Edit videos and create polished content using Cardboard's MCP server. Import footage from files or video links, trim and arrange clips, add captions, titles, music and voiceovers, search stock footage, generate media, inspect transcripts, preview edits, and export finished MP4 videos.
Cline is an open-source coding agent for the terminal and VS Code. Cline CLI adds remote MCP servers with one command and signs in to them with OAuth.
This guide connects Cardboard MCP to Cline through MewCP, so Cline can use Cardboard MCP's 51 tools, like list_projects, create_project and print_timeline, without your Cardboard MCP credentials ever reaching the app.
Cardboard MCP signs in through managed OAuth. MewCP stores the token encrypted, refreshes it before it expires and injects it only at call time. Your agent never holds it, so it can never leak it.
More apps
Crawl websites and extract clean structured data optimized for AI applications and automation. Collect content from the web at scale while maintaining high-quality outputs.
Extract structured data from websites and public web pages. Fetch content, parse information and transform online data into formats that can be consumed by applications, agents and workflows.
TinyFish MCP for running AI-driven browser automation tasks, deep web research, persistent browser login profiles, remote CDP browser sessions, and agentic wallet payments.
TypeSafe's Jev model for typed decisions — ask yes/no, multiple-choice or rubric-score questions about a block of state and get calibrated probabilities back instead of prose. Built for classification, routing and scoring steps inside a workflow
Other clients
Cardboard MCP works the same way in every client below. Pick yours for its own guide.
Free to start.
Scope access to the exact Cardboard MCP tools and accounts you allow, work and personal side by side. Anything outside that scope is refused at call time, even when the model asks for it.
Each call is logged: what ran, on which account, and whether it worked. Failures come back structured, so the agent knows whether to retry instead of guessing.
Every app sits behind one secure MCP URL, and there is no server for you to deploy or keep running. Add an integration in the dashboard and it is live on the next call, with no config to change and no new credentials to wire up. Tools load only when a task needs them, so your agent stays fast and focused however many apps you connect.
Run cline mcp, choose Authorize OAuth for mewcp, and approve access in the browser window.
Start a new chat and ask: "List the tools you can use from MewCP." You should see your connected apps. cline config mcp lists your servers.
The account’s Cardboard projects with their ids, the most recently edited first, as text that reaches only you: the person sees none of it. show_projects lists the same projects for them, in a card in the chat. Lists the newest 10 unless `limit` says otherwise, and `total` counts every one that matches; `search` keeps those whose name contains it.
Create an empty project. Where the new project goes: Personal or a workspace, by its exact name or id. With one choice, it is selected for you. With several and none selected, the result lists them: ask the person which one, then call again with its id. Never choose one for them.
The canonical timeline.tsx for a project: header (aspect, duration, media table, diagnostics, rev) and the tree. A timeline longer than one result comes back in parts by line, and `from` reads on. apply_timeline takes edits copied from this text. With `version`, the timeline of that set-aside version instead, to read only: bringing its changes across is edits to the project.
Report what apply_timeline would change for the same edits or source, without writing.
Changes the timeline. `edits` are exact { oldText, newText } replacements copied from print_timeline, with its header `rev`; text outside them stays as it is, and edits saved since that rev merge in item by item. An edit that does not match exactly once changes nothing, and print_timeline shows the current text. `source`, the whole file, is for rewriting most of the timeline.
Lint the timeline: overlaps, gaps, off-screen text, clips past their source, missing media.
Saved timeline revisions, newest first, and the project's set-aside versions: saves refused because another app held the project, or offline work that came back after it changed, each kept with who made it, where and when, for a person to bring across or dismiss.
Restore the previous timeline revision, or a specific rev from timeline_history.
Media in a project with arrival state. Uploading means the browser is still sending the file; its content cannot be read yet. Preparing means the original is stored and playback is being prepared. Decide whether to wait, work on other items, or tell the user. An original that has not been uploaded cannot be read yet. Cast reference-sheet stills are left out unless `all` is set.
Download what a link points at into a project: the video behind a post on YouTube, X, Instagram, TikTok or another video site, or a video, audio or image file. An image joins the project as a photo. It joins the project with the same thumbnails, waveform and proxy as any other media.
Screenshot a web page in a headless Chrome: the first screen, the full page, or one element. With a project the image is imported into it; without one it comes back in the result. With scroll it records the page scrolling instead, as a video clip for the project.
