Connectors (MCP)

On the Connectors (MCP) screen, an administrator connects RelayHub to an external tool server. After the connection, the tools of the server are available to the AI in chat, to your agents and to agent chains, next to the RelayHub tools. This is the opposite direction of AI Assistant Connections. There, an outside assistant connects to RelayHub. Here, RelayHub connects to an outside tool server. Both directions use the same open standard, the Model Context Protocol (MCP). You can connect each server that uses MCP.

​

When to use a connector

Add a connector when a third party gives an MCP tool server that you want the RelayHub AI to use. Examples are the search service of a vendor, a data provider or your own internal tool server. After you connect it, ask in chat, and the AI calls the external tool for you.

​

Add a connector

​

Use the tools

When a connector shows Connected, its tools are available everywhere the AI runs: in chat, in agents and in agent chains. You do not select the tools yourself. Ask for what you want, for example “look up the AI visibility for acme.com”. The AI finds the tool that it needs and calls it, as it decides when to search your files or run a skill. If a connector is not ready, the AI tells you why. For example, the server needs a new sign-in or shows an error. In chat, an administrator gets a Reconnect button. Another person is told to ask an administrator. Skills can also use these tools. Expand the row of the connector to see the tool names. In Skill Studio, add a tool name to the Platform tools of the skill. The skill can then call the tool during its run. The call has the same organization isolation and the same audit record as a call from chat.

​

Images from a connector

When a tool of a connector sends back an image, for example a chart or a screenshot, RelayHub saves the image as a file of your conversation. When the model can see images, it looks at up to 4 images of one result before it answers. When the AI wants to show you an image, it puts a link to the file in its answer. Click the link to download the image.

• RelayHub keeps at most 8 images of one result, each up to 20 MB. It does not keep SVG images, because they can run scripts.
• When the model cannot see images, it can look at a saved image with the image analysis tool.
• When a tool sends something that RelayHub does not keep, for example audio, the AI is told so. It does not act as if it had the part.
• A long text result is kept whole, as for every tool (see Chat Basics).

​

Tool approval

Each tool in an expanded row has an Approval setting:

• Ask (server default) or Server says read-only: the server decides if the tool asks before it runs.
• Always ask: RelayHub asks you each time before the tool runs.
• Run without asking: the tool runs and does not ask.

​

Manage a connector

Each row has these controls:

• Connect or Refresh: connect, or read the tool list of the server again. RelayHub also reads the tool list of each connected server again every 24 hours, and when a call fails because a tool changed. A tool that the server adds later asks before it runs, until an administrator changes its Approval setting.
• Edit connector: change the name, the address or the authentication. To keep the stored API key, leave the key field empty. If you change the authentication method, RelayHub deletes the stored credential and asks you to connect again.
• Disable or Enable: turn a connector off and keep it. When it is off, the AI cannot use its tools.
• Remove: delete the connector and its stored credentials.

The status of a row is Connected, Not connected, Disabled or Error. A row with an error shows the cause. If the server refuses the sign-in, the row shows Not connected until an administrator connects it again. If the tool list cannot be read three times in a row, the row shows Error.

​

Instructions for a connector

To tell the AI how to use a connector, open Settings > Connections and write in Connector instructions. An administrator can write Company instructions, which every person gets before their own. For more information, see Settings. When the server changes its tools. If the AI calls a tool that the server removed or renamed, or the server refuses the arguments because the tool changed, RelayHub reads the tool list of the server again at once. The AI then gets the current tools or the new arguments and can try again. RelayHub reads the list at most once in 5 minutes for each server. A tool that is new in the list asks before it runs (Always ask) until an administrator changes its Approval, unless the server says that the tool only reads. A tool that an administrator turned off is never used, and the AI tells you that it cannot use it.

​

Security and privacy

• RelayHub encrypts the credentials. RelayHub never shows an API key or an OAuth token again and does not write it to the logs.
• Each connector is private to your organization. One organization cannot see or use the connectors, credentials or tools of a different organization.
• RelayHub accepts only safe addresses. RelayHub refuses a server address on an internal or private network.
• RelayHub records each external tool call. Administrators can see the activity of connectors in the Audit Log.
• A new address deletes the key. If you change the server address to a different host, RelayHub deletes the stored credential. The old key never goes to the new address. You then connect again.

​

Limits

• The address must start with https. RelayHub refuses http addresses and internal hosts.
• Each organization has a maximum number of connectors, and RelayHub reads a maximum number of tools from each server. These limits keep the AI fast.
• If your administrator set a list of permitted hosts, you can connect only to those hosts.
• RelayHub keeps at most 8 images from one tool result.