Custom MCP Server

Connect your own self-hosted Model Context Protocol (MCP) server over HTTP.

How to connect

Go to Integrations → click Custom MCP Server.
Enter your MCP endpoint (must end with /mcp).
(Optional) Add HTTP headers, for example Authorization: Bearer ....
After connecting, Worken fetches a fresh capability snapshot and shows discovered tools, prompts, and listed resources for that server.

HTTP transport only. Local stdio and CLI transports are handled via the worken mcp developer CLI.
Bot runtime remains tools-only in v1. Prompts and resources are shown in the integration view, but are not exposed to the model runtime.
Interactive workflows that require elicitation are out of scope for v1.
URL must be reachable from Worken and use https:// in production namespaces.
Tools receive provider-safe IDs like mcp_custom__<integrationSlug>__<toolSlug>__<hash>. The format keeps a readable slug from the original name while staying within provider limits and avoiding collisions between servers.

URLs are validated against SSRF: localhost, private IP ranges, and link-local addresses (169.254.x.x) are rejected.
Redirects are blocked at the HTTP transport level.
Embedded URL credentials are rejected; secrets must be passed through headers.
Headers (for example API tokens) are never logged and are stored encrypted or in the external secret store instead of plaintext auth_data.

You have an MCP server that exposes many tools and want to plug it into multiple Worken agents.
You need to experiment with internal MCP tools without publishing them as first-class vendor integrations.

For CLI-based development or local MCP servers that are not exposed over HTTPS, use the worken mcp CLI instead of this integration.