@tinify-dev/mcp

MCP (Model Context Protocol) server for the Tinify.dev image API. Lets ChatGPT, Claude, Cursor, and other MCP clients compress, resize, crop, and convert images — with honest results.

You: compress the screenshots in ~/Desktop/launch/
Claude: 312.4 KB → 97.1 KB (-68.9%), wrote /Users/you/Desktop/launch/hero.min.png
        84.2 KB → 84.2 KB — logo.png is already as small as Tinify.dev can make it
        (optimized: false), so no file was written.

• Never lies about savings — when the API cannot shrink a file it says so (optimized: false) instead of writing a byte-identical "optimized" copy.

• Never overwrites your originals silently — results go to <name>.min.<ext> beside the input, or to an explicit output_path. Replacing a file requires overwrite: true.

• Raw numbers in structuredContent on every call, so agents can do math instead of parsing prose.

Not affiliated with TinyPNG. This server talks to the Tinify.dev API.

Requirements

• Node.js >= 20

• A Tinify.dev API key from tinify.dev/developers — free tier is 500 operations/month, no card required.

Related MCP server: tinypng-mcp-server

Setup

Claude Desktop

Add to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "tinify": {
      "command": "npx",
      "args": ["-y", "@tinify-dev/mcp"],
      "env": { "TINIFY_API_KEY": "tnf_live_..." }
    }
  }
}

Claude Code

claude mcp add tinify -e TINIFY_API_KEY=tnf_live_... -- npx -y @tinify-dev/mcp

Cursor

Create .cursor/mcp.json in your project (or ~/.cursor/mcp.json globally):

{
  "mcpServers": {
    "tinify": {
      "command": "npx",
      "args": ["-y", "@tinify-dev/mcp"],
      "env": { "TINIFY_API_KEY": "tnf_live_..." }
    }
  }
}

Remote server (hosted)

The same five tools are available as a hosted streamable-HTTP MCP server - no install, works from ChatGPT connectors, the claude.ai directory, and any registry that expects a URL:

Endpoint:  https://api.tinify.dev/mcp
Auth:      OAuth 2.1 (PKCE + dynamic client registration), or
           Authorization: Bearer tnf_live_...   (or tnf_test_...)

Two ways to authenticate:

• OAuth (for connectors) - ChatGPT connectors and the claude.ai directory only speak "OAuth" or "no auth". Add the server by its URL (https://api.tinify.dev/mcp) and pick OAuth; the client discovers the authorization and token endpoints automatically from the server's .well-known metadata, registers itself, and opens a Connect Tinify page where you paste your Tinify API key. Your key stays the credential - the client only ever holds an opaque token that maps back to it server-side.

• Direct bearer (for scripts/CLIs) - send Authorization: Bearer tnf_live_... (or tnf_test_...) and skip OAuth entirely. Unchanged.

ChatGPT developer-mode connection

1. In ChatGPT, open Settings → Security and login and enable Developer mode.

2. Open ChatGPT Plugins, select the plus button, and add https://api.tinify.dev/mcp.

3. Complete Connect Tinify with a Tinify API key.

4. Add the connection from the conversation's tools menu, attach a PNG/JPEG/WebP/AVIF, and ask:

   Use Tinify to compress this image for the web. Show the original size,
   result size, and percentage saved.

The hosted server cannot access local paths, so each image tool accepts exactly one of:

• image — the ChatGPT attachment object. ChatGPT fills this automatically because the descriptor advertises openai/fileParams.

• image_base64 — a portable fallback for other MCP clients, capped at ~28 MB decoded.

ChatGPT attachment downloads are HTTPS-only, reject private/reserved network targets and redirects, time out after 30 seconds, and are capped at the Tinify API's 40 MB limit. Successful operations return exact byte metrics and a temporary MCP result link. Existing base64 callers also receive structuredContent.image_base64 for backwards compatibility. get_usage is identical to the local version. Nothing is written to the user's filesystem.

Try it with curl:

curl -s https://api.tinify.dev/mcp \
  -H "Authorization: Bearer tnf_live_..." \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"curl","version":"0"}}}'

For local files, prefer the stdio server above - it reads and writes them directly with no base64 round-trip and no 28 MB cap.

One server, two transport adapters

This repository intentionally supports both OpenAI and Claude:

• The shared tool names, Tinify client, result metrics, errors, and annotations are platform-neutral.

• The stdio adapter uses absolute local paths and writes files for desktop/CLI clients such as Claude, Cursor, and Codex.

• The hosted adapter uses ChatGPT file attachments or base64 and returns temporary result links because hosted servers cannot read a user's filesystem.

An OpenAI-only repository would duplicate the Tinify logic and make behavior drift more likely. OpenAI-specific descriptor metadata stays as a small additive layer in the hosted adapter; a separate repository is not needed.

For public OpenAI submission, set the portal-provided domain token as OPENAI_APPS_CHALLENGE_TOKEN in /etc/tinify/mcp-http.env and deploy the matching nginx location from deploy/nginx-location.conf. The endpoint returns only that token at /.well-known/openai-apps-challenge; do not commit the real portal token.

Tools

All image tools require absolute paths (MCP servers run with an unpredictable working directory) and pre-check the 40 MB API limit locally. API errors come back as readable tool errors with the error code and request_id to quote to support.

Example prompts

• "Compress every PNG in /Users/me/site/static/img"

• "Resize /Users/me/photo.jpg to 1200px wide and tell me how many bytes it saved"

• "Convert /Users/me/hero.png to webp with best quality"

• "How many Tinify operations do I have left this month?"

Limits and privacy

• Max 40 MB / 50 MP per image; PNG, JPEG, WebP, AVIF.

• Uploaded images and results are deleted from Tinify.dev servers after 2 hours.

• The server only reads the files you point it at and only writes where it tells you it wrote.

Troubleshooting

• "TINIFY_API_KEY is not set" — the server exits at startup with the exact config snippet to fix it. Add the key to the env block of your MCP config.

• invalid_api_key — the key is wrong or revoked; create a new one at tinify.dev/developers.

• quota_exhausted (429) — the monthly quota is used up; run get_usage to see the period end.

• "path must be absolute" — pass full paths like /Users/you/img.png, not ./img.png.

• Logs go to stderr; in Claude Desktop see ~/Library/Logs/Claude/mcp-server-tinify.log.

Roadmap

• Batch tools (create/commit/wait/download) on top of the durable batch API — not in v1.

License

MIT © Stian Larsen

One-click / one-line installs

Cursor: Add to Cursor - then set your real key in Cursor's MCP settings.

VS Code:

code --add-mcp '{"name":"tinify","command":"npx","args":["-y","@tinify-dev/mcp"],"env":{"TINIFY_API_KEY":"tnf_live_..."}}'

Codex CLI:

codex mcp add tinify --env TINIFY_API_KEY=tnf_live_... -- npx -y @tinify-dev/mcp

or in ~/.codex/config.toml:

[mcp_servers.tinify]
command = "npx"
args = ["-y", "@tinify-dev/mcp"]

[mcp_servers.tinify.env]
TINIFY_API_KEY = "tnf_live_..."

Claude Code plugin (MCP + image-optimization skill):

/plugin marketplace add Stianlars1/tinify-claude-plugin
/plugin install tinify@tinify

More: https://tinify.dev/mcp