Claude Video Generation: How to Make AI Videos From Claude Through MCP

Claude video generation works through connected models: Claude cannot render footage itself, but over MCP it can run video models and hand back the finished MP4. We built that on Wireflow, the node-based AI canvas we make, as one pipeline that turns a shot brief into a keyframe, clips from two video models and an edited cut. On 16 October 2026 it ran end to end in 101 seconds for 176 credits.

Can Claude generate videos?

Not natively. Claude reads text and images and writes text and code, and none of its models output video frames. What it does well is the layer before rendering: scripts cut to a runtime, shot lists and prompts written for a specific video model, the same planning step behind the best AI storyboard generators.

Task	Claude on its own	What you add
Script, shot list, model prompts	Yes	Nothing
Motion graphics (animated text, shapes, charts)	Yes, as code	Claude Motion (Team and Enterprise beta) or Claude Code with Remotion
Realistic footage, people, camera moves	No	A video model such as Kling, Veo or Gemini Omni Flash
Voiceover and music	No	A voice or music model
Cutting clips into one video	Through code	Remotion, an editing API or an editor node

Anthropic's help article says its animation feature, Claude Motion, "doesn't create realistic footage or people". Our test of Claude motion graphics found the code route good for titles and explainers, not product shots.

How Claude calls video models through MCP

The Model Context Protocol (MCP) lets an app publish tools that Claude can call: Claude picks a tool, fills in its inputs and reads the result. Video takes from 30 seconds to minutes, so video servers start a job, return an id and let Claude check back. Our roundup of the best MCP servers for video editing covers the servers built for this today.

Most video connectors are tied to one vendor: Higgsfield's runs its own model library (Veo 3.1, Kling 3.0, Seedance 2.0 and others, per its guide) and HeyGen's makes avatar videos. Each call returns one clip from one model, so chaining image to video to edit means Claude carries URLs between calls. Our list of Higgsfield MCP alternatives covers more.

A workflow connector works one level up: Claude runs a saved graph in which the models already feed each other. Our MCP server exposes that through four tools:

• list_workflows finds the flows in your account.
• get_workflow shows a flow's nodes and the inputs it accepts.
• run_workflow starts a run and returns an execution id. This is the call that spends credits.
• get_execution returns the status and, once finished, the output URLs.

The pipeline we ran: one brief, two video models, one cut

The graph has four working nodes. Both video models start from one keyframe, so any difference comes from the model, not the composition, the idea behind comparing AI video models side by side.

1. Shot Brief. One text input that feeds every model.
2. Keyframe. Nano Banana Lite renders a 16:9 still from the brief.
3. Clip A and Clip B. Gemini Omni Flash 1.1 (5 seconds, 720p) and Kling O3 (5 seconds) each animate that same keyframe, with the brief as the motion prompt. They run in parallel.
4. Side by Side Cut. The Video Editor plays both clips back to back at 1920x1080 and labels each with its model name.

The brief, word for word:

Product shot for a cold brew coffee ad. A clear glass bottle of cold brew coffee stands on a wet dark slate surface, with ice cubes and a few roasted coffee beans around it. Soft window light from the left, shallow depth of field, photorealistic, 16:9. Camera: slow push-in toward the bottle. Motion: drops of condensation slide down the glass and one ice cube settles. No text, no logos.

That loop is the unedited Video Editor render; click it for the MP4. The edit is a node in the same graph, not a separate service as with the AI video editing APIs we have compared. Figures come from the run record.

Step	Model	Finished after	Credits
Keyframe	Nano Banana Lite	5 s	16
Clip A	Gemini Omni Flash 1.1, 5 s, 720p	40 s	80
Clip B	Kling O3, 5 s	65 s	80
Cut	Video Editor, 10 s, 1920x1080, 30 fps	101 s	0
Total		101 s	176

The pre-run quote was 176 credits and the run charged exactly that. The final render, which waits for the slower clip, took over a third of the time.

What the two models did with the same frame

Both clips kept the bottle, beans and window light and added running condensation. They differed in the camera move and in what each model invented, which no spec sheet, our Kling AI API page included, can predict for your shot.

