MCP Tools overview
MCP Tools are the third class of tool available to the AI assistant on the glasses — alongside Skills and Knowledge.
MCP Tools are the third class of tool available to the AI assistant on the glasses — alongside Skills and Knowledge. Where a Skill is a webhook YOU register and Knowledge is documents YOU upload, an MCP Tool is a tool exposed by a third-party MCP server that you've connected to your tenant.
What MCP is
MCP (Model Context Protocol) is an open protocol for letting AI models call tools on remote servers. Lots of companies now ship MCP servers — your FSM might have one, your CRM, your job-scheduling platform.
When you connect an MCP server to your TrainAR tenant, the AI assistant on the glasses can call its tools the same way it calls your custom Skills. No code, no glue, just connect.
Where MCP Tools live
Dashboard → Skills & Knowledge → MCP Tools tab. Admin role required.
The tab shows:
- Connected MCP servers — each as a card with a count of available tools
- Tools within each server — listed with their name, description, and per-tool settings
Disconnecting a server immediately removes its tools from the AI's available set on the next session.
Where MCP servers come from
There are two paths a server appears in your tenant:
Via an integration — some TrainAR integrations ship with an MCP server included. The clearest example is ServiceM8 — connecting ServiceM8 in Settings → Integrations automatically connects its MCP server at
https://go.servicem8.com/mcp. The AI then gets ServiceM8's job/customer/quote tools alongside your own skills.Direct connection — for MCP servers not yet integrated as a first-class TrainAR integration, you can connect them directly by URL. See Connecting an MCP server.
How MCP Tools differ from Skills
| Skills | MCP Tools | |
|---|---|---|
| Who owns the tool definition | You (or a Marketplace bundle) | The MCP server's operator (3rd party) |
| How TrainAR finds them | Stored in your tenant's skills table |
Discovered live from the connected MCP server (cached for ~24h) |
| Auth | Auth header you set on each skill | Whatever the MCP server requires (often OAuth via Nango on the integration) |
| Customisable | Fully — every field is yours | Overridable (display name, description, response_target, etc.) but not the underlying tool itself |
| Versioning | You version | The MCP server operator does |
The AI doesn't care about the distinction at runtime — it picks whichever tool (Skill or MCP Tool) has the best matching description for the engineer's utterance. The categorisation matters for how YOU manage them.
How the AI calls an MCP Tool
When the AI picks an MCP Tool:
- It calls the tool through TrainAR's MCP client (we proxy the call to the remote MCP server).
- The server runs the tool against its own backend (your ServiceM8 account, etc.) using whatever auth was configured at connection time.
- The response comes back through TrainAR and is delivered to the AI per the tool's
response_target/display_hint/response_scheduling(same fields as Skills).
For latency reasons, TrainAR caches the tool list from each connected server for ~24h — so a server adding/removing tools mid-day won't reflect until the cache refreshes. The AI's tool calls themselves are always live, not cached.
Where to next
- Connecting an MCP server — wire one up
- Managing MCP tools — overrides, visibility, defaults
- What are skills? — the TrainAR-native equivalent