AI导哈

Leonardo AI 图像生成 MCP 服务

连接 Claude 与 Leonardo AI 的 MCP 服务器，提供文本生成图像、图像到图像、图像转视频、超分辨率放大及模型发现等功能，适用于通过 Claude 进行 AI 图像创作和视频生成。

  • by lildaddyo
  • TypeScript
  • 混合部署
  • 更新于
在 GitHub 查看 lildaddyo/leonardo-claude-mcp

使用说明

Leonardo AI MCP Server

MCP server connecting Claude to Leonardo AI for image generation, img2img, animation, upscaling, and model discovery. Deploys to Railway (same pattern as Higgsfield MCP).

Tools (9 total)

Tool What it does
leonardo_get_user_info Account info + API credit balance
leonardo_list_models List all platform models + IDs
leonardo_generate_image Text-to-image with polling
leonardo_get_generation Poll / fetch a generation by ID
leonardo_image_to_image img2img with init image + prompt
leonardo_upload_init_image_from_url Create S3 upload slot for init images
leonardo_image_to_video Animate an image (Leonardo Motion)
leonardo_upscale_image Upscale a generated image
leonardo_delete_generation Delete a generation

Environment Variables

Name Required Description
LEONARDO_API_KEY ✅ Your Leonardo API key
TRANSPORT ✅ http for Railway, stdio for local
PORT ✅ 3000
MCP_ACCESS_KEY ✅ for http Secret that callers must present (see Authentication). Several keys may be given comma-separated to rotate without downtime. Use a long random value, e.g. openssl rand -hex 32
MCP_REQUIRE_AUTH true = refuse every /mcp request (503) while MCP_ACCESS_KEY is unset, instead of serving it unauthenticated
LEONARDO_MAX_JOBS_PER_DAY Spend guard: max paid jobs (image, img2img, video, upscale) per UTC day. Default 200
LEONARDO_MAX_ULTRA_PER_DAY Spend guard: max Ultra image jobs per UTC day. Default 25
LEONARDO_MAX_VIDEO_PER_DAY Spend guard: max image-to-video jobs per UTC day. Default 25
LEONARDO_MAX_JOBS_PER_MINUTE Spend guard: burst rate limit for paid jobs. Default 15
LEONARDO_MIN_CREDIT_BALANCE Spend guard: refuse Ultra/video jobs when the API credit balance is below this. Default 0 (off)

The spend guard is in memory and per process: counters reset at 00:00 UTC and on restart. Setting a cap to 0 blocks that kind of job.

Authentication (HTTP transport)

Every tool spends or reads this server's own Leonardo account, so the HTTP endpoint must not be public. With MCP_ACCESS_KEY set, a request is accepted only with one of:

  • Path token (works for claude.ai URL connectors): https://<your-domain>/mcp/<MCP_ACCESS_KEY>
  • Bearer header: POST /mcp with Authorization: Bearer <MCP_ACCESS_KEY>

Anything else gets 401. Keys are compared in constant time. GET /health stays public and reports auth: enforced | open | locked.

If MCP_ACCESS_KEY is unset the server still answers without a token (so an auto-deploy does not cut off an existing connector) and logs a SECURITY WARNING every 10 minutes. Set MCP_REQUIRE_AUTH=true to refuse requests instead.

Switching an existing deployment to auth without downtime:

  1. Generate a key and change the claude.ai connector URL to https://<your-domain>/mcp/<key> (accepted while auth is still open).
  2. Set MCP_ACCESS_KEY=<key> in Railway. After the redeploy, /health shows auth: "enforced" and plain /mcp returns 401.

The path token appears in proxy access logs, so treat those logs as sensitive and rotate the key if they leak.

Get Your API Key

  1. Go to app.leonardo.ai
  2. Left menu → API Access
  3. Click Create New Key
  4. Copy the key

Deploy to Railway (Baby Steps)

STEP 1 — Push to GitHub

git init
git add .
git commit -m "leonardo mcp"
git branch -M main
git remote add origin https://github.com/YOUR_USERNAME/leonardo-mcp.git
git push -u origin main

STEP 2 — Deploy on Railway

  1. Go to railway.app → log in
  2. New Project → Deploy from GitHub repo
  3. Pick leonardo-mcp
  4. Wait ~1 minute for build

STEP 3 — Set Environment Variables

In Railway → Variables tab:

Name Value
LEONARDO_API_KEY your-api-key-here
TRANSPORT http
PORT 3000
MCP_ACCESS_KEY output of openssl rand -hex 32

STEP 4 — Generate Domain

Railway → Settings → Networking → Generate Domain

You'll get a URL like: https://<service>-production.up.railway.app

STEP 5 — Connect to Claude

claude.ai → Settings → Integrations → Add MCP server

URL: https://your-domain.up.railway.app/mcp/<MCP_ACCESS_KEY>

Keep this URL private: it contains the key.

Tests

npm test builds the server and runs test/http-smoke.mjs: auth modes, id validation and concurrent requests, with a fake API key and no paid Leonardo calls.

Test in Chat

"Check my Leonardo AI credit balance"
"List available Leonardo models"
"Generate a cinematic pharma lab image with Leonardo, dark moody lighting, teal accents"
"Generate a product shot with Leonardo: floating VR headset on dark background, cyan holographic effect"

Use Together with Higgsfield

Claude can now orchestrate both in the same conversation:

  • Leonardo: Superior image quality, fine-tuned models, img2img, alchemy pipeline
  • Higgsfield: Cinematic video generation, DoP camera movements, lip-sync

Example flow:

  1. leonardo_generate_image → get a perfect product shot
  2. higgsfield_image_to_video → animate it with cinematic camera movement

相关 MCP