Skip to main content
Open source Agent presets save a model, instructions, tools, skills, and MCP integrations under a slug. Open Agents from the workspace navigation to create one, then run it from a workflow with ai.preset_agent. Agents page Each row shows the preset’s tool sources as icons, its MCP integrations after a divider, the number of tools, the model, and the last update. The count includes the tools of remote MCP integrations. Hover an icon to see its namespace or MCP server.

​
Configure a preset

Click Create new, name the preset, and write instructions in the editor. Use the tabs in the side panel for the rest of the configuration:
  • Tools: Set Model, Thinking, Internet access, and Retries, then choose the tools the agent can call.
  • Subagents: Attach other presets the agent can delegate to.
  • Skills: Attach published skills.
  • Output: Choose the format of the agent’s final answer.
Agent preset tools The Tools list groups tools by namespace. Each group shows how many of its tools the preset uses, such as 5 of 35, and expands to list them. MCP integrations appear in the same list with an MCP badge, and tools from attached skills appear under From skills. Click Publish version in the top right to save a change in any tab.

​
Add tools

In the Tools tab, click Add tools and check the tools the agent needs. An MCP server is one checkbox for the whole server. The dialog counts the selected tools against the limit, such as 16 of 128 tools. Registry actions from the preset and its skills count toward the limit. MCP tools do not. Tracecat rejects a save over the limit with a message such as This agent has 160 tools; the limit is 128. Each tool has an Auto or Ask toggle. See Human-in-the-loop to require approval before a tool runs.

​
Add subagents

A preset can delegate work to other presets in the workspace.
  1. In the Subagents tab, click Add subagent and choose a preset.
  2. Expand the row and set:
    • Alias: The name the parent agent uses for the subagent. It defaults to the preset slug and must be unique on the parent.
    • Delegate when: When the parent should hand work to this subagent.
    • Max turns: The most turns the subagent can take. Leave it empty for no limit.
  3. Click Publish version.
Each row shows the subagent’s name, model, and alias. Click the name to open that preset. A preset with its own subagents cannot be attached. A preset with tools set to Ask can be attached only when an enabled agent backend supports subagent approvals. The picker flags a preset that cannot be attached. Hover the row to see the reason.

​
Set the output format

In the Output tab, choose a Format:
  • Text: The agent returns plain text. This is the default.
  • Structured: The agent returns one typed value. Choose a Type of String, Integer, Float, or Boolean, and turn on List of values to return a list.
  • JSON schema: The agent returns JSON that matches the schema you enter.
Keep Text unless a downstream step needs structured output. To get structured output from a single workflow step without changing the preset, set output_type on the ai.preset_agent action. It replaces the preset’s output type for that run.

​
Where tools come from

Three sources on the preset side make up what the agent can call:
  • Registry actions on the preset, added in the Tools tab. The usual home for actions only this agent needs.
  • MCP integrations attached to the preset, also in the Tools tab. An MCP integration reaches a preset directly and does not need to go through a skill.
  • A skill’s metadata.tools, which takes registry actions and MCP references alike. Tools declared here travel with the skill’s instructions, so a group several agents reuse stays in one place.
The effective tool set is the union of the three. A namespaces filter set on the preset through the API narrows the registry actions in that set; MCP tools are not filtered by it. The Tools tab shows how many listed tools the filter blocks, with Clear to remove it. A workflow run can swap one piece out. The actions input on the ai.preset_agent action replaces the preset’s registry actions for that run instead of adding to them, so registry actions a skill contributed are dropped for the run. MCP tools stay either way, and an empty list is ignored. It is handy for running a preset with a different tool set in a test or an eval.

​
Availability

Open source includes presets, tool_approvals and the agent inbox for reviewing tool calls before they run, skills, secrets and variables in preset instructions, and the MCP catalog connectors. Enterprise Edition Enterprise adds:
  • See AI agent for the ai.preset_agent inputs and examples.
  • See Human-in-the-loop to require approval before a tool runs.
  • See Skills to attach reusable instructions and files to a preset.
  • See MCP servers to connect servers a preset can call.