Streamable HTTP. Every request needs a token, including listing the tools.
• Sign-in: OAuth 2.1 with PKCE (S256), and one scope, films.
  • The person signs in on KitCut's own page (Google, or a 6-digit code by email) and presses Allow. The assistant never sees the sign-in.
  • Apps register with a Client ID Metadata Document or by Dynamic Client Registration.
• Tokens:
  • An access token lasts 1 hour.
  • A refresh token lasts 30 days from its last use and is replaced every time it is used.
  • Reusing an old refresh token disconnects the app.
• Disconnecting: on the account page under "Connected assistants". The app's tokens stop working at once.
• What KitCut receives: only what the assistant sends on the person's behalf: the idea, the length and look, and pictures, text files or voice notes. It never receives the conversation.

The tools

The KitCut plugin in ChatGPT and Codex

The plugin that ChatGPT's and Codex's plugin directory lists uses a second address of the same server, https://kitcut.ai/mcp/app. It is for making films and following them:

How it differs from https://kitcut.ai/mcp:

• It has no tools for projects or YouTube, and make_film takes no people. Those stay on the website and on https://kitcut.ai/mcp, which a person can still add to ChatGPT themselves in developer mode.
• Starting a film asks for a yes, and so does making one public. make_film and remake_film are marked as changes that cannot be fully undone, because a finished film's credits do not come back; set_film_public is marked the same way, because a film that was shown in the gallery cannot be reliably un-shown. The assistant confirms before each.
• Its answers name no plan and no page that sells. get_credits gives the credits left and the account's limits without the plan's name. An account short of credits is told what it has, when its credits renew and the longest film that still fits, and nothing else.
• Its answers carry less. Each tool there has an exact list of what it may answer, and anything else is left out: a film's project, the app it was made from, a template's maker and numbers, the studio's working notes, and all but the start of a stranger's idea in the gallery.
• Its film player is plainer. At most two buttons a view (Stop asks "Stop for good?" before it stops), no scrolling inside the card, and one fixed address, so an edit to the player never strands a chat that already shows it.
• Its links never lead to an offer. A film made through the plugin, and any film of a person who has the plugin connected, shows no "Remove branding" offer on its page at kitcut.ai. On a paid plan those films are cleaned like any other (Plans and credits).
• Sign-in is the same, and a token is good only for the address it was asked for. Disconnecting is on the account page under "Connected assistants".

Rules the tools follow

• Films are link-only by default. A film made through an assistant stays out of the gallery, unless the person asks for it to be public (public: true, or set_film_public later). Films are kept out of search engines either way.
• Nothing is sold in the chat. The tools never show prices or plans, and never link to them. A film longer than the account allows is refused with the account's own longest length and no upsell.
• As long as on the website. A film, a remake, a draft and a project's default length can be anything the account's longest film allows (get_credits gives it), the same as the website's slider.
• Ask before spending. The server asks the assistant to confirm the idea, length and look before starting a film when the person hasn't said them.
• Never claim a film is ready early. A film takes minutes to hours (How a film is made). The assistant gives the film's link at once and checks on it with get_film.
• Only the person's own account and the workspaces they are in. Every tool acts on the signed-in person's films, projects and channels, and on the projects of the workspaces they make films in, and does only what the website lets them do.
• Ask before publishing. The assistant confirms the channel and who can watch before a film goes on YouTube, unless the person already said.

Several films in a row

• One at a time, as on the website: an account makes one film at a time (2 on Pro). The studio also has its own line, shared by every account: a film may wait there for its turn.
• The account's line: make_film while the account is making all it may puts the new film in the account's own line instead of refusing it. It starts by itself about a minute after there is room, in the order it was asked for, through the same steps as any film. Its credits are held when it starts, not before; the films waiting count against the credit left, so a film is not lined up that the credit could not pay for. Up to 10 films wait at once. queue: false refuses it instead, as before.
• Seeing it all: get_queue gives every film being made now, each with its stage, what the studio is doing, its place in the studio's line when it waits there, how long it has been going and roughly how long it has left. Then the films waiting in the account's line, in order, each with roughly when it starts, and the films lately started from it, or that could not start and why (no credit left, the day's limit). The times are rough guesses from each film's length.
• Taking one out: unqueue_film takes a waiting film out of the line. Nothing was held for it. A film already being made is stopped with stop_film.
• What comes with it: a draft made this way leaves the project's drafts when its film starts, and a film asked for with youtube is scheduled as it starts.
• A film waits at most 23 hours. Its pictures and files are kept a day, so one that waited longer could not be made with them; it fails, saying so.

