Skip to main content
Version: 3.6.0

Create and manage MCP catalog entries

The catalog stores reusable MCP server definitions. Each entry contains an image reference, runtime settings and input declarations. The image stays in its container registry.

This differs from connecting a server that you already run. See Connect an MCP server. To run a server from an entry, see Deploy MCP servers.

Access​

You need permission to manage MCP infrastructure on the platform. The same permission manages MCP registries and private networks. Permission to manage external MCP servers is separate and does not give access to the catalog.

Under Manage platform, open MCP Servers, then select the Catalog tab.

In a tenant, members open MCP Servers, then Catalog. They see only the entries shared with their tenant, and they cannot change them. Editors and Owners can deploy an entry from its page. An entry that an administrator is deleting is no longer listed, and opening it reports that it is not found.

Open an entry​

Select an entry's name in the list. The entry page shows what the entry declares. It does not show a running server.

  • Container image: The image the entry uses.
  • Runtime: The HTTP port the server listens on.
  • Deployment settings: The settings the entry accepts. Each row shows the name, the type, whether the setting is required, the environment variable that carries it, and the declared default.
  • Deployment secrets: The secrets the entry needs. Each row shows the name, whether the secret is required, and how it reaches the server: as an environment variable or as a file. The page shows which secrets are needed, never their values.

A value that the entry does not declare shows as Not configured.

On the platform, the catalog list and the entry page also count the entry's deployments in all tenants and the platform: all of them, the running ones, and the ones that need new input values.

Image and registry​

Each entry names one image, pinned by digest, such as registry.example.com/acme/postgres-mcp@sha256:<64 hexadecimal characters>. The image must start with its registry host. For a Docker Hub image, start with docker.io/.

  • Public image: Keep Registry set to None (public image).
  • Private image: Select the registry that holds the pull credentials. The image host must match the registry's authority. The platform treats docker.io, index.docker.io and registry-1.docker.io as the same host.

See Create and manage MCP registries to create or manage registries.

Create an entry​

Select Create entry. Fields marked * are required.

  • Display name: The name shown to users. Description is optional.
  • Container image: Select a registry, or keep None (public image), and enter the image.
  • Runtime: Enter the required HTTP port and HTTP path, for example /mcp.
  • Deployment settings: Optionally declare non-secret string, integer or boolean settings with environment-variable mappings. Declared default says whether the setting has a default value. Keep No default when it has none. Choose Default value to enter the value the platform uses when nobody gives one at deployment time. For a string setting, an empty box declares an empty text value, which is not the same as No default. For an integer setting, enter a number, such as 0.
  • Deployment secrets: Optionally declare sensitive values supplied when deploying an MCP. Choose an environment variable or a file in the platform's input directory. Do not enter secret values in a catalog definition.

If you choose File in the input directory, enter a single file name, such as database-url. Do not enter a folder path. When an MCP server is deployed from the entry, the platform writes the value to /var/run/cm-mcp/inputs/ followed by that file name. The MCP server must read the value from that file.

Within an added declaration, complete the marked fields. Required at deployment controls whether users must supply that input when deploying an MCP.

Advanced settings starts collapsed. Open it to configure these optional sections:

  • Container startup arguments: Command-line options passed when the MCP container starts. Enter each argument separately.
  • Preset environment variables: Non-secret values applied to each deployment. Users cannot change them in the deployment form.

Host name checks​

The platform reaches a deployed MCP server through an internal host name. Each deployment gets its own host name, so a catalog entry cannot name it in advance. Some MCP servers accept requests only for host names on an allow list, as protection against DNS rebinding. Such a server refuses the platform's requests with HTTP 403, and its tools do not appear.

Add container startup arguments that allow every host name or turn the check off. For example, the Grafana MCP server needs --allowed-hosts and * as two arguments. A host-name suffix pattern does not work.

This does not expose the server. Only the platform's model gateway can call a deployed MCP server.

Container requirements​

The platform runs the container without root, with all Linux capabilities dropped and no Kubernetes API token. The image must set a numeric non-root user, such as USER 1000. The platform cannot check a user given by name, so a container with a named user or with root does not start. An image that needs root access does not work either, for example one that writes to directories owned by root or listens on a port below 1024.

On OpenShift, the cluster gives the container a user from the namespace's range instead, as a member of group 0. Group 0 must be able to read the files the server reads and write the folders it writes.

The server must serve MCP over HTTP on the port and path that the entry declares under Runtime. The platform treats the server as running when it accepts connections on that port.

Edit an entry​

Open the entry and select Edit. Saving replaces the display name, description and definition. The tenant access does not change.

The form opens with the entry's current values. If the entry uses a registry that the registry list does not show, Registry shows that registry's identifier followed by (current registry). If the registry list cannot be loaded, Registry is locked and the entry keeps the registry it already uses. You can still change the other fields and save.

If another administrator changes the entry before you save, the portal refuses the conflicting save and keeps your changes in the form. Select Reload to load their version, then apply your changes again. Reloading discards the changes in the form.

If you leave the page while the form has unsaved changes, the portal asks you to save them, discard them, or keep editing.

Update the deployments​

A saved definition applies to every deployment created from the entry, in all tenants and the platform. A change to only the display name or the description does not change the deployments.

If a definition change affects deployments, the portal shows what it does before it saves:

  • How many deployments move to the new definition. Running deployments restart.
  • How many deployments need new input values, and which inputs. A new required input without a default has no value in an existing deployment. Those deployments keep running the definition they have and show Needs input, until an administrator configures them. See Configure a deployment.
  • How many tenants own the affected deployments.

Select Save and update to save, or Cancel to keep editing without saving. After the save, the entry page shows what happened. A deployment that could not be updated shows Outdated, and the next save of the entry tries again.

Share an entry with tenants​

The Access section of a platform catalog entry controls which tenants can see the entry.

  • Turn on Share with all tenants to share the entry with every tenant.
  • Otherwise, select tenants in Shared with tenants.

Select Save access. The saved access replaces the previous access.

Choosing tenants needs permission to manage tenants. Without it, the tenant list is not shown, and a save keeps the tenants that are already shared.

Removing access does not stop or change existing deployments.

Delete an entry​

Deleting an entry also deletes every deployment created from it, in all tenants and the platform, whatever its state.

  1. Open the entry and select Delete. The dialog shows how many deployments the entry has in total, and how many of them are running.
  2. Type the entry's name to confirm, then select Delete.

The dialog stays open while the platform removes the deployments. This can take a few minutes. When the entry is gone, the portal returns to the catalog list. If you leave the page, the delete continues.

While the delete runs, the entry shows Deleting. It can no longer be edited, deployed or shared. If the delete fails, the entry stays Deleting. Open the entry and select Delete again to finish it.