keyhand-mcp

Stdio MCP bridge for KeyHand — the credential dropbox for AI-powered agencies.

KeyHand stores your client API keys, admin passwords, and service-account JSONs in an encrypted vault, then lets your AI agent (Claude Code, Cursor, Claude Desktop) use them in HTTP calls without ever seeing the plaintext. This package is the stdio bridge for clients that don't yet speak the HTTPS MCP transport natively.

Install

You don't need to install — just point your MCP client at it via npx.

Use with Claude Desktop

Edit your Claude Desktop config:

{
  "mcpServers": {
    "keyhand": {
      "command": "npx",
      "args": ["-y", "keyhand-mcp"],
      "env": {
        "KEYHAND_TOKEN": "mcp_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Generate the token at https://keyhand.com/dashboard/settings/tokens.

Use with Claude Code

Claude Code can talk to KeyHand's HTTPS endpoint directly — you don't need this package. Run:

claude mcp add keyhand https://keyhand.com/api/mcp \
  --header "Authorization: Bearer mcp_xxxxxxxxxxxxxxxxxxxxxxxx"

You can still use this package via Claude Code's stdio MCP support if you prefer:

KEYHAND_TOKEN=mcp_xxx claude mcp add-stdio keyhand npx -y keyhand-mcp

What you get

Two tools the agent can call:

• list_credentials({ project_id }) — returns the names + schema metadata of every credential in a project. Never returns plaintext values.
• inject_credential({ project_id, credential_name, request: { url, method, headers, body } }) — KeyHand server-side resolves the credential, substitutes the literal string $secret in your headers / body, makes the HTTP request, and returns the response to the agent. The plaintext never enters the agent's context.

Self-hosting

Set KEYHAND_URL to your self-hosted endpoint:

{
  "env": {
    "KEYHAND_URL": "https://your-keyhand.internal/api/mcp",
    "KEYHAND_TOKEN": "mcp_xxx"
  }
}

How it works

This package opens a JSON-RPC channel on stdin/stdout, then forwards every message to the configured KeyHand HTTPS endpoint with your token as Authorization: Bearer. The upstream JSON-RPC response is written back unchanged.

That's the entire implementation. ~60 lines of Node. No dependencies beyond the built-in fetch.

Security

• Your token is read from KEYHAND_TOKEN and only ever sent to the configured KEYHAND_URL.
• All transport is TLS 1.3.
• This package never persists anything to disk.
• KeyHand's MCP endpoint logs every request to your workspace's audit log.

Full threat model: https://keyhand.com/security.

License

MIT. See LICENSE.

Links

• Site: https://keyhand.com
• Docs: https://keyhand.com/help
• Source: https://github.com/PCSAdmin081/keyhand-mcp (will go live with v0.1.0)
• Security: security@keyhand.com