Shorts from a chat

• A film from an idea in a frame of its own: make_film with frame set to 9:16 makes it a YouTube Short: tall, made to be watched on a phone, with the narration's words drawn into the picture (a Shorts feed shows no subtitle track), and its first frame its own opening rather than a cover. 1:1 makes it square. Left out, it is 16:9, as before. It works for an episode of a project too, and for a film scheduled on YouTube as it starts.
• Its length: YouTube takes a vertical video of up to three minutes as a Short, so keep seconds inside that. A Short asked for inside it stays inside it, even when its narration runs long.
• Again: remake_film keeps the first film's frame, so a Short is remade as a Short, and localize_film makes it in other languages in the same frame (A film in other languages).
• A painted Short is painted tall: its scenes come in the film's 9:16 shape and fill the phone screen, painted by the same painter as a wide film (Looks).
• On the website: the idea box and a project's episode box have the same three frames, and Tall 9:16 there is a Short too (Look, length and frame).
• A project's frame: create_project and update_project take frame, the frame its episodes start in, as its Settings do. make_film in a project without frame uses the draft's frame, else the project's. save_draft keeps a frame with the idea.

Pictures and voice notes from a chat

• picture_urls: public https links to pictures.
  • At most 6 links.
  • Each link must answer within 15 seconds with at most 3 redirects.
  • Each picture may be up to 8 MB.
  • Private network addresses are refused.
• Files attached in ChatGPT: pictures only, passed by ChatGPT.
• upload_ids: uploads made in the in-chat film maker (open_film_maker), or on an upload page (get_upload_link).
  • The upload page works on any device and takes up to 12 files for 30 minutes.
  • It also gives an address for curl -T.
• What each picture is for: picture_roles on make_film (and remake_film) is a list of {picture, role}: the picture by its upload id or by its link in picture_urls, and one of three roles, as on the website's picture tiles (Pictures):

• A picture left out of picture_roles is used as Claude sees fit, as before. A film from a template takes no picture_roles.
• Naming a picture in the idea: the pictures are @picture1, @picture2 and so on, in the order the film gets them: files attached in ChatGPT, then upload_ids, then picture_urls. Voice notes and text files are not counted. The idea says where each goes ("@picture2 on the last frame"). A picture with the role show that the idea never names, or an @picture the film does not have, is refused before any credit is used, with what to add.
• Claude can't pass a picture pasted into the chat to a tool. Use the in-chat film maker or an upload link instead.
• Voice notes: by link (voice_note_urls) or an upload page. The in-chat film maker has no recorder.
• Text files (.txt, .md): on an upload page (get_upload_link), or with curl -T and ?name=. Not by link. An assistant can also put a long brief straight into the idea, up to 30,000 characters.
• Limits: the per-film limits are the website's: 9 pictures, 3 text files and 3 voice notes (Limits).

Projects from a chat

• Naming a project: by its id (p-…), its kitcut.ai/studio/ link, or its name.
• Projects in other workspaces: list_projects lists the person's own projects, then those of each workspace they make films in, under its name. An episode of one uses that workspace's credits, and can be as long as that workspace's plan allows, as on the website. get_credits gives each workspace's credits and longest film too, and open_film_maker with project opens the film maker for an episode, with that workspace's lengths.
• An episode is make_film with project. It reads the brief and gets the project's pictures, cast (characters, places and props), brand and kept voice lines. Its length and look are the project's unless the person says otherwise.
• The cast: list_characters with project gives the project's characters: each one's name, description, how many episodes it was in, and whether it is drawn (code the episodes draw it with) or painted (a picture painted episodes are painted from, up to 4 characters in one painting). Painted characters are added on the project's Cast tab, or kept by a painted episode (Painted characters). forget_character with project leaves out either kind.
• @name in an episode's idea: make_film with project and @mom in the idea puts the cast member mom in the episode, as the cast has them. The name is the one list_characters gives.
  • A near miss (@mum when the cast has a mom) is refused with "did you mean @mom?", and no credits are used. The assistant fixes the idea and calls again.
  • Any other @word, a handle or an e-mail address, stays in the idea as written. Outside projects, @ is plain text (Characters in an episode's idea).
• A project picture's role: add_project_picture takes role (show, scene or style), as the Pictures tab's "What it is for" (Pictures). With only project, name and role, it changes the role of a picture the project already has. get_project lists each picture with its role.
• Before the first episode, put the real logo in with add_project_picture named logo, and write the brief. A project without them invents its own emblem, and later episodes keep it.
• Drafts: save_draft keeps an episode idea. make_film with project and draft makes it, and the draft leaves the list, as on the page.
• Voice lines: list_episode_lines gives a finished episode's narration line by line. keep_voice_line keeps one the person has heard and approved; the next episodes play that recording for the same words. The Studio's Voice tab plays the lines (Characters and projects).

