Skip to main content
A skill is a directory containing SKILL.md and optional supporting files. Project skills live under .polpo/skills/{name}/. Polpo Cloud uses progressive loading: the system prompt lists assigned skill names, while the automatically registered skill_list and skill_read tools expose metadata and load a relevant body on demand. The OSS runtime’s direct prompt builder injects the full body of assigned skills. In both cases an agent can only access skills assigned in its skills array.

SKILL.md

The runtime parser reads name, description, and allowed-tools. The directory name is used when name is omitted. tags and category are project index metadata stored separately in .polpo/skills-index.json; they are not runtime frontmatter fields.
A skill’s allowed-tools is metadata. Assigning a skill does not bypass the agent’s allowedTools policy; enable required tools on the agent too. Cloud’s skill_list and skill_read are added automatically for agents with assigned skills.

Assign a skill

Assign the skill in .polpo/agents/designer/agent.json:
Run polpo deploy to sync local skills and assignments. To install a repository directly through the Agent API, use POST /v1/skills/add or client.installSkills(source) with a supported source such as a GitHub repository URL.

Activate a skill for one request

Callers can explicitly prioritize one or more skills already assigned to the effective agent or loop:
Activation is additive and lasts only for that request. Other assigned skills remain available, and the agent configuration is not changed. Polpo rejects names that are not assigned to the effective agent or loop. Clients may expose shortcuts such as /frontend-design, but must send the selected names through polpo.skills; the runtime does not parse slash-prefixed message text. With the React SDK, pass the selection on the individual message:

API

See Skills ecosystem for the distinction between runtime skills and skills installed into coding agents.