Before you start
You need Claude Desktop signed in to your account, and the HTTPS URL of your Kiteworks Compliance Runtime deployment (ending in /mcp). Kiteworks agents are read-only by default; mutating actions live in separate signed bundles and require explicit approval.
Add the remote MCP connector
- Open Settings → Connectors.
- Click Add custom connector (the + control).
- Enter your runtime URL, for example
https://your-host.example.com/mcp. - Complete the sign-in prompt. Claude Desktop stores the connection and reuses it for future sessions.
The connector UI writes the same configuration that claude mcp add produces on the command line, so a connection set up here also appears in Claude Code.
Run an agent
Start a chat and ask for one of the published agents. Every tool call is recorded into the hash-chained audit ledger. See the catalog for the full list of agents.
Security prerequisites (read first)
Claude Desktop connectors are initiated from Anthropic's cloud infrastructure, so the endpoint you publish is reached from the internet. Before you publish https://your-host.example.com/mcp:
- HTTPS only. Serve the endpoint over TLS — never plain HTTP.
- Authentication required. Every connection must authenticate (OAuth 2.0 / OIDC). There must be no unauthenticated tool access: an anonymous caller reaches zero tools and zero data.
- Tenant-scoped authorization. Each authenticated user sees only their own Kiteworks tenant's data; cross-tenant access is denied server-side.
- Runtime policy enforcement. The Kiteworks Compliance Runtime applies your ABAC, retention, and redaction policy to every tool call.
- Firewall allowlisting is in addition to auth, never instead of it. If you restrict inbound access to Anthropic's IP ranges, that does not replace OAuth and tenant checks — allowlisting must not bypass authentication.