Connecting an MCP server

The Skills & Knowledge → MCP Tools tab groups what's connected to your tenant into three sections:

• Integration Tools — MCP servers carried by a TrainAR integration you've connected (e.g. ServiceM8 ships its https://go.servicem8.com/mcp server)
• Platform Tools — MCP servers TrainAR has connected platform-wide for every tenant (e.g. FerryHopper, NYC Subway Info as demo/utility tools)
• Custom MCP Servers — MCP servers you've added directly via + Add Server

Three connection paths map to the three sections.

Path A — via a TrainAR integration

If the MCP server you want is exposed by a system TrainAR already integrates with, the connection happens automatically when you set up that integration.

Currently this includes:

Integration	MCP server URL	What gets added
ServiceM8	https://go.servicem8.com/mcp	Job, customer, quote, asset, form tools

Steps:

1. Settings → Integrations → connect the integration (OAuth flow via Nango).
2. The integration record's mcp_enabled flag is true for these connectors — the MCP server is auto-discovered.
3. Skills & Knowledge → MCP Tools tab — the server appears with its tools listed within a minute.

You don't enter the MCP URL or auth credentials manually — they're carried from the integration's OAuth.

Path B — direct MCP server connection

For MCP servers not yet a TrainAR integration:

1. Skills & Knowledge → MCP Tools tab → Add Server (under Custom MCP Servers).

2. The Add Custom MCP Server modal opens:

   

3. Enter:

   • Display Name — what shows in the tab
   • Server URL — the MCP server's endpoint (typically https://<host>/mcp or /sse)
   • Description — optional, for your own reference
   • Authentication — see below

4. Click Test Connection — TrainAR calls the server's tools/list to confirm reachability + discover the tool list.

5. Click Add Server to save.

Auth types supported:

Auth type	What you provide
none	No auth — public MCP server
bearer	A bearer token
oauth_via_nango	An OAuth client config (Nango handles the flow)

For test purposes there are public no-auth MCP servers you can connect to verify the flow — e.g. https://mcp.ferryhopper.com/mcp (ferry routes) or https://subwayinfo.nyc/mcp (NYC subway info).

Default response handling

When you connect a server you set default response handling for ALL its tools:

• default_response_target — context or display
• default_display_hint — fullscreen or hud (if display)
• default_response_scheduling — INTERRUPT or WHEN_IDLE

These defaults apply unless a specific tool has its own override in Managing MCP tools. See Skill response fields for the meaning of each value.

For most third-party MCP servers, sensible defaults are:

• default_response_target: context (the AI summarises in voice)
• default_response_scheduling: WHEN_IDLE (don't interrupt for background lookups)

You can always override per-tool later.

Removing a connection

MCP Tools tab → server card → ⋯ → Disconnect. Removes the tools from the AI's available set on the next session. Doesn't delete any data on the remote MCP server — just severs TrainAR's connection.

Common issues

Symptom	Likely cause	Fix
"Connection failed" on Test	Server URL wrong or server down	Hit the URL with curl -X POST <url> -d '{"jsonrpc":"2.0","method":"tools/list","id":1}' to confirm directly
Connection succeeds but no tools listed	Server requires auth you didn't provide	Check the server's docs; may need bearer or OAuth
Tools listed but AI never calls them	Tool descriptions too vague	See Managing MCP tools — override custom_description

Where to next

• Managing MCP tools — per-tool overrides
• MCP Tools overview — the concept
• Integrations overview — Path A integrations