Video footage for a shot, found across stock libraries (Pexels, Pixabay), archives (Internet Archive, Wikimedia Commons) and, on plans with web footage, YouTube, ranked by how well each clip shows `query`. Each result has its length, license and the link download_media takes to bring it into a project. Video only; it finds no photos.
Tracks from Cardboard's music library, ranked by how well each fits `query`, `genre` and `mood`; a filter nothing matches is dropped and the result says so. Each has its length, tempo, energy, sections with their times, the stretches quiet enough to sit under speech, a volume for each kind of mix, and the link download_media takes to bring it into a project. Free.
Sound effects from Cardboard's library, ranked by how well each fits `query` within `category`. Each has its length, when it suits, a volume, and the link download_media takes to bring it into a project. Free.
What a media file is: kind, duration, size, frame rate, codec, audio.
One frame of a video or image at a source time, returned as an image.
Evenly spaced frames of a video tiled into one image, left to right then down, with the source time of each tile.
The audio of a media file drawn as an image.
Peak and RMS audio levels (0-1) sampled across a media file; for finding silence, beats or the loudest moment.
Clean voice, remove background music from speech, or split singing vocals and instruments. Creates two new project audio assets without placing them; both outputs share one usage-credit charge. Source ranges are in seconds, at most ten minutes. Place the desired returned media and mute original audio to avoid doubling.
The transcript of a media file’s audio (any video or audio in the project), with per-word timings, in seconds. It uses Cardboard’s transcription, and the result is kept with the project for every later call. `timeline: true` maps times through the clips that play the media instead of source seconds.
Find a phrase across every media of a project that has a transcript (from transcribe_media). Returns source times per media and timeline times through the clips that play them.
A <track kind="caption"> for the current timeline built from the media transcripts (transcribing any not cached yet, in seconds). Its items carry captionText and captionWordTimings, so later edits can restyle and retime them. With `apply`, the track goes onto the timeline above the footage and the result is a summary. Without it the timeline is not changed and the result is the JSX for apply_timeline; the first <track> draws on top, so a caption track placed after the footage draws under it and does not show.
For an edit, pass sourceVideo (a project video, optionally #from-to): the actual clip goes to Wan 3 and determines the duration. For a transition pass image and endImage: both frames go to Kling 3 Standard, or Pro at best quality. An unnamed model with video ref inputs routes to Seedance 2.5. Generate a video clip from a prompt with a hosted model and put it on the timeline. The clip takes its place on the timeline at `start` straight away, and the finished video fills it when the run ends. Quality `standard` (default, Kling: 5 or 10 s, 16:9 / 9:16 / 1:1, silent) or `best` (Veo 3.1: 4 / 6 / 8 s, 16:9 / 9:16, optional audio, costs more). `model: "wan-3.0"` takes any length from 2 to 30 s in one generation, 16:9 / 9:16 / 1:1 at 720p unless `resolution` says otherwise, with free synchronised audio; it cuts between listed scenes itself and keeps the same people across them, so a story of several scenes that fits in 30 s and starts from nothing is one Wan 3.0 run (name each recurring person once with their look, then the scenes in order) rather than a clip per scene. A 30 s take waits about twelve minutes. `model: "gemini-omni-flash"` is 3-8 s at 720p (1080p when `resolution` asks for it), 16:9 / 9:16, always with native audio, and holds a four-scene story in 8 s. `model: "gemini-omni-flash-1.2"` is the same model at about 1.7 times the credits a second and runs 3-15 s. `image` turns a project image into the first frame and routes on its own to Hailuo 02 (standard) or Kling 3 Pro (best), which animate a still faithfully; describe only the motion. `ref` lists stills and videos the model works from rather than starts on (a product, a place, a move to copy): name them `@Image1`, `@Image2`… and `@Video1`… in the prompt in the order given. `model: "seedance-2.5"` takes up to 9 stills and 3 videos (30 s of video in all; `clip.mp4#2-8` sends seconds 2 to 8, and each second sent bills like a generated one at a lower rate); Seedance refuses any ref showing a real-looking person, so it takes no cast; MiniMax H3 and Wan 3.0 take stills only. People who must come back in another video, or in a scene you will redo on its own: write them as `@Name` in the prompt and describe the action, not the person. The first time a name is used, pass `character: ["Name: who they are, hair, wardrobe"]`; the job draws a reference sheet, keeps it in the project (the timeline header lists the cast; the sheet stills stay out of media listings) and later shots just say @Name. Cast everyone who recurs, kids included, or they change every shot. Cast shots at standard run 5-15 s on MiniMax H3 straight from the sheet stills; at best, 3-10 s on Kling 3 Pro from a drawn first frame for one person, or 2-30 s on Wan 3.0 from the sheet for two. A shot with more than one @Name goes to best on its own unless `quality` or `model` is given. `motion` only applies to the first-frame recipe. Open a cast shot medium or closer on one action so the face is large enough to keep; a wide, fast, many-body frame comes back with the face deformed. Returns at once with the generation id; generation_status with `wait: true` waits up to about half a minute at a time and fills the pending clip when the run ends. Costs credits; the error says what to fix.
