Create images and videos from Claude with MCP

The short answer

Connect Claude to our MCP server, then ask it in plain words to create an image, a clip or a voiceover. It takes three steps:

1. Create an API key on the developers page. The key is shown only once, so copy it and keep it right away
2. Connect Claude Code to our server, at https://low-cost.ai/api/mcp, with one command you will find further down
3. Write to Claude, for example: "What would an 8 second Veo 3.1 Lite clip of a Tel Aviv street after rain cost? Show me the price and wait for my OK"

The server gives Claude a set order of work in which every generation starts with an exact, free price check, is sent with a price cap and ends with a link to the file, which is also saved in your library. The models and prices are the same as on the website: right now, for example, a clip like that costs ₪0.707.03 crVeo 3.1 Lite, 720p · 8s · with sound, and a medium quality image in GPT Image 2 costs ₪0.222.21 crGPT Image 2, 1K · 1 image · Medium quality.

What MCP is in plain words

MCP, short for Model Context Protocol, is an open standard for connecting AI applications to external systems. The standard’s official site compares it to a USB-C port: just as USB-C connects devices in one uniform way, MCP connects AI assistants to tools and data. According to the site, assistants such as Claude and development tools such as VS Code and Cursor support it.

Claude itself does not make images the way dedicated image generation tools do. According to Anthropic’s help center it builds diagrams and graphics in HTML and SVG code, but not photos or illustrations. An MCP connection gives it tools it runs by itself during the chat, and our server gives it nine of them:

The tools read the catalog live, so every model and every option in the studio is available through MCP too, with no extra setup.

API vs MCP vs CLI

All three ways work with the same key, the same models, the same prices and the same library. The difference is who sends the request:

In short, the API is the way for programmers, MCP is the way for anyone who works with an AI assistant, and the CLI is the quick way from the terminal. The connection itself costs nothing: you pay only for generations, and a price estimate is always free.

What you need before you start

You need an account with a credit balance and an API key. There is no subscription: you buy credits on the credits page, and every generation, on the website or through MCP, comes out of the same balance. You create the key on the developers page, and four things are worth knowing about it:

• The key is shown only once, when you create it. Copy it and keep it somewhere safe
• Give each key a name, for example after the computer or the tool that uses it, so you can tell it apart in the list. You can keep up to 25 active keys
• A key can create in your name and charge your balance, see jobs and files and delete files. It cannot create or revoke keys, change your password or delete your account
• Revoke a leaked key right away and create a new one. Tools that use a revoked key stop working at once

Connecting Claude Code with one command

In Claude Code, run this command once in the terminal:

claude mcp add --transport http lowcostai https://low-cost.ai/api/mcp --header "Authorization: Bearer <api-key>"

Put your own key in place of ⁦<api-key>⁩. On the developers page the command is ready to copy, and a key you created there fills in by itself. Three more things from the Claude Code documentation:

• The ⁦--scope user⁩ flag saves the connection for all your projects. Without it, by default, the connection is saved only for the project where you ran the command
• The command ⁦claude mcp list⁩ lists the servers you have set up, and inside Claude Code the command ⁦/mcp⁩ shows the connection status
• According to Anthropic, HTTP is the recommended way to connect to remote MCP servers, and that is exactly how our server works

Cursor and other MCP apps

In Cursor you add the server to the mcp.json settings file, in your home folder or in the project folder, with the address and the Authorization header, and the developers page has a ready snippet for that too. Other MCP apps that connect to a remote server over HTTP and send an Authorization header with your key can use the same address. The ChatGPT connection is not available yet, and the developers page marks it "later".

The price before every generation and the max_credits cap

The server sends Claude a set order of work, in which every generation starts with the price:

1. Choosing a model. Claude sees the list of models, and every option of the model it picked
2. Estimating the price. The estimate tool returns the exact price in credits, free and with no charge
3. Generating with a cap. The generate tool takes max_credits, the most it may charge, and Claude passes the price it got in the previous step
4. Waiting for the result. The job_status tool waits up to 25 seconds at a time, and Claude comes back to it until the generation is ready, then returns a link to the file

The cap is your protection. If the actual price is above it, the generation is refused before anything is charged, and Claude gets the real price back. If there are not enough credits, the generation is refused with a link to top up. On top of that, one account can send up to 30 generations a minute, across all the ways together.

It is also worth asking Claude to show you the price and wait for your OK before every generation, especially for video. And if a generation fails, all its credits go back to your balance automatically, as explained in the guide a failed or stuck generation.

Files: what goes in and what comes out

• Input files. A start image, reference images or an audio file go in as a public https link. You can upload a file to your library, up to 4 MB per file, and copy the link from there
• Results. Every file created comes as a link on our domain and is saved in your personal library, exactly like a generation from the website
• Storage. The first 1 GB is free, and each extra GB costs ₪10 a month from your balance. Claude can show your storage use and delete files you no longer need
• A generation that runs long. A generation that has not finished within two hours is checked one last time: if the result is ready it is saved to your library, and if not, it is closed and its credits come back

Examples of requests

Once connected, you write to Claude as you would to a person. A few examples:

• "Which video models are there, and what does a clip cost in each at the default settings?"
• "What would an 8 second vertical clip in Veo 3.1 Fast cost? Do not create it yet"
• "Create a medium quality image in GPT Image 2 of a coffee mug on a wooden table, and show me the price first"
• "Take the image at this link and turn it into a 6 second clip in Veo 3.1 Lite"
• "How many credits do I have left, and how much space do the files in my library take?"

Creating from the terminal with the CLI

The CLI is a single file with no dependencies that runs on Node 18 or newer. Download it and sign in with your key once:

curl -o lowcost.mjs https://low-cost.ai/lowcost.mjs node lowcost.mjs login <api-key>

The key is saved on your computer and sent only over https, and only to our site, even when you download a file from a link. Then check a price and create:

node lowcost.mjs estimate veo-3.1-lite "Tel Aviv street after rain" --duration 8 node lowcost.mjs gen veo-3.1-lite "Tel Aviv street after rain" --duration 8 --wait --out clips

The first command returns the price in credits with no charge. The second creates, waits for the result and downloads the file to the clips folder. In the CLI the price cap is set for you from the price estimate, unless you give one with the max flag. Every option of a model is a flag, and these are more useful commands:

An API for your software

If you build an app or an automation, you call the same services over HTTP, with the Authorization: Bearer header and your API key. The model list and the price estimate are open even without a key:

The CLI itself uses these addresses, so its code is a simple working example of the API. Whichever way you choose, the price is the same as on the website and the results are saved in the same library.

Questions and answers