Claude can generate Seedance video clips with sound through a free MCP connection, but the model behind it is Seedance 2.5 Lite, not Seedance 2.0. See the setup steps, tool parameters, polling and limits, plus the routes to take when you need Seedance 2.0 itself.
You type one sentence into Claude, and a couple of minutes later a short video with its own soundtrack comes back as a link. That is what connecting Claude to PicassoIA through MCP gives you, and it works as advertised. There is one detail most search results gloss over, though: the free video model behind that connection is Seedance 2.5 Lite, not Seedance 2.0. They come from the same family and behave differently enough that picking the wrong one costs you an afternoon.
This article sorts that out. You get the real setup steps, every parameter the connector's video tool accepts, the polling loop Claude runs on your behalf, the limits that stop a request, and the other routes you can take when the connector is not the right fit.
💡 Short version: the connector is free on the Infinite and Wonder plans, exposes four models, and lets Claude call
generate_video_seedancewithout you leaving the chat.
Once the connection is live, Claude can call list_models and get four entries back. Each one names the tool that runs it and says whether your account can use it.
| Model | Tool Claude calls | Output |
|---|---|---|
| PicassoIA Image | generate_image | Text to image |
| PicassoIA Image Editor Pro | edit_image | Prompt-based image edits |
| PicassoIA Video | generate_video_picassoia | Video from text or a first frame |
| Seedance 2.5 Lite | generate_video_seedance | Video with synchronized audio |
On the account used for this article, all four came back marked as included in the plan. The tool descriptions say the same thing in plain words: free on the Infinite and Wonder plans. The Seedance 2.5 Lite model page frames it as unlimited for Wonder members, a lightweight edition of Seedance 2.5 tuned for 480p and 720p.
So when someone searches for a free Seedance MCP for Claude, this is the thing that exists today. It is the newer sibling, trimmed for speed and volume. It takes text or an opening image and returns a clip with the audio already mixed in.
💡 Check your own plan. PicassoIA words its pricing in slightly different ways across its pages, so open the pricing page and confirm that MCP connections are included before you plan a project around them.
Seedance 2.0 is not one of the four models in the connector. It lives on its own model page, and it does things the Lite build does not. It accepts up to nine reference images for character consistency, up to three reference videos for motion style, and up to three audio clips for lip-sync work. You can pin both the first and last frame, and setting the duration to -1 lets the model pick a length. Two siblings sit next to it: Seedance 2.0 Fast and Seedance 2.0 Mini.
Here is how the three video options compare:
| Seedance 2.5 Lite (connector) | Seedance 2.0 (model page) | PicassoIA Video (connector) | |
|---|---|---|---|
| Reachable from Claude | Yes | No | Yes |
| Input | Text, first frame, last frame | Text, first and last frame, 9 images, 3 videos, 3 audio clips | Text or first frame |
| Resolution | 480p, 720p | 720p by default | 480p, 720p, 1080p |
| Longest clip | 15 s at 480p, 10 s at 720p | You set it, or -1 for automatic | 20 s at 480p, 5 s at 1080p |
| Audio | Synchronized, on by default | Synchronized, on by default | Synchronized, on by default |
If you need the same character across several scenes, or a lip-sync to a voice recording, Seedance 2.0 is the right tool and you work with it on its own page. For everything you do inside a chat, the Lite model is the quicker route.
Two things have to be true before the steps make sense. You need a PicassoIA account on a plan that includes MCP connections, and you need a Claude client that accepts connectors. Claude on the web and Claude Desktop both do, and Claude Code takes MCP servers from the command line.
Log in to PicassoIA and open the MCP page at picassoia.com/en/mcp/accounts. That is where connections are created and managed. The server address is not published on the public site, so take it from your account page rather than from a forum post. If PicassoIA already appears in Claude's connector directory, you can skip ahead and add it from there.
In Claude, open Settings, then Connectors. Pick PicassoIA from the list if it shows up, or choose the option to add a custom connector and paste the address from step 1. Approve the sign-in if Claude asks for one.