• Gemini Omni Flash 1.1 made the stronger push-in. By 4.5 seconds the bottle fills the frame and its neck is cropped off. It also generated an ambient sound track, which we muted in the cut.
• Kling O3 kept the whole bottle in frame with a gentler move, but it added an ice cube floating inside the bottle that was not in the keyframe. Its file came back at 1284x716 rather than 1280x720, and with no audio track.

Neither is wrong, which is the point of running two: keep Kling's framing for a pack shot that must show the whole bottle, Gemini's for a close opener. Each model is one node, so swapping one changes nothing else; our Kling AI alternatives list is a place to pick a third.

Set it up and run it from Claude

The connector is a remote MCP server with OAuth sign-in, so there is no API key to paste, and it also fits the AI API with Claude Code integration route in a terminal.

1. Add the connector. In claude.ai or Claude Desktop, open Settings, then Connectors, choose Add custom connector and enter https://www.wireflow.ai/api/mcp. In Claude Code, run claude mcp add --transport http wireflow https://www.wireflow.ai/api/mcp.
2. Sign in and approve. Running flows spends credits; editing flows is an opt-in permission you can leave off.
3. Copy the flow. Open the workflow linked below and use it as a template so it sits in your account.
4. Ask in plain words. For example: "Run my Claude video generation flow with this shot brief, then give me the link to the finished cut."
5. Let Claude poll. It reads the inputs with get_workflow, starts the run with run_workflow and checks get_execution until it can reply with the MP4 link.

Claude can write the brief too: ask for one shaped like ours (subject, setting, light, camera move, motion, "no text"), check it, then have it run the flow. Our guide to connect Claude to video editing uses the same habit.

Vendor connectors vs a workflow connector

	Single-vendor video connector	Workflow connector
One call returns	One clip from one model	Every node in the graph: keyframe, clips, cut
Models	The vendor's library	Any model nodes you wire in, swappable
Image to video to edit	Claude passes URLs between calls	Wired once in the graph
Repeat runs	Claude re-enters the settings	Same graph, new brief
Billing	Vendor credits	Credits per node; the editor render cost 0 on our run

A vendor connector is enough for a single clip. A graph earns its place when the same shot type repeats, when you want two models on one frame, or when the result needs an edit before anyone sees it, as in the AI video generation MCP tools we have tested.

Costs and limits to plan for

These come from our runs; the pricing page lists current credit packs.

• Each 5-second clip cost 80 credits here. Longer videos are several clips cut together, so cost scales with clip count.
• Video takes time. Our clips finished 34 and 59 seconds after the keyframe. Claude has to check back, and a chat waits while it does.
• Content filters can block a prompt. On our AI short drama run, one shot was blocked three times before a reworded prompt passed.
• Two bills, not one. Claude's tool calls count toward your Claude plan's usage; the generation is charged to the account the connector signs in to.

Try it yourself: Build this workflow in Wireflow. Edit the Shot Brief and run Side by Side Cut to get the keyframe, both clips and the labeled MP4. From Claude, one run_workflow call does the same.

FAQ

Can Claude generate videos?

Not on its own. It writes scripts, prompts and animation code, and through an MCP connector it can run video models and return the file.

Does Claude have a built-in video model?

No. Claude Motion, a Team and Enterprise beta, makes motion graphics as code, and Anthropic says it "doesn't create realistic footage or people". Footage needs a text to video AI model.

Which Claude apps can use a video connector?

Custom MCP connectors work in claude.ai, Claude Desktop and Claude Code. On managed Team or Enterprise plans an admin may need to add or allow the connector first.

How much does it cost to make a video from Claude?

Our run cost 176 credits: 16 for the keyframe, 80 for each 5-second clip and 0 for the edit. Claude's own usage counts toward your Claude plan.

Can Claude edit the clips into one video?

Yes. An editor node in the graph cuts, labels and renders the clips, or Claude Code can write Remotion code for a programmatic video generation platform approach.

Can I swap in other video models?

Yes. Each model is one node, so you can replace Kling O3 or Gemini Omni Flash 1.1 with Veo, Seedance or another model on a node-based video generation canvas without touching the rest.

Conclusion

Claude does not make video itself, but it drives the models that do: it writes the brief, starts the run and returns the file. One brief through a graph showed clear differences between two models on the same frame in under two minutes, and Wireflow's connector puts that whole AI video pipeline behind one call.