TrainAR Docs
Try Free
Skills & Knowledge · Last updated 18 May 2026 · 3 min read

Managing MCP tools

Once an MCP server is connected, you can customise each individual tool — change how the AI sees it, override its response handling, or hide it entirely.

Once an MCP server is connected, you can customise each individual tool — change how the AI sees it, override its response handling, or hide it entirely. Customisations live in the mcp_tool_overrides table; they apply only in your tenant and never affect the remote MCP server.

Where overrides live

Skills & Knowledge → MCP Tools tab → server card → tool row. Click any tool to open the override panel.

What you can override

For each tool you can change:

Field Default (no override) Override use
Visible to AI Yes Hide a tool the AI shouldn't call (e.g. destructive operations)
Custom display name The server's name Rename for your engineers ("Add Job" vs "createJobV2")
Custom description The server's description Sharpen the AI's matching signal
Custom instructions (none) Guardrails — "Always confirm before calling", "Never call inside a customer's home"
response_target The server's default_response_target Per-tool override
display_hint The server's default Per-tool override
response_scheduling The server's default Per-tool override

These overrides are stored per-tool, per-tenant. Disconnecting + reconnecting the server clears them.

Common patterns

Hiding destructive tools

Many MCP servers expose tools that delete/cancel/refund. If you don't want the AI on the glasses calling those, toggle Visible to AI off. The tool stays connected (you can re-enable any time) but the AI never sees it in its tool list.

Sharpening descriptions

Server-supplied descriptions are often generic — "Create a new job in ServiceM8". Override to a when-to-call statement specific to your team's vocabulary:

Call this when the engineer mentions creating a new job, logging a callout, scheduling a return visit, or booking follow-up work. Triggered by phrases like "create a job for…", "book this in", "schedule a return", "log a new ticket".

The AI's matching is much better with that than with the server's default.

Visual rendering for diagram-returning tools

If a tool returns a URL to a diagram or document, the default (context) just hands the URL to the AI as text — engineer hears "Here's a link to the wiring diagram, but I can't show it." Override to display + fullscreen so the URL renders on the glasses instead.

Custom instructions for safety-critical tools

For tools that affect customer billing or written records, add instructions:

IMPORTANT: Always read back the proposed action to the engineer and wait for verbal confirmation ("yes, proceed") before calling this tool. Never auto-call.

The AI treats custom_instructions as a hard-coded prefix when considering whether to call the tool.

How overrides resolve at runtime

When the AI is choosing tools:

  1. The MCP server's tool list is pulled (cached ~24h).
  2. For each tool, your overrides are layered on top:
    • If is_visible: false → tool is removed from the candidate set.
    • Display name, description, instructions, response handling all use override values if present, else server defaults.
  3. The AI sees the override-resolved tool definitions and matches against them.

There's no "merge" — an override wholly replaces the server's value for that field. So if you override custom_description, the server's description isn't appended.

Auditing

The audit log records every override change (see Audit log). Useful for tracking who hid which tools and when, especially when you have multiple admins managing the same connection.

Where to next