In Claude Code, one command does the same job:
claude mcp add --transport http picassoia <your-connection-address>
Open a fresh chat and ask: "List the PicassoIA models I can use." Claude should call list_models and answer with the four models above. If it reports no PicassoIA tools, the connector is probably switched off for that conversation. Open the tools menu in the chat and enable it.
The same connection also exposes get_account for plan and parallel limits, list_generations for your history, get_generation for checking a job, and cancel_generation for stopping one.
Start small. The defaults are 480p and 5 seconds, which is the quickest way to test your wiring. Try this message:
Use Seedance to make a 5 second clip: a hot air balloon festival at sunrise, dozens of balloons rising above misty green hills, the camera tilting slowly upward. Keep 480p and leave the audio on.
Claude calls generate_video_seedance, gets a job ID back, and checks on it until the clip is ready. You end up with a link to the finished clip, sound included. Once it looks right, ask for the same prompt again at 720p.
The video tool takes nine inputs, and only the prompt is required. Claude fills the rest from your wording, so it helps to know what exists.
| Parameter | Accepts | Default | What it does |
|---|---|---|---|
prompt | 1 to 4,000 characters | Required | Describes what happens in the video |
image | Image URL | None | Opening frame, which switches on image-to-video |
last_frame_image | Image URL | None | Closing frame, needs image as well |
resolution | 480p, 720p | 480p | Output sharpness and render speed |
duration | 5, 10 or 15 seconds | 5 | Clip length |
aspect_ratio | 16:9, 9:16, 1:1, 4:3, 3:4, 3:2, 2:3, match_input_image | match_input_image with an image, 16:9 without | Frame shape |
save_audio | true or false | true | Keeps or drops the generated sound |
seed | Integer | Random | Repeats a result when you reuse it |
| Prompt rewrite switch | true or false | false | Lets PicassoIA rewrite your prompt first, adding a few seconds |
Without an image, the tool runs text-to-video. Add an image and it becomes image-to-video: your picture is the first frame and the clip inherits its aspect ratio, so any aspect_ratio you pass is ignored. The image has to be a URL, which means a still generated earlier in the same chat makes a natural opening frame.
Add a last_frame_image as well and the clip animates from the first picture to the second. That pairing is the cleanest way to control how a shot ends instead of leaving the ending to chance.
The rule is simple: 480p goes up to 15 seconds, and 720p goes up to 10 seconds. A 15 second clip at 720p is outside what the tool allows, so ask for 10 seconds at 720p or 15 at 480p.
The sample renders on the model page took between about 100 and 190 seconds, with the longer clips on the slow end. Draft at 480p, and spend 720p only on the take you plan to keep. Reuse the same seed on the second run so you start from the same point and can compare.
💡 The numbers disagree across pages. The Seedance 2.5 Lite model page talks about clips of 5 or 10 seconds, while the connector's tool description lists 5, 10 or 15. Trust the tool schema Claude shows you, because that is what your request is checked against.
Two smaller switches are worth knowing. Setting save_audio to false gives you a silent clip, handy when you plan to lay your own music over it. And the prompt rewrite switch trades a few extra seconds for a more detailed prompt, which helps when your own description is thin.
Video generation is asynchronous, so the first call never returns a video. It returns a predict_id and an estimated time. Claude then calls get_generation with that ID on a schedule the tool itself dictates: first at half the estimate, again when the estimate is reached, and then every 2 seconds.
Each check returns one of four states:
You do not have to run this loop yourself. If Claude stops early, say "keep checking until it finishes and give me the final link."
Changed your mind mid-render? Ask Claude to cancel the job and it calls cancel_generation with the same ID. One catch: a video the GPU is already rendering can no longer be canceled, so cancel early or let it finish.
Plan for five concurrent predictions per account. That pool is shared across MCP connections and API secrets, so a batch script running in the background can starve your chat. Prompts are capped at 4,000 characters in both the connector and the API.
💡 Batch politely. Ask Claude to submit three or four clips, then collect them together. Filling all five slots leaves no room for a retry.
