Local-first agent capability manager

技能酿造局

把散落在视频、网页、README 和推荐清单里的 Agent 能力，酿成可验证、可安装、可追踪的一套本地能力库。
Turn scattered agent tips into a verified local capability cellar.

查看仓库 / GitHub 快速开始 / Start

Agent 能力不是越装越强，是越清楚越强。

很多人真正缺的不是更多工具，而是一个把工具从“看过、收藏过、装过但忘了”变成可管理资产的中间层。SkillBREW manages the messy space between “I saw a useful tool” and “my agent can reliably use it.”

收藏吃灰Saved a tool tip, lost the trail.
重复安装Same capability, different name.
配置漂移Claude Code and Codex do not share the same paths.
黑箱决策The AI says “done,” but what changed?
调用遗忘Installed, then forgotten at call time.

8 步管线，把原料酿成能力。

每一步都能单独跑，中间产物都留在本地目录。Not a one-shot magic trick, but an inspectable craft: collect, understand, verify, judge, dedupe, install, and record.

01采集
Collect
02理解
Understand
03消化
Digest
04溯源
Verify
05判断
Judge
06去重
Dedupe
07安装
Install
08记录
Record

它真正管三件事。

从素材理解到安装落地之间，SkillBREW 始终保留人工授权门：AI can advise, but the final install is yours to approve.

判断价值 / Judge value

recommend decides whether to install, cherry-pick, or skip the source. No more “install the whole list because it sounded cool.”

检查重复 / Find overlap

dedup scans local Skills, MCPs, repos, and the registry to catch duplicates before they become clutter.

留下台账 / Leave a ledger

record writes RECORD, DASHBOARD, and an installed index so each capability has a source, status, and invocation hint.

从一条链接开始。

默认不会写入系统配置。Start with a dry-run plan, then add --approve only when the plan looks right.

git clone https://github.com/Rookage/skillbrew.git
cd skillbrew
pip install -e .

skillbrew doctor
skillbrew run <url-or-text>
skillbrew verify data/sources/<id>/
skillbrew recommend data/sources/<id>/
skillbrew dedup data/sources/<id>/

# 先看计划
skillbrew install data/sources/<id>/

# 确认后真实安装
skillbrew install data/sources/<id>/ --approve
skillbrew record data/sources/<id>/

不是空壳介绍，是已经有骨架的工具。

项目还早，但已经有端到端管线、测试门禁、运行时识别、真实 MCP 种子和通用安装器的雏形。Early, yes. Vaporware, no.

已验证范围 / Verified ground

  • B站、抖音、YouTube、网页、文本输入源。Bilibili, Douyin, YouTube, webpage, and text inputs.
  • Skill / MCP / repo / 配置 / Prompt 多形态计划。Multi-form install plans.
  • playwright、filesystem、sequential-thinking、context7、github 等 MCP 种子。Verified MCP seeds.
  • fetch-mcp 通过 AI 推断安装链路验证。fetch-mcp passed the AI-inferred install path.

仍在打磨 / Still rough edges

  • memos MCP 暂缓，缺稳定官方 npm 包。No stable official npm package yet.
  • MoneyPrinterTurbo 暂缓，依赖额外凭证和运行环境。Needs credentials and runtime setup.
  • MCP Marketplace 对接仍在 RFC 阶段。Marketplace integration is still an RFC.
  • cli / install / installer 模块需要继续拆分。Core modules need cleanup.

Claude Code 出生，Codex 也要能住。

SkillBREW 的定位不是某一个运行时的小工具，而是跨 Agent 的能力资产管理层。Born near Claude Code, built to grow across Codex and MCP workflows.

运行时 Skill 目录 MCP 配置 clone 缓存
Claude Code ~/.claude/skills/ ~/.claude.json ~/.claude/clones/
Codex ~/.codex/skills/ ~/.codex/config.toml ~/.codex/clones/