docs: Fix incorrect navigation paths and misleading instructions in "MCP servers in the AI Catalog"
Summary
The documentation page MCP servers in the AI Catalog contains two categories of inaccuracies that mislead users trying to set up and manage MCP servers.
Problem 1: Wrong menu path — "Build" should not be referenced
Multiple sections instruct users to navigate via Build > AI Catalog, including:
- "Add an MCP server to the AI Catalog"
- "Edit an MCP server"
- "Connect an MCP server to a custom agent"
- "View MCP servers connected to a custom agent"
- "Disconnect an MCP server from custom agents"
The AI Catalog does not live under the Build menu. The navigation path in the docs is incorrect and users cannot find the AI Catalog by following these instructions.
Problem 2: Misleading instructions for connecting/disconnecting MCP servers
The sections "Connect an MCP server to a custom agent" and "Disconnect an MCP server from custom agents" describe actions as if they can be performed from the AI Catalog. However, the actual OAuth connection/authentication step (which is what makes an MCP server truly "connected" and usable at runtime) must be done from a completely different location: AI > MCP servers at the group or project level.
This is already documented separately in the "Authenticate with an MCP server" section, but the connect/disconnect sections do not reference it or clarify the distinction, leaving users confused about why their agent cannot use an MCP server even after associating it in the AI Catalog.
Proposed fixes
- Correct all navigation paths that reference Build > AI Catalog to use the accurate menu location.
- In the "Connect an MCP server to a custom agent" section, add a clear note that associating an MCP server with an agent in the AI Catalog is only one step — users must also authenticate via AI > MCP servers for OAuth-enabled servers before the agent can use them at runtime. Cross-link to the "Authenticate with an MCP server" section.
- Clarify the "Disconnect an MCP server from custom agents" section to reflect where this action actually takes place.
Impact
Users following the current docs are unable to find the AI Catalog and, even when they do find it, do not understand why their agent cannot call Linear or Atlassian tools after associating the MCP server. This is a significant source of confusion for the MCP feature which is already experimental.
References
- Doc page: https://docs.gitlab.com/user/gitlab_duo/model_context_protocol/ai_catalog_mcp_servers/
- Related feedback issue: #593219
- Related docs improvement issue: #595938