Templates from a chat

• Finding one: list_templates lists the public templates (optionally one category); get_template takes a template's slug, id or kitcut.ai/templates link (a link-only one too), says what it is, gives an example of what to ask for, and says how many pictures it takes and its frames.
• Using one: make_film with template (its slug or id), idea and frame (16:9, 1:1 or 9:16, as the template offers; its first frame if left out). A template is an example, not a form: idea says what the film is for, and a few words are enough (an event's name or its website). Pictures are optional and passed as for any film (picture_urls, files attached in the chat, upload_ids), up to the template's own number. Text files and voice notes can be attached too. Claude finds the rest itself and changes whatever is asked; the length and look are the template's, and project, draft and people are refused. The film is link-only unless the person asks for it to be public. See Templates.

YouTube from a chat

• Channels are connected on the account page, where Google asks which channel. list_youtube_channels shows the connected ones: a person's own and their workspaces', with the projects on each, the films planned there, and whether KitCut may read its numbers and change its videos.
• Several channels at once: every tool takes a channel by its id, title or @handle. publish_to_youtube and schedule_film take channels, a list, or ["all"]; each channel gets its own title and description in its own voice, and its own next free day. get_youtube_schedule and get_youtube_views show every connected channel side by side when none is named. Up to 8 channels a call.
• The channel: the one named, else the film's project's channel, else the only one connected.
• The words: draft_youtube_post writes the title, description and tags from the film, in the voice of the channel's latest videos. It also makes four thumbnails, and shows them as pictures where the app can. Asking again is free.
• Publishing: publish_to_youtube is unlisted unless the person says otherwise; publish_at schedules it instead (private until then, when YouTube makes it public). It uses thumbnail 1 unless another is picked, or "youtube" lets YouTube choose. The assistant confirms the channel and who can watch first, unless the person already said.
• Playlists: an episode goes into its project's playlist on the channel by itself, as from the publish dialog (Playlists). publish_to_youtube and schedule_film take playlist: a playlist's title or id on that channel, or "none". A name the channel does not have is answered with the channel's playlists. The channel must have been connected with the permission to manage its videos.
• A film goes to a channel once: publish_to_youtube is refused for a film that is already on that channel, or on its way there, and says which video is there. With again set to true, when the person asks for it, a second video goes up (a film's new version goes up this way) and the first one stays (A film goes to a channel once).
• Scheduling a film while it is being made: schedule_film puts a film on one or more channels' calendars for a day and time (publish_at, or "next" for each channel's next free day). A finished film is published at once, scheduled for that time. A film still being made holds its day: the calendars show it as planned, and the next free day is worked out after it. The minute it is finished, KitCut writes its title, description and thumbnails for each channel (unless they were given), and publishes it, scheduled for that time. If it is finished too late for that time, it goes at the same time on the next day that is free. make_film with youtube does the same as the film starts. Scheduling it again on a channel moves its day; unschedule_film takes it off. A film that does not finish is not published, and its plan says so.
• Scheduling: get_youtube_schedule gives a channel's calendar as the publish dialog shows it: what it published lately, what is scheduled, the pace it keeps (the series' own pace for an episode) and the next free day at that pace. publish_to_youtube with publish_at set to "next" takes that day, so a run of episodes scheduled one after another each get the following one. move_youtube_schedule moves a video that is already scheduled to another day and time. The pace and the suggested day are KitCut's own reading of the channel's publish dates, not YouTube's (Scheduling).
• Following it: get_youtube_post reports sending, processing, live or failed, and whether the thumbnail is on. A publish also finishes with nobody asking: KitCut checks on it every minute.
• A project's own schedule: update_project (or create_project) with auto_publish, such as "18:00 Europe/London", has every finished episode scheduled on the project's channel by itself, public at that time the day after it is finished, as the project's Settings do (Publishing every new episode automatically). "off" turns it off. It needs the project's channel, and it publishes as the person who turned it on.
• A channel's numbers: list_channel_videos gives the channel's totals (subscribers, views, videos) and its latest videos with their views, likes and comments; with order set to most_viewed it ranks the latest 50 by views.
• YouTube's own numbers: get_youtube_views reads YouTube Analytics for a connected channel, every video on it included, not only the ones from KitCut. It gives views, minutes watched, how long a view lasts and how much of a video is watched, subscribers gained and lost, likes, comments and shares, over the last 7, 28, 90 or 365 days, beside the same number of days before. For one channel it adds its views a week, its most watched videos in that time, where the views came from (search, suggested videos, browse features…), and the YouTube search terms that brought them: the top 25 terms people typed before watching, each with its views, minutes watched and share of the views from search, which is how a channel picks the next topics and titles people already search for. With video, post or film, it gives one video's numbers, with that video's own search terms. If YouTube does not give the terms, the answer says so and the rest of the numbers still come. Nothing of it is kept. It needs the channel's permission to read its numbers; a channel connected without it is named, with how to connect it again.
• One video in full: get_youtube_video takes a video's YouTube link or id, a post id, or a film. It gives the title, description, tags, length and who can watch it, YouTube's numbers, and, for a video KitCut follows, YouTube's own numbers of its last 28 days and how long a view lasts. It also names the film it was made from and that film's views on KitCut, labelled as KitCut's. Only videos on the person's connected channels can be read.
• Thumbnails: draft_youtube_post makes four from the film and shows them as pictures. set_youtube_thumbnail puts one of the four on a video that is already up, or the person's own picture by link (picture_url: a JPG or PNG of 2 MB at most; 1280×720 is best), on each of the film's videos or the one on a named channel, and shows the picture it set. A picture with no public link (a file the assistant made or holds) goes in two calls: upload: true answers with an address to send it to with curl -T, good for 30 minutes and a few pictures, and changes nothing yet; once it is sent, thumbnail: "own" puts it on. Before a film is published, the picture is kept for it, and thumbnail: "own" uses it when it goes up; publish_to_youtube and schedule_film also take thumbnail_url directly. get_youtube_video shows a video's thumbnail as YouTube has it. YouTube lets a channel use its own thumbnails only once the channel is verified.
• Changing a published video: update_youtube_video changes the title, description and tags on YouTube, or who can watch it. privacy set to public on a scheduled video publishes it now; unschedule takes it off the schedule and keeps it private. It works on videos KitCut published or linked to a film, on a channel that gave its fourth permission, as on a film's page in the Studio. The assistant confirms the change first, unless the person already said.
• The same limits as the page: 20 publishes a day per account, and publishing only where it is open (YouTube).

