Generate images and video directly from Claude, step by step

If you already do most of your work inside Claude, leaving it to make an image is a strange detour. You describe what you want to the assistant, then open another tab, paste the description into a different tool, download the result, and drag it back. MCP removes that loop. Connect a media server once, and Claude can generate images, video, voiceover and music as a tool call, right in the conversation where the idea came up.

This is a walkthrough of what that actually looks like with Dolly's MCP connector. If you want the background on what MCP is and why it exists, start with our explainer on MCP media tools and come back.

Step 1: connect the server

In Claude, add Dolly as a connector and approve it. Claude Desktop and claude.ai use the standard OAuth flow: you sign in with your Dolly account and the connection persists. For Claude Code or scripts, an API key works the same way. The MCP page has the exact server URL and current setup steps for each client.

Once connected, Claude can see a small set of tools: list the available models, quote a price, generate, check on a running job, and manage uploads. You do not need to memorize any of them. You just ask for what you want.

Step 2: ask, and check the price first

Try something like: "generate a photorealistic image of a ceramic espresso set on a walnut table, morning light." Claude picks a suitable image model and calls the generate tool. Two things are worth knowing about how this is built:

• Every generation has a spending cap. The generate call requires a maximum spend, so your agent states a ceiling on every request. It cannot quietly run up a bill.
• Quotes match charges. The quote tool and the actual charge run through the same pricing code. If Claude tells you a clip costs 250 credits, that is what it costs.

A typical image runs 5 to 15 credits (5 to 15 cents at face value). A default video clip is 250 credits. You can always ask Claude "what would a 10 second version cost?" before committing, and it will quote without spending anything.

Step 3: keep working with the result

The result comes back into the chat. From there the useful part starts, because the conversation keeps its context. "Make the lighting warmer" regenerates with your change. "Animate it" hands the image to a video model. "Now narrate this script over it in a calm British voice" brings in voice generation. You never re-explain what the project is.

Two shortcuts make chained work smoother. @last refers to your most recent output, so "animate @last" just works. And uploads can be saved as named assets: store your product shots once as @product, and every future generation can reference them by name. That sounds minor until the third campaign, because reusing the same references is how you keep results consistent across a whole series. There is more on that in the on-brand images guide.

What about ChatGPT and other clients?

MCP is an open standard, and support is spreading. Any client that speaks MCP can use the same server and the same account. Claude and Claude Code are the clients we verify against; developer tools like Cursor speak the same protocol, and as assistants like ChatGPT roll out their MCP support, the connector comes along rather than starting over. The practical benefit of the standard is that your media setup is not welded to one assistant. If you switch tools next year, the connector comes with you.

For scripted work outside a chat, the same account also drives a command line tool. Same credits, same models, same prices.

When the chat is the wrong place

Honest limits: if you are iterating visually on a dozen variants, a chat transcript is a clumsy gallery. The studio is better for heavy visual triage, and both surfaces share your account and history. The pattern that works for most people is to explore in whichever surface fits the task, then let the agent handle the repetitive generation once the direction is set.