Skills

A skill is saved know-how your agents reuse: write a task down once and they follow it every time. Some skills are instructions, some call a tool in a connected system, and some are built in. Agents only use the skills you grant them.

Where Skills live

Open the user menu at the bottom of the sidebar and choose Skills under Manage. The menu entry shows for admins and owners. The Skills page shows the catalog with tabs All, Custom, Connectors and Built-in, grouped as:

Instructions: skills you write (Write a skill).
Tools: skills that call an external system (Connect a tool).
Connectors: skill packs that come with a connected app.
Built-in: platform skills such as Search the web, Generate a file, Query structured data, Look up a record, Publish to site, Market data, and Email escalation.

Click New skill to add one. Each skill has an Enabled or Disabled pill for the whole organization. Turning a skill on means "Bound agents can use it on their next answer."; turning it off means "Bound agents lose it immediately."

Connector skills need a live connection

Connector skill packs exist for Epicor, SAP, Business Central, HubSpot, Salesforce, Dynamics 365, Zendesk, Jira, Intercom, ActiveCampaign, ClickUp, Google Calendar and Outlook. Calendar packs connect per teammate.

Until the app is connected in Sources, its pack stays locked. The Skills page tells you: "Connect [app names] in Sources to unlock their skills." If nothing is connected yet you'll see "No connector skills yet" with a Browse connectors button. Each pack shows Connected, Per-teammate or Not connected.

In the agent builder, connector skills whose app is not connected are simply not shown. If a connection is removed later, agents stop seeing those skills immediately.

Granting skills to an agent

Open the agent and go to the Skills step ("What it may do beyond answering.").
Skills are grouped as Platform, Custom and Connectors. Each connector group shows "N of M on" with Turn on all and Turn off all.
Toggle the skills this agent may use. Every skill shows a risk label: Read-only, Drafts, Takes action, or Admin. Skills your role cannot attach show a restricted badge.
On a saved agent, toggles save instantly without waiting for Set live. On a new agent, platform skills that need no connector are pre-selected and attach when the agent is created.

For each skill on a saved agent you can also set:

Available to: Team & customers, Team only, or Customers only.
Customers must be: Anyone, Verified by email, or Signed in.

If a skill that takes real action is open to anyone on a customer channel, the builder warns you and suggests requiring a verified or signed-in customer. Some skills are always team only and never run on customer channels.

Web search is not a skill toggle. It is the Public web search switch on the Knowledge step.

Using a skill from chat

In the chat composer, click the + button ("Add files & capabilities").
Under "[agent name] can also", pick a skill. The first five are listed; Show N more reveals the rest. The last item is Add files or photos.
If the message box is empty, it fills with "Use [skill name] to" so you can finish the sentence and send.

Only skills the current agent has, and whose connection is live, appear in the menu.

Approvals before actions

Read-only skills run on their own. Skills that write to an external system or need admin rights always pause for approval:

A card appears in the chat with the status Approval needed and two buttons, Approve and Decline.
A dock above the composer reads Needs approval while something is waiting.
After you decide, the card updates to Action completed, Action failed, Action declined, or Action outcome unconfirmed. Approvals that nobody acts on expire after 24 hours and show Approval expired.

Action-taking skills are not available in private chats, because an approval needs a saved conversation.

Test runs

Open a tool skill on the Skills page to find its Test button (disabled while the skill is off). It opens a dialog titled Test "[skill name]" with a Parameters (JSON) field and a Run test button. The dialog is explicit: it invokes the skill against the configured external system using stored credentials, and side effects on the upstream API are real. The result shows Success or Failed, the HTTP status, the time in milliseconds, and the response headers and body.

Testing requires an admin role, and each test run is billed like a standalone skill run.