Ask explicitly to use Cast to create presenter videos inside your agent chat. Generic video requests continue to use the regular Melius canvas workflow. Choose or create an avatar and describe your video. Your agent writes the script, validates the settings and generates through your client’s normal tool approval.
Cast tools appear when you connect with a team. Access is checked again whenever a tool runs, including actions in the image picker.
This page covers how Cast works in chat. The tool reference lists every Cast tool, and Cast covers the Cast studio in Melius.
Choose an avatar
Connect to the hosted Melius MCP server and describe the video you want. Name an existing avatar—“make a UGC video of Loopy talking about chemistry class”—to use it directly without a picker. If you haven’t chosen an avatar, browse the library or create four candidates and select your favorite. Your agent writes the script in chat and infers the generation settings from your request.
Confirm your avatar in the inline picker. It collapses to a card showing the
avatar you selected. If your client puts the selection message in the chat box,
press Send to continue. Your agent uses that choice before preparing the
script; it does not pick a different avatar while waiting.
Naming an existing avatar in your first request already selects it. Your agent can prepare or generate using that name directly, without opening the library or asking you to confirm it. Matching ignores capitalization. If the name is missing or matches multiple avatars, the agent asks for clarification instead of substituting another avatar. After you confirm an avatar in a picker, your agent continues with that selection. If your client places the selection message in the chat box, press Send to continue. The agent can recover the named selection without opening another picker.
To make new avatars without describing anyone, ask your agent to generate some for you, up to 12 at once. It writes different people, each with their own name and voice, and shows them in one card. Every person starts ticked: untick anyone you don’t want, rename them, play and change their voice, then save them all in one step. Each avatar reads in the voice it was saved with.
Change voice opens the same voice picker as Melius, already narrowed to voices like that person’s. Search, filter by gender, age, language and more, switch between the Standard library and your team’s voices, play samples and heart favorites. In clients without inline cards, your agent lists the people and their voices in chat and changes a voice when you describe the one you want.
An avatar made from your own description comes with a suggested voice. Play it or choose another before you save.
Saved avatars use a headshot cropped from the same identity image used by the Cast studio. No extra image generation is needed.
Write the script
Ask your agent to draft or edit the script in chat. Draft-only requests stop before generation. For generation requests, it infers settings from your brief. The default is Seedance 2.5, your avatar’s saved voice, portrait 9:16, the model’s default resolution and no subtitles. An avatar saved without a voice reads in the studio’s default voice, as it does in the Cast studio. You can request any supported model, aspect ratio, resolution, subtitles, Auto voice or specific voice in chat. You can also ask for a tone, such as warm, deadpan or hype, or describe your own. Explicit settings take priority; incompatible choices need clarification. Technical settings are not shown as inline controls. Video style and speech delivery stay internal and do not appear as review fields.
Scripts support both spoken lines with visual direction and separate silent action sections. Your agent can add gestures, demonstrations and reactions when they help your brief. For example, ask: “Explain how the Moon formed, act out the impact silently, then continue speaking.” Actions appear in the script review before generation and are not read aloud. You can also supply exact dialogue and stage directions in chat.
For a name that needs a particular pronunciation, tell your agent how to say it: “Pronounce Emmaus as ee-may-us.” The agent can pass this as a pronunciation mapping in the video’s creation settings. Cast uses the spoken spelling in voiceovers and generated dialogue while keeping the original script and visual directions. Pronunciation hints guide the model but do not guarantee its delivery; listen to the finished video. Automatic subtitles transcribe the audio and may need spelling correction separately.
Voice, background and props
Voice, background, prop and template pickers appear only when you ask for them. Otherwise, the agent uses your brief, your avatar’s saved voice, an automatic scene, and any supplied selections or attachments. MCP Apps clients show avatar choices and the finished video with Download video inline. Other clients receive numbered choices and media links. Attachments require file access and transfer support in your client.
