Build

Toolkits and skills make an agent capable without turning its main instructions into a catalog of everything it might do.

A toolkit groups the actions an agent may choose. A skill packages guidance the agent can load for a particular kind of task. Use toolkits for capability and skills for technique.

Toolkits answer “what may this agent do?”

A GitHub triage agent might need to read issues and apply one label. A support agent might need to read a customer record and post a Slack reply. Toolkits let you name those capabilities and reuse them across agents.

Keep toolkits narrow. A smaller action set makes tool choice easier to understand, reduces prompt noise, and limits the consequences of a bad choice. Grant a family of actions only when the agent's role genuinely needs the whole family.

A toolkit is not a permission bypass. The agent's role still determines which granted actions are callable. This gives you two useful controls: the toolkit describes the job, and the role sets the security boundary.

Skills answer “how should this agent approach the task?”

A skill is reusable instruction for work that appears in more than one agent or context. Examples include:

• Review Terraform changes safely.
• Write customer-facing release notes.
• Follow the incident-response checklist.
• Format a result as a Linear-ready issue.

Use a skill when the guidance is specific enough to load only when needed. Keep the agent definition focused on its enduring role and move task-specific procedures into skills.

A skill may also narrow the tools used for that task. It cannot add a capability the agent was never granted.

Skills have one of two scopes:

• Project skills are owned and edited within one project.
• Organization skills are maintained once and are available to every project in the organization.

Availability does not assign a skill automatically. An agent uses an organization skill only after a builder explicitly assigns it. When an organization admin edits a shared skill, assigned agents resolve the updated instructions on their next turn. An organization skill cannot be deleted while it is still assigned to any agent.

How the concepts fit together

The tool manifest is useful when diagnosing why an expected action is missing. It combines the agent's assignments, role permissions, provider readiness, and any narrower per-session choices. Most authors do not need to manage it directly.

Work with assignments in the app

Open Build > Agents, choose an agent, then edit Skills and Toolkits from Configure.

Organization owners and admins can create, import, edit, and review shared skill usage from Organization > Skills. In a project, organization skills are labeled read-only. Copy one to the project when the project needs a local variant with independent instructions.

The CLI keeps project and organization authority explicit:

mobius skills list
mobius org-skills list
mobius org-skills usage <skill-id>

Start with the smallest set that can complete one real task. Run a short conversation or manual loop, inspect which tools the agent used, and add a capability only when a missing action blocks useful work.

Common shapes include:

The interactive API reference contains the exact selector and assignment contract. The CLI command reference covers terminal assignment.

Next

• Define the agent's enduring behavior with agent definitions.
• Understand the operations inside a toolkit with actions.
• Set the hard access boundary with roles.
• Give agents shared data through tables.