The open, vendor-neutral toolchain + registry for AI agent skills.
Scaffold · validate · test · install · publish skills for Claude Code · Cursor · Codex · OpenCode · Cline — from one CLI.
Millions of developers are now writing agent skills (the SKILL.md packs that
Claude Code, Cursor, Codex, OpenCode and Cline consume). But everyone is shipping
either a static list of skills or a vendor-locked format. There is no open,
portable way to author, validate, and share them.
SkillFoundry (forge) is the missing toolchain. It defines an open
SKILL.md standard (with a published JSON Schema any client can adopt), ships a
curated starter pack, and gives you init / validate / test / install /
publish — all vendor-neutral, all from one forge command.
| Static skill repos | Vendor tools | SkillFoundry | |
|---|---|---|---|
| Validate a skill | ❌ copy-paste | ❌ | ✅ forge validate |
| Test a skill offline | ❌ | ❌ | ✅ forge test |
| Write to every client | ❌ manual copy | ❌ one vendor | ✅ forge add --client all |
| Open standard (JSON Schema) | ❌ | ❌ | ✅ forge schema |
| Share / discover skills | ❌ | ❌ | ✅ forge publish + registry |
| Quality gate vs "AI slop" | ❌ | ❌ | ✅ built-in rules |
pipx install skillfoundry # recommended
# or, zero-install:
uvx skillfoundry@latest
# or, from source:
pip install -e .forge init my-skill --template advanced --description "when the user needs X"
forge validate my-skill # ✔ passes the open SKILL.md spec
forge test my-skill # ✔ offline structural tests green| Command | What it does |
|---|---|
forge init <name> |
Scaffold a validation-ready skill (--template basic|advanced) |
forge validate [path] |
Check a skill against the spec; CI-friendly exit codes (--links, --strict) |
forge test [path] |
Offline structural tests: validation, examples, entrypoint smoke |
forge doctor |
Diagnose detected clients, installed & broken skills |
forge search <q> |
Search the registry (--tag, --client) |
forge add <name> |
Install a skill into a client (--client all|claude-code|...) |
forge list |
List skills installed into each client |
forge publish [path] |
Add to local registry, or prep a community PR (--dry-run) |
forge schema |
Print/export the open-standard JSON Schema (--out file) |
A skill is a directory with a SKILL.md whose YAML frontmatter is validated by
forge and described by a published
JSON Schema
— so any editor or client can adopt it.
---
name: security-audit # kebab-case, unique
description: Use when reviewing code for vulnerabilities and proposing fixes. # 20–300 chars, say WHEN
version: 1.0.0 # SemVer
license: MIT # SPDX id or 'Proprietary'
tags: [security, refactor] # <=10, suggested vocabulary
compatible_clients: [claude-code, cursor, codex, opencode, cline]
examples:
- "Audit this route for SQL injection"
---
# Security Audit
## When to use
...Run forge schema --out skill.schema.json to drop the spec into your own repo.
Fifteen ready-to-use skills, each passing forge validate + forge test:
security-audit · write-tests · explain-codebase · pr-review · doc-gen ·
commit-msg · dependency-upgrade · sql-optimize · i18n · api-client-gen ·
refactor-smell · changelog · test-data-gen · explain-error · release-notes
Install one into every detected client:
forge add security-audit --client allforge publish ./my-skill --registry localwrites it to your localregistry.json.forge publish ./my-skill --registry communityrecords it and (with the GitHub CLI) gives you the exact PR to contribute it to the public registry.
No server to run, no account to create — the registry is just a git repo of
registry.json. Network effects do the rest.
-
forge test --live(LLM evaluation harness) - Editor snippets / VS Code extension for authoring
-
forge syncto keep installed skills at the latest registry version - Web registry browser
Fork → forge init → write → forge validate && forge test → open a PR.
See CONTRIBUTING.md. Good first issues are tagged.
MIT © SkillFoundry Contributors