Other MCP clients

The Connect page covers claude.ai and Claude Code specifically, and OpenClaw and Hermes each get their own deeper, harness-specific guide with a distributable skill baked in. This page is the generic version underneath all of those: point Cursor, or anything else that speaks MCP, at the same endpoint with the same key, and it gets the same 5 tools and the same check_action discipline.

Point it at the endpoint

Most MCP-capable clients take a remote server as a URL plus headers, either in a config file directly or a settings UI that writes one for you. The shape below is what a remote HTTP server entry looks like; if your client's config uses different key names, the three fields it's asking for are always the same: a name for the server, this URL, and the bearer header.

Code
{
  "mcpServers": {
    "gnt-brain": {
      "url": "https://api.gntai.dev/mcp/",
      "headers": {
        "Authorization": "Bearer gnt_live_xxxx"
      }
    }
  }
}

Get a key

gnt keys create mints an org-scoped MCP key and prints the endpoint alongside it. The key is shown once, so paste both straight into the config above.

Code
$ gnt keys create

MCP URL: https://api.gntai.dev/mcp/
Key:     gnt_live_xxxxxxxxxxxxxxxx

# Shown once -- copy it now.

Teach it the check_action discipline

Connecting the endpoint only makes the 5 tools available. Nothing makes a harness actually call check_action before it acts. Paste this into the client's system prompt or custom instructions field.

Code
Before any action that sends a message, moves money, deletes data, or is
otherwise hard to undo, first call the check_action tool with a plain-
English description of what you are about to do.

- verdict "allowed": proceed.
- verdict "blocked": do not proceed. Tell the user why, citing the rule.
- verdict "needs_human": stop and ask a human to approve before acting.

Never treat a missing or unclear verdict as permission to act.

For company-specific context beyond a single action -- answering a
question about policy, or working from a standing set of rules instead
of one check -- call search_rules with a plain-English query, or pull a
compiled skill pack with list_skill_packs / get_skill_pack. Only rules a
human has approved are ever returned.

If your client supports a hook layer, something that runs before a tool call regardless of what the model decides, the way a PreToolUse hook does, pair this prompt with that hook pattern on the Enforce page for defense in depth: the prompt guides the model, the hook enforces regardless.