Generate a still image from a prompt with a hosted model and put it on the timeline as an image item. The still takes its place on the timeline at `start` straight away, and the finished image fills it when the run ends. Quality `standard` (default) or `best` (holds fine detail and spells lettering, signage and title text reliably; costs more). `image` lists project images (id or file name) to redraw or take from, such as a product shot to put in a new scene; `standard` takes up to 6, `best` up to 4, and an edit at `best` costs more than drawing from text. `durationSeconds` is how long the still sits on the timeline (default 5). Returns at once with the generation id; generation_status with `wait: true` waits on the run and fills the still when it ends. Costs credits; the error says what to fix.
The state of a generation started by generate_video or generate_image. `wait: true` waits up to about 30 seconds: once the run ends it fills the pending clip it was placed for, otherwise it says how far the run has got, and a later call waits again.
Narrate text with a hosted ElevenLabs v3 voice and place it on the timeline: a pending audio item is filled with the MP3, and its word timings are stored as the media’s transcript. `voice` is an ElevenLabs voice id or the exact name of one of the account’s voices (default: its most recent); `docs` topic `voice` has the detail.
Compose instrumental music with Sonilo and place it on a "Music" track: a pending audio item is filled with the finished track. `sections` change the mood at seconds into the track, not timeline time: the first starts at 0, each at least 5 s after the one before, the last at least 5 s before the end. `auto` builds them from the timeline instead: quiet where people speak (cached transcripts), energy elsewhere, an accent where a title appears. `dryRun` returns the duration and sections it would send, free, without generating. Costs credits.
Generate a sound effect with ElevenLabs and place it on an "SFX" track: a pending audio item is filled with the MP3. Place one per moment, with `start` at the timeline second it belongs to. Costs credits.
The account’s own voices for generate_voice: favourites, recent voices and its clones.
Search the ElevenLabs voice library, most-used voices first; needs a query or a filter. Page with `cursor`.
Silent stretches in a media file’s audio (offline). Source seconds.
Integrated loudness (LUFS), loudness range, true peak and mean/max volume of a media file (offline).
Shot changes in a video, as source seconds (offline).
Separate the subject of a video or image from its background on a hosted GPU (costs credits); video mattes find people, image mattes the main subject. Returns a grayscale matte file, white = subject: MP4 for video, PNG for an image. Cached per media, stretch and mode.
Composited frames of the timeline as the player shows them, returned as images.
Render the project in Cardboard's export pool, from the originals in Cardboard's storage. Hands back a job at once (an export takes minutes for a long edit): job_status follows the render and returns the export's id and a link to the file, and cancel_job stops it. Cardboard's card plays the project's newest preview or export and offers its newest export's download; show_preview opens it. For when the user wants the file; check an edit with look and verify_timeline instead.
Read a file of the recipe a project was started from: its brief (recipe/RECIPE.md), its data such as recipe/structure.json, the person's answers (recipe/inputs.json, which lists their own files under the media field as they upload them), a file you wrote, or one of its pictures, which comes back as an image. A folder lists its files.
Cardboard's documentation by topic. `cardboard` is Cardboard's editing guide, the one its own editor agent follows, with these tools in place of `cb`: the editing loop, elements, seeing and hearing an edit, captions, generated media, color, the player in the chat and replies after an edit. The other topics each cover a command or technique (source, look, code, synced-captions, motion-graphics, color-grading, 3d-scenes, ...). No topic lists them.
Waits on a job a Cardboard tool handed back in place of its result (export_video does after a few seconds, any other tool once it runs past about 40 seconds). Returns that tool's result once the job ends; a job still running after about 40 seconds comes back as it stands, with its progress (e.g. "Rendering 40%"), and a later call waits again.
Stop a running job: an export stops and leaves no file. Work that cannot stop once started says so and ends on its own; a generation always runs to the end.