Planning the next videos from a chat

• How the films are doing: get_views gives the Views page's numbers for the whole account, one project or one film, over the last 7, 28 or 90 days or the whole time. YouTube's own numbers (views, minutes watched, subscribers gained, how much of a video is watched) and KitCut's page views are kept apart, each with the period before. The films come ranked by views.
• Everything to plan from, in one answer: brainstorm_videos gathers, for a project (or the whole account):
  • the brief, the characters and pictures every episode gets, and whether episodes are published by themselves;
  • the latest episodes, each with the start of its idea and how it did;
  • the drafts already saved, and what is being made;
  • the channel's totals, its most viewed recent videos, what is scheduled, its pace and the next free day;
  • the credit left.
• The ideas are the assistant's and the person's. brainstorm_videos writes none itself. The assistant thinks them up from it with the person, save_draft keeps each one they like, and make_film with the project and the draft starts it. Once it is finished, the project's own schedule puts it out, or publish_to_youtube with publish_at set to "next" takes the channel's next free day.
• A ready-made prompt: the server offers the prompt plan_videos (a project, how many ideas, and what they should be about). Apps that show an MCP server's prompts list it with the server's tools. It walks through the whole loop: what worked, the ideas, the drafts, the film, and its day on the channel. It asks before anything that uses credits or publishes.

Research films from a chat

A research film answers one question people are already searching for, from its sources, with every number labelled and checked, the evidence on screen, both sides, and what the data does not show. An assistant can make one from start to finish, on a schedule too:

• The method: get_playbook gives it, step by step: choosing a topic by what people search for, researching it with the page and the exact words behind every number, writing the three documents the film is made from in the shapes the checks read, checking them, making the film and publishing it. The server also offers the prompt research_film (a project, the subjects to look for, and whether to publish).
• A project for them: create_project with format: "research" writes the brief: the research look (a RESEARCH strap and the date of the data, a label on every number, explainers drawn for newcomers, charts built on screen, pages read under a moving camera), the voice, and the end on the open question.
• The documents: make_film takes up to 3 documents as text: the fact sheet, the chart data and the script. A script written word for word (lines - L01. with the spoken words in quotes) is said exactly as written: the narration is checked against it as the film is made, and the film is as long as its lines take.
• Checks before any credit is spent: check_research reads the three documents together: every number the narration says must be on the fact sheet, every fact needs a label, a source, a link and a quote, no phrase the fact sheet forbids may be said, and every chart value must be a fact. It also says how long the script runs. check_sources reads each fact again on its page, 20 at a time, and says whether the quote is there, something close is, only its number is, or it is not. Neither uses credits; pages are read on the public internet only, and no login or paywall is got around.

Not available through assistants