Skip to main content
A toolkit groups tools around a service or local resource. Use the provider-specific guide for credentials, configuration, supported operations, and cleanup. This page shows how to combine configured toolkits in an Agent and choose the appropriate integration.

Compose configured toolkits

Host factory: pass configured toolkit instances and a model. The factory constructs an Agent from their public tools. The host owns toolkit clients and closes them according to each toolkit’s lifecycle after all consumers finish; it also closes the returned Agent.
Start with one read operation, inspect its tool result, then add effects deliberately. getTools() exposes the toolkit’s collection; filter it before passing tools if your application needs a smaller surface. A prompt that says “read only” does not remove write tools or authorize access to a service record. For a complete lookup and approval example, use tool patterns. Import concrete toolkits from @agentium/core/toolkits or the documented individual subpath, not the core root.

Messaging and collaboration

Also see Gmail, WhatsApp, and Google Workspace. Service authorization and application record ownership remain host responsibilities.

Data, files, and execution

A filesystem base path and command restrictions are not OS isolation. Choose local, E2B, or Daytona execution according to the workload and ownership contract. Account for database pools, subprocesses, and remote sessions in shutdown.

Search, media, and judgments

“No API key” does not mean offline. Tools that search, scrape, or download still perform network I/O. Test live-service behavior separately from type-checking your configuration.

More patterns

Find every documented capability in the integration directory, add MCP tools for protocol servers, or compose selected tool capabilities in a research harness. Use approval and execution policy where effects need a host decision.