Skip to main content
Two directions, and they are independent:
  • Connectors are things an agent can call - a search API, a mailbox, an issue tracker.
  • Channels are surfaces an agent can be triggered from - a Slack channel, a Telegram bot, the in-app playground.

​
Connectors

Connect a service once in Integrations; its operations then appear as tools any agent can reference. Credentials are stored encrypted per workspace and are validated before they are saved, so a bad key fails at connect time rather than mid-run. Available today: Tavily, Exa, Gmail, Google Calendar, Slack, GitHub, Linear, Notion, Databricks, Salesforce, Apollo, LinkedIn, and Aurora (Solana portfolio). Gmail and Google Calendar connect over OAuth using your workspace’s own Google OAuth app, not a shared platform one - so consent, scopes, and revocation stay under your control. Several connectors (GitHub, Linear, Notion) are backed by the vendor’s MCP server, which means their tool surface tracks the vendor rather than being reimplemented here.

​
Channels

Inbound webhooks are verified before they run anything - Slack by signing secret, Telegram by a secret token echoed in the request header. An unverified call never reaches the agent. A channel run is an ordinary run: it appears in AI Monitoring with a full trace.

​
Subagents

An agent can be exposed to another agent as a tool. Delegation is then a normal tool call, so it shows up in the trace like any other step and the child’s run is metered on its own. Use it to keep one agent focused - a router that hands specialised work to specialists - rather than growing a single agent’s tool list until the model loses track of it.