MCP Servers
===========

MCP (Model Context Protocol) is an open standard for connecting AI models to external tools and data sources. MCP servers expose one or more tools over the protocol, and any AI platform that supports MCP can discover and call those tools in a standardised way. OneByZero Neo supports both platform-managed MCP servers (deployed and operated by the platform) and user-managed servers (deployed by your organisation).


MCP Server Cards
-----------------

Each MCP server appears as a card on the **MCPs** tab of the Integrations page. Cards display:

* **Name** — the identifier for the server.
* **Description** — what capabilities the server provides.
* **Status** — deployment status of the server (e.g. ``Deployed``).
* **Deployment source** — whether the server is system-managed (deployed by the platform) or user-managed (deployed by your organisation).
* **Enabled / Disabled** toggle — controls whether the server's tools are available for agents to use.
* **Tags** — organisational tags.
* **Last updated** timestamp.

Clicking a card opens the detail view, which shows:

* Full description.
* For **user-managed** servers: the Docker image URL and port used for deployment.
* For **system-managed** servers: configuration is managed by the platform.

Adding an MCP Server
---------------------

MCP servers are added to the platform by administrators with appropriate permissions. The process depends on the deployment source:

**System-managed servers** are provisioned through the platform's MCP server management interface (typically under admin settings). Once provisioned, they appear in the MCPs tab for workspace users who have the `canViewMcpServers` permission.

**User-managed servers** require you to provide:

* **Docker image URL with tag** — the fully qualified image reference for your MCP server container.
* **Port** — the port the MCP server listens on.
* **Environment variables** — any configuration the server requires at startup (API keys, endpoints, etc.).

Once deployed and enabled, user-managed servers appear alongside system-managed servers.

Enabling MCP Tools for Agents
------------------------------

Once an MCP server is enabled in the Foundry, its tools are available when building Agentic Flow agents. Add an MCP tool node to the flow canvas and select the desired MCP server and specific tool from its exposed tool list.

In NeoGPT, the **Integrations** toggle panel lists MCP tools alongside other tools attached to the agent's flow. Individual tools can be toggled on or off per conversation thread. See the :doc:`/11_command_center/neogpt` documentation for details.

Best Practices
--------------

* Disable MCP servers that are not actively in use. A disabled server's tools are not available to agents, which reduces the LLM's tool selection scope and can improve response quality.
* For user-managed servers, use specific image tags rather than ``latest`` in the Docker image URL. This prevents unexpected behaviour when a new image version is pushed.
* Test MCP servers after deployment by running a simple agent flow that calls one of their tools before exposing them to production agents.