Agent skills

anyCode aligns with common Agent Skills conventions: each skill is a directory with a SKILL.md file whose YAML frontmatter includes name and description. An optional executable run in that directory is invoked by the Skill tool (same broad risk class as Bash — approvals / sensitive-tool policy apply).

Layout

• User-wide default root: ~/.anycode/skills/<skill_id>/
• Project overrides (no startup scan): <cwd>/skills/<skill_id>/ or <cwd>/.anycode/skills/<skill_id>/ — resolved when the Skill tool runs if the id is not already in the catalog.
• skill_id must match the directory name and the frontmatter name (ASCII letters, digits, ., _, - only). Mismatches are skipped with a log warning.

Minimal SKILL.md:

---
name: my-skill
description: One line for the model and for `anycode skills list`.
---

# my-skill

Longer documentation for humans (optional).

Optional run (e.g. bash): must be a regular file; executed with the skill directory as cwd, with optional CLI args passed through.

Config (~/.anycode/config.json)

Under skills:

Field	Meaning
enabled	When true, scan skills.extra_dirs then ~/.anycode/skills at startup; build catalog and inject ## Available skills into the default system stack (skipped when system_prompt_override is set).
extra_dirs	Extra scan roots (lower precedence than ~/.anycode/skills; later roots override same id).
allowlist	If set, only these ids appear in the catalog and prompt.
run_timeout_ms	Subprocess timeout for run (minimum enforced in code).
minimal_env	When true, only a small env whitelist (PATH, HOME, USER, etc.) is passed to run.
expose_on_explore_plan	When true and enabled, explore / plan agents also get the Skill tool (default false to limit code execution surface).

CLI

anycode skills list   # id, has run, description, root
anycode skills path   # effective scan roots + skills.enabled
anycode skills init <name>   # ~/.anycode/skills/<name>/ + SKILL.md + run template

Model visibility

When skills are enabled and the default system prompt stack is used, the prompt includes an Available skills section listing ids and descriptions. Execution still goes through the Skill tool with {"name": "<id>", "args": [...]}.

Related

• Config & security — approvals and sandbox
• Discovery & test-security — list-tools
• Architecture — bootstrap and ToolServices

Chinese: Agent skills（中文）.