SkillForge
Turn YouTube videos into structured, reusable skill files that make any AI agent smarter.
Credential Requirements
SkillForge requires no API keys or credentials of its own. It produces structured skill files from YouTube transcripts; transcript synthesis uses whatever LLM the user's OpenClaw agent is already configured with (e.g. Claude via ANTHROPIC_API_KEY). SkillForge never stores, proxies, or manages provider credentials.
File System Behavior
SkillForge writes to ~/.skillforge/ on first use. Before creating this directory, the CLI prompts for explicit user consent.
Contents of ~/.skillforge/:
- -
library/ — Skill files organized by creator slug and topic (e.g. library/alex-hormozi/lead-magnets/SKILL.md) - INLINECODE5 — User preferences (trusted creators list, consent flag)
- INLINECODE6 — Saved scan proposals for deferred builds
- INLINECODE7 — SQLite full-text search index used by the MCP server and
recall command
SkillForge does not write outside ~/.skillforge/ unless the user provides an explicit --output path.
MCP Server Behavior
The skillforge serve command starts an MCP server using stdio transport (StdioServerTransport from @modelcontextprotocol/sdk). It communicates via stdin/stdout as a subprocess — there is no HTTP server, no port binding, and no network exposure. The host agent (e.g. Claude Desktop) launches SkillForge as a child process and exchanges JSON-RPC messages over the stdio pipe.
The server exposes two tools:
- -
skillforge_recall — Full-text search across indexed skills - INLINECODE15 — List all skills grouped by creator
On startup, the server rebuilds the SQLite index at ~/.skillforge/skills.db from skill files on disk.
Commands
| Command | Writes to INLINECODE17 | Description |
|---|
| INLINECODE18 | Yes | Build a skill from a single YouTube video |
| INLINECODE19 |
Yes | Build from URL, channel, topic, or proposal |
|
scan <url> | Yes | Score videos by relevance and save a proposal |
|
serve | Yes | Start the MCP server (stdio transport) |
|
forge | Yes | Alias for build |
|
trust add/remove | Yes | Manage trusted creators |
|
recall | No | Search the skill library by intent |
|
list | No | List all saved skills |
|
check | No | Check if a skill exists for an intent |
|
suggest | No | Search YouTube for channels by topic |
|
share | No | Export a skill with attribution |
|
check-auth | No | Verify provider authentication |
SkillForge
将YouTube视频转化为结构化、可复用的技能文件,让任何AI智能体变得更聪明。
凭证要求
SkillForge 自身不需要任何API密钥或凭证。它通过YouTube转录生成结构化技能文件;转录合成使用的是用户OpenClaw智能体已配置的任何LLM(例如通过ANTHROPICAPIKEY配置的Claude)。SkillForge从不存储、代理或管理提供商凭证。
文件系统行为
SkillForge首次使用时写入~/.skillforge/。在创建此目录前,CLI会提示用户明确同意。
~/.skillforge/的内容:
- - library/ — 按创作者别名和主题组织的技能文件(例如library/alex-hormozi/lead-magnets/SKILL.md)
- config.json — 用户偏好(信任创作者列表、同意标志)
- proposals/ — 保存的扫描提案,用于延迟构建
- skills.db — MCP服务器和recall命令使用的SQLite全文搜索索引
除非用户提供明确的--output路径,否则SkillForge不会写入~/.skillforge/之外的位置。
MCP服务器行为
skillforge serve命令启动一个使用stdio传输(来自@modelcontextprotocol/sdk的StdioServerTransport)的MCP服务器。它通过stdin/stdout作为子进程进行通信——没有HTTP服务器、没有端口绑定、也没有网络暴露。宿主智能体(例如Claude Desktop)将SkillForge作为子进程启动,并通过stdio管道交换JSON-RPC消息。
该服务器暴露两个工具:
- - skillforgerecall — 对已索引技能进行全文搜索
- skillforgelist — 按创作者分组列出所有技能
启动时,服务器根据磁盘上的技能文件重建~/.skillforge/skills.db处的SQLite索引。
命令
| 命令 | 写入~/.skillforge/ | 描述 |
|---|
| watch <url> | 是 | 从单个YouTube视频构建技能 |
| build [url] |
是 | 从URL、频道、主题或提案构建 |
| scan
| 是 | 按相关性对视频评分并保存提案 |
| serve | 是 | 启动MCP服务器(stdio传输) |
| forge | 是 | build的别名 |
| trust add/remove | 是 | 管理信任创作者 |
| recall | 否 | 按意图搜索技能库 |
| list | 否 | 列出所有已保存的技能 |
| check | 否 | 检查某个意图是否存在技能 |
| suggest | 否 | 按主题搜索YouTube频道 |
| share | 否 | 导出带有归属信息的技能 |
| check-auth | 否 | 验证提供商身份验证 |