Skip to main content

Tools

A tool is something the agent can actively do — search the web, read a calendar, query a database, control a device. Tools turn a model that just talks into one that can act.

The three kinds of tools​

Each tool in the dashboard has a coloured type tag:

TagTypeWhat it isWhere it runs
LocallocalA tool implemented in device firmware (volume, reminders, time, emotion detection, etc.).On the ESP device
CloudcloudA Knowledge Base surfaced to the LLM as a tool.Platform
Remoteremote_fetchAn MCP server you point the platform at.Where the MCP server lives

Local tools come from the device firmware — you enable or disable them per agent on the dashboard, but you don't define them here.

Cloud tools are managed in the Knowledge Base Integration section, not as standalone tool entries. They appear as read-only entries in the tool list.

The kind you'll spend most time configuring is Remote (MCP).

Two places, two jobs​

Remote (MCP) tools require setup in two separate places:

WhereWhat you do
Agent editor → Tool Integration sectionDeclare the tool on the agent: URL, type, auth method.
Tool Configuration page (/tool-config)Authorize the tool (if OAuth), fetch the tool list from the server, enable the tools you want.

You reach the Tool Configuration page from the agent's View more dropdown → Configure Tools (wrench icon).

warning

Declaring a tool in the editor does not make it usable. You must also complete setup on the Tool Configuration page before the agent can call it.

Discovering MCP servers​

Beside the Add Tool button in the Tool Integration section there is an unlabeled icon button (question mark). Click it to open the More Tools dialog, which links to:

  • Claude Documentation — remote MCP server list at docs.claude.com.
  • Remote MCP Servers — curated catalog at mcpservers.org/remote-mcp-servers.

Use this when you need to find an MCP server for a specific capability.

Adding a remote (MCP) tool​

Step 1 — declare it on the agent​

  1. Open the agent editor → scroll to Tool Integration.
  2. Click the + button:
    • Pre-defined Tool — pick from the catalog of available tool templates.
    • From Scratch — fill in a blank form.
  3. Fill in the fields:
    • Name — a label for this tool integration.
    • URL — the MCP server endpoint.
    • Auth — OAuth 2 or none. (API Key appears in the dropdown but is currently disabled — do not use it today.)
    • For OAuth: choose Automatic (Dynamic Client Registration / CIMD) if the server supports OAuth DCR, or Static (Pre-registered client) if you have a pre-registered OAuth client ID.
    • Timeout — optional, in seconds.
  4. Save the tool entry (✓ button on the row).
  5. Click Update Agent at the bottom.

The tool is declared but not yet usable until you complete Step 2.

Step 2 — configure it on the Tool Configuration page​

  1. Go back to Your AI Agents.
  2. On the agent's card → View more → Configure Tools (wrench icon).
  3. You are now on the Tool Configuration page.

For OAuth tools:

  • Click Configure on the tool entry.
  • For Automatic OAuth: you are redirected to the provider's consent screen. After approving, you are returned to the dashboard and the token is stored encrypted.
  • For Static OAuth: you may be asked to enter a Client ID first, then click Configure to start the redirect.

For no-auth tools:

  • Skip the authorization step and proceed to fetching.

Fetching the tool list:

  • Once authorized (or if no auth is required), click to fetch tools. The dashboard calls the MCP server's discovery endpoint and lists every tool it exposes.
  • Each discovered tool has an enable/disable toggle.
  • Enable only the tools the agent needs — fewer tools means the LLM picks more accurately.
  • Click Save.

Re-fetching when the server changes​

If the MCP server adds new tools later, click the Reload agent data (fetch latest KB list and tools) button at the top of the Tool Configuration page. New tools appear with their switches off by default — enable the ones you want and save.

OAuth in detail​

OAuth typeWhen to use
Automatic (Dynamic Client Registration / CIMD)The MCP server supports OAuth DCR — the platform creates the OAuth client automatically.
Static (Pre-registered client)You have manually registered an OAuth client with the server. Provide the Client ID (and optionally Client Secret).

After authorization, tokens are stored encrypted with KMS. Admin tokens authorize tool discovery; per-end-user tokens authorize runtime calls during conversations.

API Key (not currently enabled)​

The Auth dropdown lists API Key for forward compatibility, but it is currently disabled. Selecting it has no effect and there is no save flow for a key. Use OAuth (or no auth) for now. If your MCP server only supports an API key today, run it behind an OAuth-aware proxy or contact your platform admin.

How the agent decides which tool to call​

The LLM picks tools based on their name and description — both come from the MCP server's discovery response. If two tools sound similar, the agent may pick the wrong one. To improve accuracy:

  • Disable tools you don't need.
  • Add guidance in the System Prompt: "When the user asks about X, use the Y tool."

Tips​

  • Less is more. A focused tool list beats a kitchen-sink one.
  • Test with a fresh conversation. Newly added tools sometimes need a prompt nudge before the LLM uses them.
  • Disable, don't delete. If a tool is misbehaving, switch it off on the Tool Configuration page rather than removing it — easier to flip back on.
  • Use the More Tools button. Click the ? icon beside Add Tool to discover pre-built MCP servers.

On this page