Source: Agent Skills overview (Claude Platform docs)
Anthropic's reference for how skills load across the API, Claude Code and claude.ai: three-level token costs, VM/filesystem architecture, surface differences, security and data retention.
Contents
- What: official documentation
- Publisher: Anthropic
- Date: living document, read 2026-11-14
- Original: Agent Skills
Summary
Explains the runtime architecture behind skills on each Anthropic surface and the constraints that differ between them.
Key claims
- Skills run in a VM with filesystem and bash; the model reads
SKILL.mdwith bash when triggered and runs bundled scripts, receiving only their output. - Token costs: level 1 metadata about 100 tokens per skill, always loaded; level 2 instructions under 5k tokens when triggered; level 3 resources cost nothing until read.
- Pre-built skills:
pptx,xlsx,docx,pdf(API, claude.ai, Claude Platform on AWS, Foundry); not in Claude Code, which bundles its own (e.g. the Claude API skill). - Custom skills do not sync across surfaces: claude.ai uploads, API uploads and Claude Code filesystem skills are separate. API skills are workspace-wide.
- Runtime differences: API has no network access and no runtime package install; claude.ai network access varies by admin setting; Claude Code has full network access.
- Security: treat a skill like installed software; audit all bundled files; skills that fetch external URLs are especially risky since fetched content can carry instructions.
- Data handling (as of 2026-11): Agent Skills are not ZDR-eligible and not HIPAA-eligible.
- Messages API: list
skill_idsin thecontainerparameter together with the code execution tool; Managed Agents attach skills in askillsarray.