The engine behind every skill your agents run.
Anthropic Skill Engine is the technical companion to the AgentSkills marketplace: the canonical spec for skill definitions, a linter that catches broken manifests before they ship, and a CLI that takes you from skill init to skill publish.
#One definition format. Three tools.
Agent skills were drifting: every framework invented its own manifest, its own schema dialect, its own permission model. The Skill Definition Format (SDF) fixes the contract — one JSON document that describes what a skill is, what it takes, what it returns, and what it's allowed to touch. This site is the engine room for that contract.
01The Spec
Every field of the SDF v1 manifest — name, version, inputs, outputs, permissions, runtime, examples — with types, required/optional markers, and annotated examples.
02The Validator
Paste a skill definition into the browser and get real lint feedback instantly: semver checks, slug rules, schema sanity, permission review. No signup, no server round-trip — your JSON never leaves the page.
03The CLI
skill init, skill validate, skill test, skill publish — the full command reference with flags, exit codes, and install instructions.
#SDF v1 in 30 seconds
A skill is a single JSON document. Five fields are required; everything else is opt-in. Here's the smallest valid skill:
// the smallest skill that passes validation
{
"name": "echo",
"version": "1.0.0",
"description": "Echoes the input text back to the caller.",
"inputs": {
"type": "object",
"required": ["text"],
"properties": {
"text": { "type": "string" }
}
},
"outputs": {
"type": "object",
"properties": {
"echo": { "type": "string" }
}
}
}→Strict identity
name is a lowercase slug, version is semver 2.0.0 with no leading v. If either is malformed, validation fails hard — marketplaces key everything off this pair.
→Schema-first I/O
inputs and outputs use JSON-schema-style shapes. required entries must exist in properties; unknown types are errors, not guesses.
→Capabilities, not trust
permissions declares exactly what the skill may touch — net.http, fs.read, exec. Undeclared access is denied at runtime; privileged scopes trigger manual review on publish.
#Install in one line
The skill CLI runs the same checks as the in-browser validator, plus scaffolding, local test runs, and publishing. macOS, Linux, and Windows (WSL) supported.
# install the skill CLI
curl -fsSL https://anthropicskillengine.com/install.sh | sh
# verify it
skill --version
# → skill 0.4.0The installer drops a single static binary into ~/.skill/bin and adds it to your PATH. Prefer a package manager? Documented alternatives are on the CLI page — Homebrew and npm shims included.