The Loadout MCP server: ask your AI about MCP servers
Everything on this site about MCP servers (which ones exist, what each can do, which tools can change things, the exact config for Claude Code, Claude Desktop and Cursor) is also available as an MCP server you can add to your own AI assistant. Then you can just ask it. It is free, needs no account, and only reads.
Add it to Claude Code: run this once.
claude mcp add --transport http loadout https://loadout.pages.dev/mcp
We ran exactly this in a fresh folder. claude mcp list showed ✔ Connected, and Claude used the server's tools in a real session. Other apps that can connect to a remote MCP server by URL should work with the same address, but we have only tested Claude Code.
What you can ask
| Tool | What it does |
|---|---|
list_mcp_servers | The 14 servers we have tested, with what each does, whether it needs an API key, and how many of its tools can change things. |
get_server_tools | One server's full tool list, as the server itself reported it, with each tool labelled read-only or able to change things, plus what to watch out for. |
get_server_config | Copy-paste config for a server in Claude Code, Claude Desktop or Cursor. API keys are referenced by environment variable, never written in. |
check_command | Whether the Safe setup guard would block a shell command, and why. Nothing is run. |
check_file_path | Whether the guard would block an AI from editing a file, such as .env or a lock file. |
Some things we asked it in a real session, and what came back:
- We had Claude check
git -C . push origin main. The tool answered "blocked", with the reason the guard gives Claude. - We asked how many of the Memory server's tools can change things. Claude looked it up and answered 6 of 9, which matches the server's own declarations.
Ideas: "Which of these servers need an API key?", "Give me the Cursor config for Playwright", "Which Notion tools can delete things?", "Would this command be blocked?"
What it does and doesn't do
- Read-only. All five tools are marked read-only, and none of them runs, reads or writes anything on your machine. The checker tools only analyse the text you give them.
- Nothing you send is stored. There are no accounts. We keep one anonymous daily total of how many tool calls were made, with no content, in the same counter table the website uses.
- Inputs are limited to 10,000 characters, which is where Claude Code itself stops analysing a command. Longer input is refused. We also timed the checks on a range of hostile inputs (long runs of quotes, backticks, nested brackets and so on) and every one finished in well under half a second.
- Not a guarantee. "Not blocked" from
check_commandmeans no rule matches, not that the command is safe. A text check only knows its patterns. See why a hook can't be a sandbox. - Data has dates. Tool lists were collected by starting each server, and each carries the date it was collected. We have no tool list for the GitHub server (it needs Docker, which we could not run), so it returns toolsets instead.
How it was tested
- 268 automated tests: every server's config in every app matches the config builder exactly, the guard gives the same answers as the pack for all 195 shared cases, bad and oversized input is rejected, and the adversarial inputs stay fast.
- The official MCP Inspector, first against a local copy and then against this live address.
- Raw protocol requests for bad input, an unknown tool, and a 2 MB body. All were refused cleanly.
- A real Claude Code session, as described above.
It is built with the Cloudflare Agents SDK, as a stateless server, so there is no session to keep and nothing to log in to. To check any MCP server yourself, see testing a server with the Inspector.