skillbrew

Skillbrew

Curate, adapt, and update a personal collection of Agent Skills from across GitHub.

Skillbrew vendors each skill as ordinary files, records its exact upstream revision, and optionally applies a semantic patch with the Cursor Agent SDK. Updates preserve local intent while incorporating upstream improvements, producing a normal Git diff for review.

Install

Any agent harness

The recommended installer is the open-source skills CLI. It detects installed harnesses and places Skillbrew's generated skills in their native skill directories:

npx skills add redox/skillbrew

Install every Skillbrew skill globally in every supported harness without prompts:

npx skills add redox/skillbrew --skill '*' --agent '*' --global --yes

Target particular harnesses or install only one skill:

npx skills add redox/skillbrew --agent cursor --agent codex --agent claude-code
npx skills add redox/skillbrew --skill brainstorming

The installer supports Cursor, Codex, Claude Code, OpenCode, Pi, Antigravity, Kimi, GitHub Copilot, and many other Agent Skills-compatible tools. These are external installation commands; they consume Skillbrew's generated output and do not require Bun.

Native plugin systems

Skillbrew also ships native manifests for harnesses that install plugins directly from Git repositories.

Claude Code:

/plugin marketplace add redox/skillbrew
/plugin install skillbrew@skillbrew

Antigravity:

agy plugin install https://github.com/redox/skillbrew

Factory Droid:

droid plugin marketplace add https://github.com/redox/skillbrew
droid plugin install skillbrew@skillbrew

Kimi Code:

/plugins install https://github.com/redox/skillbrew

Pi:

pi install git:github.com/redox/skillbrew

For OpenCode, add the Git-backed package to opencode.json; see the OpenCode instructions.

Develop the collection

You only need Bun when adding, updating, or validating the collection itself. Install the development dependencies, then use the Skillbrew CLI through the project scripts:

bun install
bun run start list

Skillbrew commands in this repository always begin with bun run start:

bun run start list
bun run start add <source>
bun run start update [name]

Updating skills

Update one skill:

bun run start update brainstorming

Update the entire collection:

bun run start update

Preview an update without writing files:

bun run start update brainstorming --dry-run

An update resolves the recipe's update_ref, validates the generated skill, writes the new output and provenance data, and advances the pinned ref. Skills that are already current are skipped without invoking a model.

Review updates with Git. Commit the generated diff to keep it, or restore it to reject it.

Adding a skill

Add a skill from a GitHub directory URL:

bun run start add https://github.com/owner/repository/tree/main/skills/example

Or add a committed skill from a local Git repository:

bun run start add ../another-repository/skills/example

add reads the upstream frontmatter, resolves the exact Git commit, creates a minimal recipe, imports the skill, and writes its lock entry. A plugin or repository URL also works when it contains exactly one nested SKILL.md or one Claude agents/*.md definition; Skillbrew discovers and normalizes it automatically.

Available options:

--name <name>       Install under a different local name
--license <spdx>    Record the upstream license
--ref <ref>         Override the branch, tag, or commit
--path <path>       Override the skill path inside the repository

--ref and --path are useful for repository URLs or branch names that cannot be inferred unambiguously from a GitHub URL:

bun run start add https://github.com/owner/repository \
  --ref feature/agent-skills \
  --path skills/example

Local skills must be committed before they are added. This ensures the generated recipe can be rebuilt from its pinned revision.

Semantic patches

A recipe is a complete build definition for one skill:

name: diagnosing-bugs

source:
  repository: https://github.com/example/skills.git
  path: skills/diagnosing-bugs
  ref: 71e0a6b1c7f00000000000000000000000000000
  update_ref: main
  license: MIT

transforms:
  - type: llm
    mode: semantic-rebase
    prompt: |
      Adapt this skill to our engineering environment.

      - Prefer GitHub Issues over Linear.
      - Assume a Rails, Rust, Kubernetes, and DuckDB stack.
      - Preserve upstream wording where it already fits.
      - Never invent commands or internal infrastructure.
    allowed_paths:
      - SKILL.md
      - references/**
    forbidden_paths:
      - scripts/**

validation:
  require:
    - SKILL.md
  max_skill_tokens: 5000