Skip to main content
A capability is an integration your worker can actually use — not just a connected account sitting in a list, but a tool it can call while it works. Connect Gmail, Slack, your CRM, or any of CogniAgent’s 2,700+ integrations, and your worker can send the email, post the message, or update the record itself.

What a capability is

Capabilities reuse the same Connection Manager and connected accounts that power the rest of CogniAgent — there’s no separate “worker integrations” system to learn. When you connect an app to a worker, the Connection Manager turns the actions that app supports (send an email, post a message, add a row) into callable tools, and the worker picks up those tools the next time it runs a task.
A capability is different from a skill. A skill teaches the worker how to do something; a capability gives it something to do it with. A worker often needs both — for example, a “customer follow-up” skill that relies on a Gmail capability to actually send the message.

Connecting an account

You attach capabilities from a worker’s Capabilities tab. The flow is one click for the common case, with more control when you need it.
1

Add an existing connection

Pick from your workspace’s already-connected apps and add it in one click. Every action that app supports is allowed by default — the card shows “All tools.”
2

Or connect a new app inline

Don’t have the app connected yet? Search the catalog, pick the app, and authorize it (OAuth or an API key) without leaving the worker page. It’s the same connect flow used everywhere else in CogniAgent.
3

Configure exactly which tools are allowed

Open “configure” on any attached connection to turn individual actions on or off, search a long action list, and optionally add a note on when the worker should use a given action.
Start with “all tools” for a trusted app and narrow it later. If you only ever want the worker reading Gmail, not sending from it, configure the connection down to the read-only actions.

Example capabilities

A worker can hold as many capabilities as its job needs. A few common ones:

Gmail

Send emails, reply to threads, and read messages from a connected inbox — the same actions available to email nodes elsewhere in CogniAgent.

Slack

Post updates to a channel, send a direct message, or react to a message as part of finishing a task.

CRM

Look up a contact, update a deal stage, or log an activity in Salesforce, HubSpot, or another connected CRM.

Google Sheets

Read a spreadsheet for source data, or append rows to log results as the worker completes a task.
Connections aren’t one-directional: the same Gmail or Slack connection that lets a worker send also lets it receive. See Channels for how a worker gets messages over these connections while a task is running.
Some actions leave CogniAgent and change something in the outside world — sending an email, posting a message. Depending on your permission settings, the worker may pause to confirm before it does. See Human-in-the-loop for how approvals work.

Search, browse, and interact with the web

Two related capabilities give a worker web access, from light research to hands-on browsing:
  • Web search & extract — a built-in toggle that lets the worker search the web and pull clean content from a page. Best for research and lookups.
  • Browser — a built-in capability that gives the worker a live cloud browser it can navigate, click through, and fill in forms on. Unlike search and extract, this is for pages that need to be acted on — logging in, submitting a form, paging through results.
The Browser capability streams a live view of what the worker is doing into the cowork session, so you can watch it work. When it hits something it can’t get past on its own — a login screen, a captcha, a two-factor prompt — it hands control to you rather than guessing. You act in that same live browser, then hand it back and the worker picks up where you left off. This uses the same pause-and-resume mechanism as Ask-a-Human — the worker’s turn suspends until you respond. A named browser remembers where it’s logged in. The first time a worker needs a site, you sign in through a handoff; after that, the worker reopens the same browser already authenticated, so you’re not re-entering credentials on every task.
Other ways to drive a browser or desktop — beyond the built-in provider — are on the roadmap. The tools a worker calls (open, act, hand off, close) are designed to stay the same as more providers are added underneath.

The always-on CogniAgent connection

One connection is different from the rest: every worker gets it automatically, with nothing to add. It’s what lets a worker:
  • Invoke and manage Applications and workflows — list them, kick one off, and check on a run, the same way a person would from the Applications area.
  • Start and manage Conversations — begin a Conversation Flow, check on one in progress, or read back what happened.
  • Ask a human a question — the Ask-a-Human tool, which pauses the task and asks you directly when the worker needs information or a judgment call only you can provide.
  • Delegate to another worker — the same reach covered in Multi-worker collaboration.
Because this connection is always attached, none of the above requires attaching a capability first — it’s simply part of what every worker can do.

Tools load on demand

A worker can accumulate a lot of capabilities over time — several integrations, each with its own set of actions. Loading every possible tool into every conversation would waste context and slow the worker down, so tools load as needed: a small toolset is available immediately, and a larger one stays out of the way as a searchable list until the worker actually needs one of those tools mid-task. Either way, only the actions you’ve enabled for a connection are ever available to call — configuring a connection down to a handful of actions is also what keeps the worker focused. See Context & memory for how this fits into what the worker keeps in view more broadly.

Moving files between integrations and the sandbox

Integration actions often deal in files — an emailed attachment, an exported spreadsheet, a document to upload. Those files need to reach the worker’s task files (the same files you see in the cowork session and that the sandbox works with), and files already in the task need to reach the integration when the worker calls it.
Automatic file exchange between integrations and task files is on the roadmap and not yet available. For how files work in a task today, see Execution, sandbox & files.

Next steps

Workers overview

See how capabilities fit alongside instructions and skills on a worker.

Skills

Teach a worker how to use its capabilities well.

Human-in-the-loop

Understand approvals, Ask-a-Human, and permission settings.

Multi-worker collaboration

Let workers delegate tasks to each other.