Skip to content

Custom Tools & MCP

Connect your agents to your own systems. A Custom Tool is a webhook the agent can call mid-conversation (look up an order, create a ticket); an MCP Server exposes a whole catalog of external tools at once.

Where: left nav → Tools (/tools) and MCP Servers (/mcp-servers).

Open Tools → New Tool (or Import to bulk-import from an OpenAPI spec or a cURL command). Configure:

  • Tool name — a lowercase function name the agent calls (e.g. create_ticket).
  • Description — tells the agent when to use the tool (≥ 10 characters).
  • Endpoint — method (GET/POST/…) and an HTTPS URL. You can use :id path tokens and insert {{secrets.X}}, {{call.X}}, {{metadata.X}} template variables.
  • Parameters — the values the agent fills in (name, type, required, description, constraints). Map each to the path, query or body. Up to 30 parameters.
  • Headers — static headers (values can reference secrets; encrypted, never re-shown).
  • Fixed values — values the platform always sends but the agent never sees (constants or templated data).
  • Authentication — None, Bearer, API key, Basic, OAuth2 client credentials, or JWT bearer.
  • Response Mapping — extract fields from the response into variables (dot-path, e.g. response.order.status), or leave empty to pass the whole response to the agent.
  • Advanced — timeout (1–120 s), retries (0–3, exponential backoff), Silent mode (hide the response from the agent), and Speak while running (play a short spoken filler while the request is in flight).

On an existing tool (with no unsaved changes), open Test to fire a real request with sample values and inspect the Request / Response / Extracted tabs — confirm it works before enabling it on an agent or in a flow. Admins can also see an execution History tab of recent real invocations.

Enable the tool on an agent (Tools tab → Custom Tools) or add a Tool node in a flow.

Instead of defining tools one by one, connect an MCP (Model Context Protocol) server and the agent gains its whole tool catalog.

  1. MCP Servers → New MCP Server.
  2. Enter a name, the server URL (HTTPS), and pick the transport — Streamable HTTP (URLs ending /mcp) or SSE (/sse).
  3. Add auth headers if needed.
  4. Optionally list allowed tools to expose only a subset (empty = all).
  5. Save, then enable the server on an agent or flow.
  • Secrets belong in Secrets; reference them as {{secrets.KEY}} in tool URLs, headers, fixed values and auth.
  • Duplicating a tool never copies secret values — re-enter credentials on the copy.
  • Disable an MCP server (instead of deleting it) to temporarily pull it from agents without losing its config.