MCP Registries in Catalog

An MCP Registry is a centralized publishing and discovery endpoint for MCP servers within your organization.

As organizations experiment with AI agents, MCP servers are often embedded directly into agent code, stored in local configuration files, or scattered across repositories. Over time, this can lead to MCP sprawl, making it difficult to understand:

• What MCP servers exist
• Which agents should use which servers
• How those servers should be governed

MCP Registries extend Konnect Catalog to provide a structured, standards-based way to register and manage MCP servers, helping Platform Teams maintain visibility and control as AI adoption scales.

This feature is built on top of the MCP Registry API specification defined by Anthropic’s open source project. The specification allows flexibility in how registries are structured, so you can model them according to your organization’s governance needs.

MCP Registries in Catalog are currently available in tech preview via Konnect Labs.

1. In the Konnect sidebar, navigate to Organizations.
2. In the Organization sidebar, click Labs.
3. Click Catalog - MCP Registry.
4. Click the toggle at the top right to enable the integration.

You can access the MCP Registry by navigating to Catalog > MCP Registries in the sidebar.

Send a POST request to the /registries endpoint:

curl -X POST "https://klabs.us.api.konghq.com/v0/registries" \
  -H "Authorization: Bearer $KONNECT_TOKEN" \
  --json '{
    "name": "internal-mcp-registry",
    "display_name": "Internal MCP Registry",
    "description": "Registry for MCP servers approved for internal AI agents"
  }'

An MCP server represents an agent-facing service definition. It describes:

• What the server does
• Its version
• How agents can connect to it

To add an MCP server, send a POST request to the /registries/{registryIdentifier}/v0.1/publish endpoint:

curl -X POST "https://klabs.us.api.konghq.com/v0/registries/internal-mcp-registry/v0.1/publish" \
  -H "Authorization: Bearer $KONNECT_TOKEN" \
  --json '{
    "name": "com.example/expense-reimbursement",
    "description": "MCP server that allows agents to submit expense reports and check reimbursement status.",
    "version": "1.0.0",
    "packages": [
      {
        "registryType": "npm",
        "identifier": "@example/expense-reimbursement-mcp",
        "version": "1.0.0",
        "transport": {
          "type": "stdio"
        }
      }
    ],
    "remotes": [
      {
        "type": "streamable-http",
        "url": "https://mcp.internal.example.com/expense"
      }
    ]
  }'

MCP servers are defined independently from a single deployment. They may include multiple delivery paths.

Packages

Packages describe run-it-yourself distribution options, such as:

• npm packages
• PyPI packages
• OCI artifacts

These allow teams to install and run the MCP server within their own agent environments.

Remotes

Remotes describe hosted MCP server endpoints that agents can connect to over the network, such as:

• streamable-http
• sse

An MCP server can define multiple packages and multiple remotes simultaneously. This allows organizations to support different runtime environments without duplicating server definitions.

Registry endpoints assume a Konnect authentication context. Only authenticated clients with appropriate access tokens can query the registry URL:

curl -X GET "https://klabs.us.api.konghq.com/v0/registries/internal-mcp-registry/v0.1/servers" \
  -H "Authorization: Bearer $KONNECT_TOKEN"

Future enhancements will allow publishing MCP servers to Dev Portal, enabling controlled exposure to broader audiences, including users and agents outside of Konnect.

We are continuing to evolve MCP Registries alongside the broader MCP ecosystem.

Planned enhancements include:

• Publishing MCP servers to Dev Portal
• Auto-registration of MCP servers created via Konnect-native workflows such as AI Gateway
• Additional governance and lifecycle controls

Because the MCP specification is still evolving, we are committed to iterating in partnership with customers who have already begun developing MCP servers and experimenting with agent-based workflows.