Skills server

Skills are discovered from the workspace:

.agents/skills/<skill-name>/SKILL.md
.claude/skills/<skill-name>/SKILL.md

The canonical project location is .agents/skills. The .claude/skills location is supported as a vendor adapter; canonical definitions take precedence on collisions.

Start the server:

tuls skills /work/project \
  --allow skills.read

Tool

Tool	Capability	Purpose
activate_skill	skills.read	Load one discovered skill's full instructions and resource manifest

Supporting files are not auto-injected

A skill's supporting files are not automatically injected into the model context. The activation result provides resource paths so the model can read only what is needed.

Example layout

.agents/
└── skills/
    └── rust-review/
        ├── SKILL.md
        ├── checklist.md
        └── examples/
            └── review.md

Activation result

activate_skill returns structured content with the skill's name, description, skillDir, full instructions, and a resources manifest of supporting files.

Related

• Workspace layout — where skills live.
• Limits & bounded behavior — skill and catalog limits.