A video can hold up to six props, within the selected model’s total reference-image limit. Avatars and background images share that limit, so additional background images can leave room for fewer props. A single prop appears throughout the video unless an action places it in specific sections. With several props, each appears only in the sections whose action uses it, as in the Cast studio. For example: “Show the serum while I talk about mornings and the mask while I talk about evenings.”
Your Cast review includes the selected prop image so you can verify it before generation. For an image already attached or present in your Cowork folder, the agent uses that file directly without searching Assets. If your team has access to Assets, you can select an existing image as a prop. Semantic library search requires Pro or above; uploading an image from chat does not. Otherwise, attach an image in your chat or skip the prop. Attachments require a client that can access and transfer the file; the remote MCP server cannot read your device’s files itself.
Generate
Generation uses your MCP client’s native tool approval. With auto-approve enabled in the client, a request such as “make 10 videos using Loopy” needs no additional Generate click for each video. Your agent reuses the named avatar, prepares all requested scripts and checks their combined estimate, then submits every video without waiting for earlier renders. Your plan and the generation queue determine how many render simultaneously. Multi-video requests use one batch review and one live batch card containing every video, completion/failure counts, downloads and View all in Cast. Each request supports up to 20 videos. Single videos retain their individual card. Existing credits and usage limits apply. Keep the conversation active until all requested videos are submitted. Accepted renders continue on the server; the agent cannot resume an unsubmitted batch after the client closes.
Each finished video comes with its Cast link, shared the way Share in Cast shares it: anyone you send the link to can watch the video.
Clients that support MCP Apps show avatar choices and a video player with Download video. Other clients receive numbered choices and media links. Voice, background, prop and template pickers appear only when you explicitly request them.
In Apps clients, the card opened by an accepted generation tracks that video through rendering and subtitles, then becomes the finished player. Your agent does not need another result call. Text-only clients poll accepted videos and return media links. Loading cards update automatically without an elapsed-time counter. Progress checks never restart generation. Each wait call uses a short window; an unfinished window continues checking the same video. A failed video stops waiting and shows its error. Requested subtitles finish before Download video retrieves the file. If subtitles still fail after retries, the original is available with a warning and the subtitle credits are refunded. You can ask the agent to check an existing run again without creating another video or card. If a status check fails, the card reports that the current status is unavailable and provides View in Cast. Batch cards check only unfinished videos and slow their retries during outages.
Batches
Ask explicitly for Cast, for example: “Create 10 Cast videos using Loopy, each visiting a different European city.” Your agent prepares one review with all scripts, settings and the total estimated credits. Native MCP permission approves the batch, including your client’s automatic approval setting.
One batch card shows every video, ready/failed/pending counts, individual downloads and View all in Cast. If the server connection is interrupted, the card reports that current status is unknown and retries. It never generates replacements. Batches support up to 20 videos per request and retain normal credit and queue limits. Accepted videos keep running after a disconnect; unsubmitted items are reported separately.
History and remixes
Ask about past videos to see your own Cast history, or your team’s finished videos. To remix one, ask for the same video with your changes: your agent reuses its avatar, voice, tone, props, script and settings, and tells you anything it cannot carry over. You can also rename or delete your team’s saved avatars and backgrounds, and save a photo you already have as a background without generating one. Your agent deletes only when you ask it to, and confirms the name first.
Usage and credits
Cast follows the same MCP usage policy as the other tools. Ordinary Cast calls are recorded in your team’s MCP usage; cast_result, cast_wait and cast_batch_result polling have the same counting exemption as other generation polling tools. Image and video generation use your team’s credits. Checking an existing run never restarts the video. For a video requested with subtitles, Melius adds them once the original video is ready, whether or not anything is checking. A subtitle pass that fails or stalls is retried, up to five passes in total, and checking never starts an extra one. If some candidates fail, the available images remain selectable with their original option numbers. Last modified on November 16, 2026