The LenderSkills connector

At a glance

The three tools

Searching and reading are read-only and need no credential — a catalogue nobody can browse gets recommended by nobody. Installing is the one that writes, and it is marked that way in the protocol so your client asks you first rather than firing it on its own.

Setup

Replace YOUR_KEY with your install key from lenderskills.ai/install — a work email and a six-digit code, about thirty seconds, no password and no card. You can also connect without a key and add one later; search and read will work in the meantime, and an install will tell you how to get one.

claude mcp add --transport http lenderskills https://lenderskills.ai/api/mcp --header "Authorization: Bearer YOUR_KEY"

[mcp_servers.lenderskills]
url = "https://lenderskills.ai/api/mcp"
http_headers = { "Authorization" = "Bearer YOUR_KEY" }

Claude, ChatGPT, Cursor, Windsurf and everything else

Most clients read a JSON config. Some connector interfaces take only a URL and no header — paste the endpoint on its own and search and read will work.

{
  "mcpServers": {
    "lenderskills": {
      "type": "http",
      "url": "https://lenderskills.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_KEY"
      }
    }
  }
}

Try it with these

Four prompts that exercise the tools in the order a real session uses them.

1. “What LenderSkills skills are there for reviewing a construction draw request?”

   search_skills — Searches the catalogue and reads back the matches with who wrote each one. Nothing is written and no key is needed.

2. “Tell me what the payoff demand builder actually does before I install anything.”

   get_skill — Pulls the full listing and the opening of the skill file so you can judge it. Still read-only.

3. “Install the payoff demand builder into this project.”

   install_skill — Writes .claude/skills/<slug>/SKILL.md, then walks you through the [BRACKETED] placeholders — your lender name, your state, your fee schedule. The author is emailed that you took it.

4. “Find me an underwriting skill for DSCR deals, show me the best one, and set it up.”

   all three — The normal shape of a session: search, read, then install the one you picked.

What you are installing

A skill is one Markdown file: instructions, worked examples, and [BRACKETED] placeholders where your own numbers go. No code, no binaries, no scripts — the screening pipeline rejects them outright. Every submission clears five gates before it publishes: archive inspection, an executable block, static analysis, then two independent model reads, one adversarial and one editorial. Every skill page shows its own screening record. How screening works →

Skills are written by working lenders and are not legal, tax or lending advice. Every package carries that disclaimer. Read one before you run it, the way you would skim a contract before signing.

Keys, limits and support

• Your key installs skills as you and does nothing else. It cannot read your leads, change your byline, or touch your account. Treat it like a password anyway.
• Lost or leaked? Email mila@helveticagroup.com and we will rotate it.
• Rate limits. Installs are capped per account per hour. Search and read are capped by address. Going over returns a plain message, not a failure.
• Repeat installs of the same skill inside fifteen minutes count once, so a retry does not send the author a second introduction.
• Something wrong with a published skill? Every skill page has a report link, open to anyone.