Opens Cardboard's upload card in the chat, where the person picks video, audio or photo files (JPEG, PNG, HEIC, WebP or GIF) on their device and uploads them into a project; the card says when each has uploaded, and photos join the project as images. It is how footage and photos on the person's device reach Cardboard; video, audio and photos already attached to the chat go in with add_media instead. Documents such as a brief, script or shot list are not uploaded: the card takes none, and one attached to the chat is read there. On a phone, Photos hands over a copy of a video, which is fine for previews: before export_video, ask whether to replace those clips with their originals, and if so call this with `replace` for each.
Adds video, audio and photo files (JPEG, PNG, HEIC, WebP or GIF) attached to this chat to a project: Cardboard copies each into the project and prepares it as it does an upload from upload_media's card, so the person doesn't pick it again there. A photo joins as an image. The result names each file's media id, or why it wasn't added. Documents such as a brief, script or shot list stay in the chat, where you read them.
Opens Cardboard's player in the chat on a project's latest edit, and on its export once one finishes: the person watches the export there and saves the file with its Download button. apply_timeline, undo_timeline and caption_from_transcript with apply show nothing in the chat by themselves, so this player is how the person sees what they changed: one call once a reply's edits are done shows them the result.
How a project's uploads, edits, preview renders, playback in the chat and exports went in the last 6 hours, as the person's cards showed them, and why something failed when Cardboard can tell. `now` is where things stand; `events` is what happened, oldest first, each with how long ago. In an event, `status` and `cause` are Cardboard's verdict (`cause` from a fixed list); `error` is the failing part's own message, as evidence; `measurements` are the numbers behind it; `involves` names the file, timeline item id or second it concerns; `retryMayHelp` says whether the same work again could go differently. Its fields are for working out a cause. The person sees the cards, not this: what they want from it is a sentence or two on what they'll notice and what happens next, not the events, codes, field names or timings. Opens no card.
Lists the person’s Cardboard projects for them, in a card in the chat: the most recently edited first, each with a picture from its footage, how long it runs and when it was last edited. They can search the list there and pick a project. The result tells you how many there are and names the newest six, with when each was edited. In ChatGPT a pick puts "Let's work on …" in their message box for them to send; in Claude the card turns into that project's player and notes which project they opened, which reaches you with their next message.
For Cardboard's card: a fresh session and the project's latest preview and export.
Sends a report to Cardboard's team: something in Cardboard that went wrong (`bug`), something the person asked for that Cardboard can't do yet (`request`), or anything else the person wants the team to hear (`feedback`). People at Cardboard read each report with the person's name and email, so it goes only once the person says yes. The result gives the report's reference.
The person's Cardboard subscription and what it has left: exports, usage, generative credits and when they renew. For questions like "what plan am I on?", "how many exports do I have left?" or "I want to upgrade", and before work their subscription may not cover.
What Cardboard can make: shows the person Cardboard's recipes in a card in the chat, the answer when they ask what Cardboard can make or do: finished edits by Cardboard's team, each a starting point that makes the person's own footage into that kind of video (a travel montage, a podcast clip, a brand launch film, captions for a talking video, an effect). Each plays its edit with its name, what it makes and its size; the person can open one to watch it with sound, switch to a kind, search them, and pick one there. When the person has said what they want to make, read the docs topic `recipes`, which lists every recipe with what it makes, and pass `picks`: only the recipes that make that whole kind of video, best first (one or two is fine; an effect, a title or a caption style is not a whole video), with `heading` saying what they are for; `also`: recipes of other kinds the same footage would suit, so they can change their mind; and `filters`: a few short guesses at what they want, each with the recipes it fits. The card shows those first and keeps every other recipe a tap away. Call show_recipes without them first if you do not yet know the slugs. `search` is for words the person asked to search for, not for finding what fits their idea: it finds recipes by what the words mean as well as the words themselves, and leads the card with them. The result names the matching recipes with their slugs, for start_recipe. A recipe marked usesGeneration makes some shots with generation, which costs credits.
Makes a new project from one of Cardboard's recipes: the recipe's brief, graphics and data in its recipe/ folder, the recipe's own media in the project, and the person's answers in recipe/inputs.json. When the recipe takes media, the result opens upload_media's card on the new project unless `upload` is false. `answers` holds the person's answers to the recipe's questions by id, as far as they have given them; the result names the questions still open. Where the new project goes: Personal or a workspace, by its exact name or id. With one choice, it is selected for you. With several and none selected, the result lists them: ask the person which one, then call again with its id. Never choose one for them.