Skip to main content
A connector gives your agents access to one of your company’s systems: your own HTTP API or an MCP server. Each thing the agent can do there is an action, and each action has its policy: whether it runs on its own, the customer confirms it or the team approves it. The URL and the key stay on the app’s side: the agent never sees them.

​
Creating a connection

In Connectors, tap New connection:
  • Name: for example, Orders system.
  • MCP server URL (Streamable HTTP, over HTTPS) or API base URL (HTTPS and public hosts only; actions add their path to this URL).
  • Authentication: No authentication, Token (Bearer), Username and password, Key in a header or OAuth2 (client credentials). Secrets are stored encrypted and never shown again.

​
Actions

Tap Read tools: each of the server’s tools shows up as an action, turned off. Review it, pick its policy and enable it.If the server changes, tap Read tools again. New ones are added; changed ones are marked Changed on the server until you review them; the ones that are gone are turned off.
On each action:
  • For the agent: its Name (lowercase, numbers and _, like check_stock) and When to use it: what it does, when it helps and what it returns. Without a description, the agent won’t know when to use it.
  • When it runs: the policy (see below).
  • Data the agent fills in: a JSON Schema (on MCP, defined by the server).
  • Response: the Fields it reads (one path per line, like order.status or items[].price), the Maximum items and the Maximum wait (s). Less is better: faster, cheaper and no extra data.
  • Test: runs the action with test data and shows what the agent would read. If the action changes your system, the test really does it.

​
Policies

An action that changes your system with the Automatic policy runs without anyone confirming it. Use this only if it’s harmless.

​
Giving actions to an agent

An enabled action isn’t used by anyone yet: on the agent, Tools → Actions in your systems, turn on the ones it may use. See Agents.