Connect and verify an MCP server

Connect an external tool#

MCP servers expose tools to a conversation. This guide covers a local filesystem server and the public Microsoft Learn documentation server, using Onevium's actual configuration fields.

Open the right configuration surface#

Open Plugins → MCP from the left sidebar. Choose the intended project in the scope selector, then Add → Add manually to open the configuration form. Choose Personal · all projects when the server should be configured at user scope. The selected scope determines where a new manual entry is saved.

Add → Add with AI, or @MCP in a conversation, asks the assistant to help with configuration. The AI entry prepares a conversation draft; add the service details and send it before setup starts.

The older right-side MCP Servers → Add Server shortcut still creates a global/user entry. For project scope, use the unified page above or the .mcp.json example below.

For the local example, install Node/npm and prepare a folder containing hello.txt with the text MCP demo ready. Use an absolute folder path and keep unrelated files outside it.

Add the local filesystem example#

Fill the form as follows:

-y
@modelcontextprotocol/server-filesystem
/absolute/path/to/mcp-demo

Replace the last line with your prepared directory, for example C:/mcp-demo on Windows. Click Add Server, confirm the entry is enabled, then open a new test conversation. The package and allowed-directory behavior are documented by the official filesystem server.

Add a remote HTTP server#

Microsoft Learn's official MCP server uses Streamable HTTP and requires no authentication. Add a second entry:

Save it and ask: “Use microsoft-learn to find Microsoft's documentation for creating a .NET console app; return the official source link.” A browser GET to this endpoint may return 405; test it through MCP instead.

Form and JSON fields#

For example, the remote entry's JSON tab can contain:

{
  "type": "http",
  "url": "https://learn.microsoft.com/api/mcp"
}

Keep a server in one project#

The simplest path is Plugins → MCP, choose the project, then Add → Add manually. After saving, keep that project selected and confirm the entry appears there.

To maintain the same configuration as a project file, create or merge the entry into .mcp.json at the project root. Preserve other servers already in the file:

{
  "mcpServers": {
    "microsoft-learn": {
      "type": "http",
      "url": "https://learn.microsoft.com/api/mcp"
    }
  }
}

Reopen the project's conversation and check that the entry is shown under that project. Project JSON needs the outer mcpServers map; the Add Server dialog's JSON tab does not. Keep credentials out of committed example files.

Verify the tools, not just the entry#

For docs-files, ask the assistant to list its actual allowed directories and read hello.txt. Expect the prepared directory and MCP demo ready. For Microsoft Learn, inspect the returned documentation link and tool source.

The filesystem server also has write tools. Its allowed roots can change when a client supplies Roots; verify the returned directories rather than assuming the launch argument is the final boundary. The server row provides enable/disable, Edit, Restart, and Delete actions.

Choose which conversations load which tools#

Open Plugins → MCP → Tool loading and context cost. These controls determine the tools loaded into conversation requests. The Personal / Project selection used when adding a server determines where its configuration is stored; these are separate choices.

1. Choose All sessions, This project, or This session first. Project and session buttons are unavailable without the corresponding context.
2. Choose This session for a rule limited to the current task, or This project for project-wide behavior. When an override exists, Use inherited settings restores the parent rule.
3. Adjust the server's Load tools switch if needed. A server disabled in its original configuration must be enabled there first; runtime rules cannot override that disablement.

Changes apply to new conversation requests after the current run ends. Running tools are not interrupted. After changing a rule, send the next request and check the actual tools.

Inspect and select only the tools you need

For this guide's docs-files server:

1. Click Inspect tools. This connects to the configured service and reads its tool list without calling its file-reading or writing tools.
2. Click Choose tools, search by name or description, and keep the directory query and text-reading tools needed for this exercise.
3. Selection changes save automatically in the chosen scope. Use all tools restores all available tools; Use inherited settings removes the current scope's override.
4. In the next request, read hello.txt and confirm that the response still matches the practice file.

Enlarge imageReal 1.2.2 MCP controls with an isolated docs-files example and a simulated tool list. No external server was connected and no real files were read or written. Actual tool names and counts depend on the service.

Load tools on demand when supported does not behave identically on every connection. Native OpenAI and Gemini connections still use upfront loading; selecting fewer tools can reduce their tool-definition overhead. Schema token counts estimate the size of tool definitions. They are not billable usage or guaranteed savings.

When to export a CLI and skill

For a service exposing many tools, consider Export CLI and skill if you want the assistant to use command-line calls. Check the selected scope first, then read the returned skill location and export result.

Export validates the CLI, stores required credentials in a private local file, and disables MCP tools in the selected scope to avoid duplicate loading. It retains the original server configuration. This does more than download instructions: do not share or commit files containing credentials. To restore MCP use, re-enable the tools or restore inherited settings in the same scope, then verify with the next request.

Troubleshoot a server#

Next steps#

Use plugins for packaged integrations, permissions for action controls, and settings reference for network and tool settings.