Cloud/Connectors/AI Capabilities

AI Capabilities

Ask AI about Vinkius

The switchboard of your connector: every tool it exposes, one by one, with its HTTP method and path, a master switch for the whole set and the power to silence any single capability before your AI ever sees it.

A connector is a bundle of AI Capabilities: each tool it exposes is a capability your agents can call. The AI Capabilities tab is the switchboard for that bundle, and its subtitle states the job: "Review and toggle which capabilities are available to connected clients." This was the third move in the Dashboard's playbook, and it is where least privilege stops being a slogan and becomes a list of switches.

R

AI Capabilities

Review and toggle which capabilities are available to connected clients.

5/6 ENABLED
POSTcreate_issue/repos/{owner}/{repo}/issues
GETlist_issues/repos/{owner}/{repo}/issues
GETget_file_contents/repos/{owner}/{repo}/contents/{path}
POSTcreate_pr_comment/repos/{owner}/{repo}/issues/{n}/comments
GETlist_pull_requests/repos/{owner}/{repo}/pulls
GETsearch_code/search/code
AI Capabilities, live. The master switch with the enabled counter and every tool with its colored method badge and path: flip create_pr_comment off and the counter drops to 5/6.

The mockup is the real table: flip any tool off and watch the counter drop, then use the master switch to silence or restore the whole set at once.

The master switch and the counter

The header holds the fleet move: a master On switch beside the counter 5/6 ENABLED. Flip it one way and every tool goes dark in a single action (fired per tag behind the scenes); flip it back and the full set returns. When you want almost everything except one dangerous capability, turn the master on and switch off the exception.

One row per tool

Each row is one capability with three facts on display:

  • the On switch, your per tool kill switch: a tool switched off disappears from the connector's capability set, so connected clients never see it and never call it;
  • the Method badge, color coded like the console: GET in emerald (reading is safe), POST in sky (writes are actions), amber and red reserved for the heavier verbs;
  • the Name and the Path, both in monospace, so create_issue points at /repos/{owner}/{repo}/issues and there is zero mystery about what a tool touches.

A switched off row dims to 40% and stays in the list: silenced, not deleted. The empty state is the console's own: "No tools available for this connector."

Why this screen exists

Every capability you leave on is a capability your AI will use eventually, because agents use what they can see. The power move is subtraction: if your agents should read GitHub but never comment on pull requests, create_pr_comment goes off and the risk goes with it, while the rest of the connector keeps working. Least privilege, one switch at a time.

The perimeter, now complete

With this tab the connector's identity block is complete: Dashboard to watch it, Credentials to identify it, Connection Tokens to hand out keys and AI Capabilities to define exactly what those keys can do. The remaining tabs cover the build side: Playground to try tools, Deployments to ship changes, Audit Logs and Log Streaming for the record, Settings and Monetize for the business.