The tool describes the prompt as "what happens in the video," and the model page says cinematic, chronological descriptions work best. Name the subject and starting pose, then the motion over time, then the camera move, then the light.
| Weak prompt | Stronger prompt |
|---|---|
| A woman by the sea | A woman in a flowing red dress stands at the edge of a cliff. Wind lifts her hair and hem as she walks forward. The camera pulls back in a slow wide arc at golden sunset. |
| A busy kitchen | A sushi chef rolls a maki at a steamy counter. An overhead camera drifts down toward his hands. Warm restaurant light, soft clatter of dishes. |
Short prompts still work. One of the samples on the model page is just "a cat steals a fish from a supermarket and escapes," rendered at 720p for 5 seconds. Detail simply gives you more control over the shot.
Since audio is generated alongside the picture, mention sound the way you would brief a sound designer: footsteps on gravel, wind, a crowd murmur.
You have a writing partner already sitting in the chat, so use it. Three requests that work well:
If you would rather draft in PicassoIA itself, a model like Claude Sonnet 5 is available there for writing shot lists and prompt variations.
Most failures come from a short list of causes.
| Symptom | Likely cause | Fix |
|---|---|---|
| Claude says it has no PicassoIA tools | Connector off for this chat, or sign-in expired | Enable it in the tools menu, then reconnect in Settings |
| A 720p request for 15 seconds is refused | 720p stops at 10 seconds | Use 10 seconds at 720p, or 15 at 480p |
| Prompt rejected | Over 4,000 characters | Trim it and keep only motion, camera and light |
| Job comes back failed | The render itself failed | Start a new generation, since a failure is final |
| New jobs sit waiting | Five predictions already running | Wait, or cancel one you no longer need |
| A model shows as unavailable | Your plan does not include it | Check list_models and get_account |
💡 Do not retry in a loop. If the same prompt fails twice, change something: shorten it, lower the resolution, or swap the opening image.
Claude Code accepts the same connection with the command from step 2, which is handy when you want video generated as part of a coding session, such as placeholder clips for a landing page.
For scripts, PicassoIA also has a REST API. The base URL is https://api.picassoia.com/v1, and requests carry a bearer secret that starts with pia_sk_. You create secrets from the API section of your PicassoIA account, up to two per account. The endpoints follow the Replicate style:
| Action | Request |
|---|---|
| Create a prediction | POST /v1/models/{owner}/{name}/predictions |
| Fetch a prediction | GET /v1/predictions/{id} |
| Cancel a prediction | POST /v1/predictions/{id}/cancel |
| List predictions | GET /v1/predictions |
The model ID for the free video model is picassoia/seedance-2.5-lite. Jobs run asynchronously, so you create one, poll it, then fetch the result. Limits mirror the connector: five concurrent predictions, 4,000-character prompts, a 10 MB request body and a 3 hour timeout. The payload shape follows the Replicate-style convention these endpoints mirror, so check the API docs for exact fields before you ship anything:
curl -X POST https://api.picassoia.com/v1/models/picassoia/seedance-2.5-lite/predictions \
-H "Authorization: Bearer $PIA_SECRET" \
-H "Content-Type: application/json" \
-d '{"input": {"prompt": "A hot air balloon festival at sunrise", "resolution": "720p", "duration": 5}}'
The docs describe API predictions as free for now, yet they also mention an Infinite plan, so confirm what your own account allows.
Go to the Seedance 2.0 page when the job needs references. A few rules to remember there:
[Image1], [Video1] and [Audio1].For a longer run, look at Seedance 2.5, which the platform lists for clips up to 30 seconds. And if you want 1080p straight from Claude, PicassoIA Video through the same connector goes up to 1080p for 5 second clips.
You now have the full picture: a free connector that runs Seedance 2.5 Lite, a tool with a short list of parameters, and clear fallbacks when you need Seedance 2.0 itself. The fastest way to find out whether it fits your work is to run one prompt.
Open Picasso IA, connect it to Claude, and ask for a 5 second clip at 480p. Then build a still with PicassoIA Image, polish it with PicassoIA Image Editor Pro, and hand it to Seedance 2.5 Lite as the opening frame. One afternoon of experiments will show you which route suits your projects. Try it, change one variable at a time, and keep the takes you like.
