Skip to main content
Open source Skills let you package reusable agent instructions and supporting files. Attach published skill versions to a preset agent when the agent needs the same playbook, reference material, or helper files across runs.

​
Create a skill

Open Skills from the workspace navigation. Click Create new, then choose Skill to start with an empty working copy or Upload to import an existing skill directory. Skills list Skill names use lowercase letters, numbers, and single hyphens. Tracecat seeds new skills with a root SKILL.md file.

​
Write SKILL.md

Use SKILL.md for the skill metadata and instructions. The root file must include frontmatter with name and description.
Tracecat uses the frontmatter name as the published skill name. Keep published skill names unique among skills attached to the same preset agent. Write the body as instructions the agent can follow during a run. Skill editor

​
Declare tools in frontmatter

A skill can carry its own tools under metadata.tools in the SKILL.md frontmatter.
Each entry is a registry action name (core.http_request) or an MCP reference (mcp.<slug> for the whole server, mcp.<slug>.<tool> for one tool). Per-tool references work for HTTP integrations; a stdio integration is granted whole with mcp.<slug>. A skill can declare at most 64 tools. When the skill is attached to a preset agent, its tools are added to the preset’s own Tools selection. The agent’s effective tool set is the union of the preset’s registry actions, the MCP integrations attached to the preset, and the tools declared by every attached skill.
If the preset restricts namespaces, that filter also applies to the registry actions a skill declares. A skill’s registry action outside the allowed namespaces is dropped from the run and reported under tool_policy.blocked_tools. MCP tools are not filtered by namespaces, whether a skill declares them or the preset attaches the integration directly.
Skill tools are granted when the preset configuration is resolved, whether or not the agent ever opens the skill. Only the skill instructions and files load on demand.

​
Add files

Add supporting files when the agent needs more than instructions. Use files for examples, checklists, schemas, helper scripts, or reference data. Editable text files open inline in the skill editor. Other file types can be replaced, deleted, or downloaded.

​
Publish a version

Save the working copy, then click Publish version. Publishing creates an immutable version from the current files and makes the skill available to preset agents. Published skill version Publishing requires:
  • A root SKILL.md file.
  • Valid YAML frontmatter in SKILL.md.
  • A frontmatter name that follows the skill name rules.
  • No duplicate paths or path conflicts.

​
Attach skills to a preset agent

Open the preset agent, then use the Skills tab to attach published skills. An attachment records the skill version that was current when you attached it. Attach skills to a preset agent Only published skills can be attached. When a preset agent runs, Tracecat resolves each attached skill to that skill’s current published version, so publishing a new version takes effect on the next run without editing the preset.

​
How agents load skills

When a preset agent runs, Tracecat stages the resolved skill versions into the agent runtime. The agent can discover the skill by its SKILL.md metadata and use the bundled files during the run. Use skills for reusable agent behavior. Use preset instructions for behavior that belongs only to one preset agent. A skill’s metadata.tools fits a group of tools that several agents share, so the instructions and the tools travel together. Tools only one agent needs fit on the preset. See Where tools come from for the full picture, including what the ai.preset_agent actions input does to a skill’s tools.