Cleanor

Cleanor MCP: a zero-auth MCP server for image optimization and dev utilities

A hosted, zero-auth MCP server with 22 tools: optimize and convert images, generate QR codes, and run the deterministic dev utilities LLMs get wrong.

No API key. No signup. No OAuth. Point any Model Context Protocol client at one URL and 22 tools appear, all read-only and safe to hand to an autonomous agent.

Install

Pick your client. Every option below points at the same hosted server, so there is nothing to build, install or key.

Claude Code

claude mcp add --transport http cleanor https://mcp.cleanor.app/mcp

Then run claude mcp list to confirm it connected. Full walkthrough: docs/add-an-mcp-server-to-claude-code.md.

Cursor

Add to .cursor/mcp.json (per project) or ~/.cursor/mcp.json (global):

{

"mcpServers": { "cleanor": { "url": "https://mcp.cleanor.app/mcp" } } }

VS Code

Run code --add-mcp '{"name":"cleanor","type":"http","url":"https://mcp.cleanor.app/mcp"}', or add the server to .vscode/mcp.json in the workspace:

{

"servers": { "cleanor": { "type": "http", "url": "https://mcp.cleanor.app/mcp" } } }

Claude Desktop

Open Settings, then Connectors, then "Add custom connector", and paste the endpoint:

https://mcp.cleanor.app/mcp

Or run the server locally over stdio from claude_desktop_config.json:

{

"mcpServers": { "cleanor": { "command": "npx", "args": ["-y", "@cleanor/mcp"] } } }

The npm build is the same tool registry compiled for Node. It uses sharp (an optional dependency) for local image encoding; the other 21 tools need nothing.

Raw Streamable HTTP

The endpoint speaks the MCP Streamable HTTP transport, so you can drive it with curl or any HTTP client:

curl -s https://mcp.cleanor.app/mcp \

-H 'content-type: application/json' \ -H 'accept: application/json, text/event-stream' \ -d '{"jsonrpc":"2.0","id":1,"method":"initialize", "params":{"protocolVersion":"2025-06-18","capabilities":{}, "clientInfo":{"name":"curl","version":"1"}}}'

There is also a plain REST surface for non-MCP callers: GET /v1/capabilities (feature detection, machine-readable tool list) and POST /v1/optimize (one image in, optimized bytes out).

Tools

22 tools, every one of them carrying an input and an output schema, so a client gets typed structuredContent back and not just a wall of text.

Image

Cited data (Cleanor Labs research)

Dev utilities

The operations an LLM is unreliable at: hashing, counting, radix math, timezones. Pure compute, no network, no state.

Every tool is annotated readOnlyHint: true. Only optimize_image touches the network (openWorldHint: true); the other 21 are pure functions of their input.

Full input and output schemas: docs/mcp-tools-reference.md.

Tool index (CSV)

data/tools.csv is the same 22 tools as a flat, machine-readable table: name, title, category (image, data, dev), summary, required and optional parameters, and output fields. Handy for diffing releases, feeding an MCP directory, or comparing this server against another.

The file is generated, not hand-written: every value is read out of the tool's Zod schema in src/tools/. If you add or change a tool, regenerate it.

npm run tools:csv        # rewrite data/tools.csv from src/tools/

npm run tools:csv:check # exit 1 if data/tools.csv is out of date

Docs

FAQ

Is Cleanor MCP free?

Yes. It is free to use, with no account, no API key and no paid tier. The hosted endpoint is rate-limited per IP (120 requests per minute in general, 30 per minute for image optimization) so one client cannot starve the rest, and the source is MIT licensed if you would rather run it yourself.

Does it need an API key?

No. Auth is none, which you can verify for yourself at GET https://mcp.cleanor.app/v1/capabilities. There is no signup, no OAuth flow and no token to rotate. Point your client at https://mcp.cleanor.app/mcp and start calling tools.

What data does it send?

Only what you pass to a tool. 21 of the 22 tools are pure functions evaluated on the server, and optimize_image additionally fetches the public image URL you hand it. There is no account, no user identifier, and no database or object store in the server: it is a single Cloudflare Worker with no persistence binding for your inputs. If your data is sensitive, run the stdio build locally with npx -y @cleanor/mcp and nothing leaves your machine.

Can I self-host it?

Yes, in two ways. Run it locally over stdio with npx -y @cleanor/mcp (the same tool registry, with sharp doing the image encoding instead of Cloudflare Images), or deploy this repo to your own Cloudflare account with wrangler deploy, since the Worker entry point and wrangler.jsonc are both in the tree. The Dockerfile builds the stdio server for registries and scanners that prefer to start it themselves.

How is a hosted MCP server different from a local MCP server?

A local (stdio) server is a process your client spawns on your machine: it can reach your filesystem, but you have to install it, keep it updated and carry its dependencies. A hosted (Streamable HTTP) server is just a URL: no install, no version drift, and the heavy work runs on someone else's CPU. Cleanor MCP is offered both ways, and the hosted endpoint is the one to prefer for image optimization, because Cloudflare Images does the encoding and you never install sharp.

Which clients work with it?

Any MCP client that speaks Streamable HTTP: Claude Code, Claude Desktop (as a custom connector), Cursor, VS Code, and the agent frameworks with MCP support. Clients that only speak stdio can run the npm package instead. Both paths expose the identical 22 tools, because both register the same tool array from src/tools/.

Where does the research data come from?

storage_capacity and image_format_savings return numbers from Cleanor Labs' publish