agent-skills

Harness-agnostic agent skills by MotionDevLab. One repo, every harness, free-local-first.

Skills ship as plain folders (skills/<name>/ with SKILL.md + skill.json) that load unmodified in Claude Code, OpenCode, and OpenClaude. No lock-in, no registry account.

Skills

Skill	Cost	Backend	When to use it
laya	free, local, offline-capable	Ollama preferred, llama.cpp fallback	Default: routing, triage, scoring, guardrail verification
jev	very cheap, paid API	TypeSafe native endpoint, Vercel Gateway fallback	Escalation: low confidence, high stakes, multilingual quality gaps

Routing rule: default to Laya for zero-cost/offline; escalate freely to very-cheap Jev on low confidence (cost is negligible, so the threshold stays low) or explicit request. See skills/_shared/decision-contract.md.

One-command install

Windows (PowerShell):

irm https://raw.githubusercontent.com/MotionDevLab/agent-skills/main/scripts/install.ps1 | iex

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/MotionDevLab/agent-skills/main/scripts/install.sh | bash

What it does: detects OS, ensures Python 3.10+, downloads the Laya model weights (--NoModel skips this), then symlinks-or-copies each skill into every harness it finds (~/.claude/skills, ~/.config/opencode/skills, ~/.openclaude/skills). Re-running is safe: already-installed skills are reported, not duplicated. Details: docs/INSTALL.md.

Add a skill in 3 steps

1. Copy the template: templates/new-skill/ to skills/<name>/, fill in the TODOs.
2. Register it in root skill.json and check the installer picks it up (install.ps1 -WhatIf).
3. Run scripts/lint-repo.ps1 (must PASS) and open a PR. Full guide: docs/ADD-A-SKILL.md.