Create a capability

This walks through creating a new capability that calls an external API. The capability starts as a Draft, and is ready to be published once you’re happy with it. There is no sandbox — a test run goes to the real endpoint you configured.

Before you start

• A user role that can edit Capabilities (typically Admin)
• A connection to the system you want to call. If you have none, add one first from Settings → Capabilities — Add connection offers Parse from documentation (paste a docs URL, or paste the docs text), Set up manually (base URL, endpoints by hand), or MCP connection. Credentials are set on the connection, not on the capability; see Authentication options.
• A sample request that you know works — Postman, curl, or your engineering team can supply one

Steps

1. Open Settings → Capabilities.
2. In the Capabilities card header, open the ⋯ menu (More capability actions) and choose New capability (manual). The editor opens on a blank capability. (Two other doors sit in the same header: Add capabilities turns several endpoints of one connection into capabilities in a single pass, and Guided setup walks connection → action → tier → publish as one flow.)
3. Under Name your capability, fill in Capability name and Description. The name is what specialists see when you assign capabilities to them, so make it task-shaped: Look up order, Cancel subscription, Track delivery.
4. Under Choose an action, select which API connection and which endpoint this capability calls.
5. Under Customer interaction, work through each parameter the endpoint declares and say whether the customer would know it and what they call it. (The section only appears once you have picked an endpoint that declares parameters.)
6. Under Response filter, choose which response fields your Agent Stacks are allowed to see. Fields you leave out never reach the AI.
7. Under Action tier, choose how far this capability may go — Public, Read, Act or Transact. Atender proposes one; you confirm it. The proof each tier requires is built in and not configurable, and the enforcement level is derived from your choice and from the connection — see security levels.
8. Under Guidance for the Agent Stack, answer When should the Agent Stack use this? — the situations that should trigger a call. Publishing requires non-empty guidance.
9. Click Save in the editor header. You’re not live yet — Drafts aren’t callable from any specialist.

Verify it worked

You should see your new capability in the list at Settings → Capabilities with a Draft badge and its tier in the Tier column. Open it and confirm:

• The parameters and the response filter match what you intended
• The Before it goes live checklist has no outstanding rows, and the When this goes live card describes the behaviour you want
• The description clearly explains when the capability should be used

The capability isn’t usable yet — Drafts can’t be assigned to specialists. Move it through testing and publish it next.

Troubleshooting

• Symptom: Parsing the documentation doesn’t pick up any endpoints. Fix: Documentation parsing lives on the connection, not here. Reopen Add connection → Parse from documentation. If the URL yields nothing — most modern doc sites render with JavaScript — copy the endpoint documentation out of your browser and use Paste documentation instead. Auto-discover from OpenAPI only searches a base URL for a spec; it has no file upload.

• Symptom: Authentication fails when you try a test call. Fix: Fix it on the connection, not on the capability. Confirm the detected auth type matches what the API expects (Bearer vs API Key is the most common mix-up) and that the credential value doesn’t have whitespace or a wrapping "Bearer " prefix already in it. A key sent in a custom header or a URL parameter can’t be saved from Connect your account — finish it in Advanced settings.

• Symptom: The AI calls the capability with the wrong values. Fix: Revisit Customer interaction — for each parameter, say plainly whether the customer would know it and what they call it — and sharpen When should the Agent Stack use this?. Both are what the AI reads when it decides what to pass.

See also

• Test and publish a capability
• Node types reference
• Authentication options
• Order lookup recipe