Skip to navigation
Organize and automate

Agents

Define reusable AI behavior, choose tools, and control triggers.
View as Markdown

These operations require an enabled connection exposing the corresponding agents_ tools. They are not available in every MCP client connection.

Create an agent

  1. Use agents_listAgents to check for an existing agent.
  2. Discover available models with agents_listAvailableModels.
  3. Create the agent with agents_createAgent, supplying its title, model, and instructions using the current schema.
  4. Read it back with agents_getAgent.

Create a campaign research agent that summarizes supplied sources, distinguishes facts from suggestions, and does not publish social content.

Keep its purpose and expected output specific. Clone an existing agent when you want a reusable starting point.

Give the agent capabilities

Inspect agents_listToolkits and agents_listAvailableTools. Attach a toolkit using its returned name with agents_attachAgentToolkit. Use agents_updateAgentToolkit to adjust the enabled tools where supported.

Use agents_listAgentSkills and agents_addAgentSkill to manage the agent’s own instructions. These are separate from skills installed in your desktop assistant. Inspect whether an existing skill is editable before updating it; built-in instructions may be protected.

Configure a trigger

Discover supported trigger kinds with agents_listTriggerKinds, then create the selected kind with agents_createTrigger and its required configuration. Read the result before enabling a recurring workflow.

agents_runTrigger queues a run; it does not mean execution has finished. Inspect agents_listTriggerRuns for results. Pause and resume a trigger when you want to stop and restart its automatic activity without deleting the definition.

Verify the agent before automating it

Read the configured model, instructions, skills, and attached toolkits back through their respective tools. Check that its available tools match the intended task. Attaching a toolkit does not establish that its external accounts or integrations are authorized.

Use a narrowly scoped test input before relying on a recurring trigger. Inspect the resulting trigger run and its outputs; a successful trigger submission only means it was queued.

Maintain instructions and automation separately

Use agents_updateAgent to revise the agent definition and the skill tools to change its editable instructions. Use agents_getTrigger and agents_updateTrigger to inspect or change trigger configuration. Pausing a trigger stops that automatic entry point; it does not prove that every already-running execution has stopped.

Clone an agent with agents_cloneAgent when you want to preserve the original configuration. Inspect the clone before attaching new triggers. System-managed resources can have restrictions on editing or deletion; read the returned error rather than attempting to bypass them.

Add searchable context

A knowledge base stores indexed sources. Configure the agent to use the intended collections through supported configuration and tools, then test a question whose answer is present in those sources. Merely creating a collection does not attach it to an agent.

API reference

For request fields, response shapes, and direct HTTP integrations:

These endpoint pages describe REST requests. For MCP calls, use the current tool schema, including any additional convenience fields.