MCP server

The Createa Meme API ships an MCP server, published on npm as @createa/meme-mcp. That means any MCP-compatible AI assistant (Claude Code, Cursor, Cline, Continue, and a growing list of others) can generate memes for you with no extra integration work — just paste your API key into the client config and ask.

You: "Make me a meme about Monday morning meetings"
Claude: [calls createa_meme_generate] Here you go: https://...

Quick install

Pick your client below, paste the snippet into its MCP config, restart, done.

Claude Code

~/.claude.json (or your project's .claude/mcp.json):

{
  "mcpServers": {
    "createa": {
      "command": "npx",
      "args": ["-y", "@createa/meme-mcp"],
      "env": {
        "CAM_API_KEY": "cam_live_xxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "createa": {
      "command": "npx",
      "args": ["-y", "@createa/meme-mcp"],
      "env": { "CAM_API_KEY": "cam_live_xxxxxxxxxxxxxxxxxxxx" }
    }
  }
}

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows. Same shape as Claude Code above. After restart, look for the 🔨 hammer icon in the input bar — clicking it shows the available createa tools.

Cline, Continue, anything else stdio-MCP

Same shape — command: npx, args: ["-y", "@createa/meme-mcp"], env.CAM_API_KEY.

Verify the install (works in any client)

Different clients show MCP differently — some have a /mcp slash command, some show a hammer icon, some don't surface the server visibly at all. The reliable test is just ask in natural language:

"Use createa to find meme templates about disappointment."

If install worked, the assistant calls createa_template_search and replies with a list of tpl_* IDs. Real templates returned = it's working. You don't need any UI.

If nothing happens or the assistant says it can't access createa, see Troubleshooting.

Tools exposed

Power tool: createa_meme_remix

The remix tool is the answer when you want to edit an existing image rather than generate from scratch. Pass an image URL plus an edit prompt and the model produces a modified version. Useful for:

• Character consistency: pass a reference photo, edit costume/setting/expression
• Style transfer: "make this corporate headshot look like a 90s anime"
• Caption-aware edits: "swap the background to a meme template scene"

Example: "use createa_meme_remix with the URL [...] and the prompt 'turn this person into a firefighter holding a hose, photorealistic'". The MCP picks up the tool, the API hits GPT Image-2, and you get a 1-2 minute round-trip back.

Brand watermarks (Starter+ plans)

Every generation tool — generate, caption, ai_image, remix — accepts an optional watermark. It stamps your brand onto the output. Provide either text or a logo image:

// Text watermark
"watermark": { "text": "Green Frog Labs", "position": "bottom-right", "opacity": 0.65 }

// Logo watermark
"watermark": { "logo_url": "https://yoursite.com/logo.png", "scale": 0.15 }

In practice you just say it in natural language: "…and watermark it with 'Green Frog Labs' in the bottom right." Watermarks are a Starter+ feature — free-tier keys get a 402, no per-call upcharge on paid plans.

Get an API key (free, 20 credits/month) at the dashboard.

ELI5 FAQ

What's an MCP server?

It's a way for AI assistants like Claude Code or Cursor to call external services. Think of it as a USB port for AI: instead of every assistant having to write custom integration for every service, anyone can plug into anyone. Anthropic invented the spec; everyone else adopted it.

Why does this matter to me?

You don't have to build anything. If you already use Claude Code or Cursor, installing our MCP takes 30 seconds and from then on you can ask your assistant for memes mid-conversation — same as asking it to read a file or run a command.

Do I still need an API key?

Yes — the MCP server is just a thin wrapper that hands your prompts to our API. You get a key free at dashboard.createa.meme, paste it into the config, and you're done. The free tier (20 generations/month) works for casual use.

Will this slow my assistant down?

Only when you actually ask for a meme — and the cost is the API call itself (typically 2-12 seconds depending on the endpoint). When you're not using meme tools, the MCP server sits idle.

What happens if I run out of credits?

The tool returns an error message and your assistant tells you. Nothing breaks. Your free credits reset on the 1st of each month.

Can my assistant accidentally rack up huge bills?

No. The free tier hard-caps at 20 credits/month. The most expensive call is meme_ai_image at 3 credits. Even if your assistant burned through every credit in one session, you'd see ~6 image generations max — then it stops.

Is my data safe?

The MCP server runs locally on your machine. Your prompts are sent to the Createa API over HTTPS. We log requests for billing + abuse-detection (per our privacy policy) but don't train models on your prompts and don't share them with anyone besides the model providers (OpenAI, Anthropic, xAI) that generate the captions.

What if I'm not using Claude or Cursor?

The protocol is open. Any MCP-compatible client works. The list is growing fast — Cline, Continue, Goose, and several smaller agents all support it today.

What can't this do yet?

Stuff on our roadmap — batch generation, GIF/video memes, branded watermarking. We'll add tools as those endpoints ship.

How do I uninstall?

Delete the createa block from your client's MCP config and restart. There's nothing else on your system — npx doesn't leave anything behind.

Troubleshooting

"The assistant says it doesn't know createa"

Most likely the config didn't load. Two checks:

1. Fully quit and relaunch the client. Running sessions don't pick up new MCP servers; on macOS this means Cmd+Q (not just closing the window).
2. Test the server runs standalone. This proves the binary works and your key is valid:

   CAM_API_KEY=cam_live_xxxxxxxxxxxxxxxxxxxx npx -y @createa/meme-mcp

   You should see one line on stderr:

   [createa-meme-mcp 0.1.0] connected via stdio

   Press Ctrl+C to exit. If it errored before that line, paste the output to us.

"JSON syntax error" or config silently failing

Most clients fail silently on broken JSON. Validate it: paste the file's contents into jsonlint.com — common mistakes are missing commas between entries or trailing commas.

"Invalid API key"

Mint a fresh one in the dashboard and update the config. Don't include the word Bearer — just the key starting cam_live_.

"The call timed out" (MCP error -32001)

Generation is synchronous and can run long — caption ~1s, generate 3-15s, ai_image 10-30s, remix 60-120s. Most MCP clients default to a ~60s per-request timeout, which remix can exceed.

If you're timing out on remix, do these in order:

1. Raise your client's tool timeout — this is the fix that reliably works. For Claude Code, launch it with MCP_TOOL_TIMEOUT=300000 in the environment (e.g. echo 'export MCP_TOOL_TIMEOUT=300000' >> ~/.zshrc, open a fresh terminal, relaunch). We verified empirically that current Claude Code enforces its ~60s cap even while receiving progress notifications.
2. Update the package — clear the npx cache (rm -rf ~/.npm/_npx) and restart your client so it pulls ≥ 0.4.0 (progress display, request queueing, honest timeout errors).
3. Don't blind-retry a timed-out remix — the generation usually still completes server-side (and bills). Check your dashboard for the finished image before retrying, or you'll pay twice for one meme.

Parallel heavy generations are queued (2 at a time) since v0.4.0 rather than racing each other into provider rate limits.

Package

Published on npm: @createa/meme-mcp

The source isn't currently public, but the install command above is the entire surface — npx fetches the published npm package and runs it on your machine. Bug reports and feature requests welcome at jaz@createa.meme.