AI Tools

Can Claude Generate Images? How to Make YouTube Thumbnails with Claude

Claude can't draw, but it can call real image models. Here is the setup that turns Claude Code or Claude Desktop into a thumbnail generator.

No, Claude cannot generate images on its own. It can analyze images you upload and it can write SVG, HTML and charts, but it does not produce photos or illustrations natively, as Anthropic's help center confirms. To make a real YouTube thumbnail from inside Claude, you connect an image-generation tool through MCP.

What Claude Can and Can't Create

Claude is strong at reading images, writing code and producing markup. Since September 16, 2026 it can also create designs, decks and docs in any conversation through Claude Design, per Anthropic's release notes. Those outputs are layouts and documents, not photographic thumbnails. For a photoreal face, accurate headline text or a styled product shot rendered as a JPEG, Claude needs an external image model.

That is exactly what MCP servers are for. Claude Code and Claude Desktop can call a remote tool that runs the image models and returns finished files.

The Fix: Connect ThumbnailCreator's MCP Server

ThumbnailCreator.ai exposes a streamable-HTTP MCP server at https://thumbnailcreator.ai/api/mcp. You sign in at thumbnailcreator.ai, open the account menu and create a key under Agent keys, then authenticate with Authorization: Bearer <key>. The server has four tools: list_models, generate_thumbnails (a prompt, up to 4 models, up to 4 images per model and at most 8 images per call, in 16:9, 1:1, 9:16 or 4:5 formats - it returns image URLs valid for one hour), get_generation and account_usage. Every image is billed per image to the key's account.

Claude Code

Paste this in your terminal, replacing the placeholder with your key:

claude mcp add --transport http thumbnailcreator \
  https://thumbnailcreator.ai/api/mcp \
  --header "Authorization: Bearer tc_your_key"

Claude Desktop

Add this to claude_desktop_config.json:

{
  "mcpServers": {
    "thumbnailcreator": {
      "command": "npx",
      "args": ["mcp-remote", "https://thumbnailcreator.ai/api/mcp",
        "--header", "Authorization: Bearer tc_your_key"]
    }
  }
}

Agent Skill (optional)

If you want Claude Code to know when and how to use the tool without prompting it every time, install the Agent Skill:

mkdir -p ~/.claude/skills/thumbnailcreator
curl -o ~/.claude/skills/thumbnailcreator/SKILL.md https://thumbnailcreator.ai/skill.md

The claude.ai web and mobile limitation

One caveat: custom connectors in the claude.ai web and mobile apps sign in with OAuth, per Anthropic's connector docs. ThumbnailCreator uses API keys today, so use Claude Code or the Claude Desktop config file. On claude.ai web, you can still have Claude write and refine your prompt, then paste it into the web studio at thumbnailcreator.ai.

Which Model to Ask For

Different models have different strengths. Tell Claude which one to use, or let it run several at once and compare:

  • GPT Image 2 - the most reliable for readable text
  • GPT Image 2.5 Flare - the cheaper, faster option for text
  • GPT Image 2.5 Sunburst - the most precise detail
  • Nano Banana Pro - faces and reference photos
  • FLUX.2 Max - photoreal scenes
  • Ideogram 4 - bold typography
  • Grok Imagine - cheap exploration

You can also upload up to 4 reference images - your face or your product - and GPT Image 2, GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, Nano Banana Pro, Nano Banana 2 and Seedream 5 Lite will use them.

Example Workflow

  1. Ask for concepts: "Give me 3 thumbnail concepts for a video about my app finally compiling, with overlay text of 5 words or fewer."
  2. Generate: "Generate concept 2 with gpt-image-2 and nano-banana-pro in 16:9." Claude calls the MCP server and returns image URLs.
  3. Pick and download: open the URLs (they are valid for one hour), keep the winner, upload to YouTube.

A prompt Claude might write for that second step:

YouTube thumbnail: tired developer at a desk at 2 AM surrounded by coffee cups, huge bold text "IT WORKED" in yellow with a black outline, dark room lit by monitor glow

Claude Alone vs Claude + MCP vs ChatGPT Images

Claude alone (SVG / Claude Design) Claude + ThumbnailCreator MCP ChatGPT Images
Photoreal images No Yes Yes
Readable headline text As markup, not in an image Yes, rendered in the image Yes
Models available None for images 9 models OpenAI's image model
Output SVG / HTML layout 1920x1080 JPEG under 2 MB Image file
Cost Included in Claude plan Per image, $0.026-$0.195 (as of October 2026) Included in ChatGPT plan

What It Costs

Images on ThumbnailCreator cost $0.026 to $0.195 each depending on the model (as of October 2026). New accounts get $0.50 of free credit - no card needed - then $10 or $25 prepaid credit packs. Credits never expire and failed images are free.

Frequently Asked Questions

Can Claude generate images?

No. Claude can analyze images and write SVG, HTML and charts, but it does not produce photos or illustrations natively. To generate images from inside Claude, connect an image-generation MCP server like ThumbnailCreator's.

Does Claude Design make thumbnails?

Claude Design creates layouts, decks and docs - useful for mocking up a thumbnail concept, but the output is a document, not a photographic JPEG. For the rendered image itself you still need an image model.

Does this work in the Claude mobile app?

No. Custom connectors on claude.ai web and mobile use OAuth sign-in, and ThumbnailCreator authenticates with API keys. Use Claude Code or Claude Desktop instead, or generate in the web studio at thumbnailcreator.ai.

What does it cost?

$0.026 to $0.195 per image depending on the model (as of October 2026), with $0.50 of free credit for new accounts. Credit packs are $10 or $25, credits never expire, and failed images are free.

Make Your First Thumbnail from Claude

Connect the MCP server, ask for a thumbnail, and get back a YouTube-ready JPEG.

Start with $0.50 Free

No card needed. Pay per image after that.