Give your agents access to knowledge bases, integration actions, and custom tools.
Out of the box, an actor can have a conversation — but it can’t do much. Capabilities are how you give each actor superpowers: search your knowledge, send a Slack message, run a refund calculation, kick off a workflow.There are three capability families:
Knowledge bases
Search your indexed documents — product docs, policy PDFs, FAQs.
Integrations
Call actions on connected apps — Gmail, Slack, Salesforce, Shopify, 2,700+ more.
Custom tools
Run a Python snippet, hit an HTTP endpoint, or trigger a CogniAgent workflow.
Each capability becomes a tool the actor can call during its turn. The LLM decides when to use it.
Pick a knowledge base from your workspace’s Knowledge Base library. The actor gets a search_knowledge_base tool it can call whenever it needs to look something up.Fields when you add a KB:
Field
What it does
Knowledge Base
Pick from the dropdown.
Description
One-line note about when this actor should search it. The LLM uses it to decide.
Active
Toggle to enable / disable without removing.
Apply Reranking
More accurate but slightly slower searches. Default on.
Add multiple knowledge bases per actor when they serve different purposes. The LLM sees them as separate tools (e.g. search_product_docs and search_policy_documents) and picks the right one.
Integrations give an actor the ability to act on connected apps.
Three-step flow:
1
Pick an app
Browse the catalog (2,700+ apps). Search by name — Gmail, Slack, Salesforce, etc.
Browse or search the integrations catalog. Each tile shows how many actions the app exposes.
Narrow the catalog by typing a name — e.g. 'gmail'.
2
Pick a connection
Choose an existing connection or set up a new one. Connections live in the Integrations page of your workspace.
3
Pick which actions to expose
Each app has many actions. You choose which ones this actor can call.For example, Gmail has send-email, list-messages, search-threads, etc. If this actor only ever needs to send emails, give it only send-email. Narrow scope = safer behaviour.
Toggle individual actions on or off — only checked actions are exposed to the actor as tools.
Each chosen action becomes a named tool the actor can call (e.g. send-email). The LLM picks the tool’s arguments — to, subject, body — based on the actor’s instructions and the conversation context.
Instructions:“After collecting the user’s name, email, and request, send them a confirmation email summarizing what we’ve captured. Use the send-email tool.”
When the user has provided their details, the actor calls send-email with to=<user's email>, subject="We received your request", and a body containing the summary. The email lands in the user’s inbox; the actor confirms in chat.
The LLM provides the args based on the conversation; the snippet returns structured data back.Best for: calculations, formatting, data transformations.
Execute Code form — name, description, parameters, and a Python editor. The LLM fills args; you write what to return via output_data(...).
Kicks off another CogniAgent workflow by id. Optionally waits for it to finish and returns the result.Best for: deterministic multi-step processes you’ve already built as workflows — refund processing, lead scoring, report generation.
Execute Workflow form — pick the deployed workflow and map the LLM's arguments to its inputs.
Make an HTTP call to an internal API or third-party endpoint. Configure method, URL, headers, body schema; parse the response.Best for: lightweight integrations to APIs you control, or one-off endpoints that don’t justify a full integration.
HTTP Request form — method, URL, headers, parameters, and an optional response parser.
Independent of what you configure, every actor automatically has the routing tools its focus mode allows: done (signal completion), hand off to a sibling, escalate to global, ask user to clarify. You don’t see these in the capabilities drawer — they’re managed by the platform.
Don’t expose every action from an app. If you give an actor “all of Gmail”, the LLM will call delete-thread when “delete this conversation” comes up in chat. Narrow scope to the actions the actor genuinely needs.
Don’t put behaviour into tool names. Tool names should describe what they do, not when to use them — “when to use” goes in the actor’s instructions. The LLM reads instructions, then picks tools.
