# Add and manage tool servers

Tools can come from installed apps or connected tool servers. Open **Tools** to inspect what is available and whether an action asks for approval.

## Add an MCP server
Choose **Add tool server → MCP server**. Give it a unique name, then select Command or Remote URL.

For a command server, enter the executable and arguments. Hosted commands start inside Gideon’s private workspace, not on your laptop. Put non-secret settings in Environment and save credentials in Connections using the names the server expects.

For a remote server, supply a reachable endpoint. Choose Streamable HTTP unless the server specifically requires SSE. If authorization is needed, enter the **credential name**, not its secret value, and the required prefix such as Bearer. Save that secret under MCP in Connections.

## Add a tool API server
Choose **OpenAI tool server** for the supported tool-API protocol. Enter its endpoint, a readable name, and a credential name when needed. Use Tool filter to expose only named tools; leaving it empty exposes all tools returned by that server. This connects tools, not a customer-selected inference model.

## Check discovery and permissions
Re-probe MCP servers or reconnect the selected server after setup. Inspect the discovered tools and run a small read-only task. Tools needed by platform features cannot be disabled through the ordinary toggle. Approval indicators tell you when a tool asks before running; they are not proof that a request will succeed.

## Remove a server
Inspect workflows, agents, and schedules that depend on it, then use Remove server. Revoke external credentials separately where needed. The [inbound editor MCP interface](/docs/self-hosting/remote-access-and-ide) is a different, self-hosted read-only capability.