OpenClaw 完整配置指南46 分钟 Skills、Voice、Memory 核心设置“treat openclaw like an untrusted virtual assistant”01:33—— 这是整个配置过程的核心安全原则 视频来源标题OpenClaw Full Course: Setup, Skills, Voice, Memory More频道Tech With Tim观看量131,108 views发布时间1 month ago链接https://www.youtube.com/watch?vvte-fDoZczE时长46:53核心观点这期视频是 OpenClaw 的完整配置教程作者花费 60-70 小时亲自配置后总结出的最佳实践。核心观点是将 OpenClaw 视为不受信任的虚拟助手在享受自动化便利的同时保持安全边界。视频作者明确指出“if you want to go extreme and make sure that you’re very secure, I have a full 50-minute long video that breaks down every step you can take to harden and secure your OpenClaw instance”01:00-01:06这意味着安全配置可以非常深入但本视频聚焦于基础安全 核心功能配置。关键建议运行在 VPS 上而非本地电脑更安全、更便宜、更易管理使用多模型策略Opus 4.6 处理复杂任务Codex 处理编码任务通过技能系统扩展功能配置记忆系统实现长期上下文使用心跳和 Cron 实现自动化任务 安全最佳实践核心原则视频作者强调的安全理念“treat OpenClaw the exact same way. Make sure that if you are connecting it to external resources that those are secure and that you set up separate accounts for”02:00-02:08具体做法不要给 OpenClaw 访问主电脑不要给加密密钥或主邮箱访问权限为 OpenClaw 创建独立账户如独立的 GitHub 账号只连接你愿意承担风险的资源VPS 部署优势视频推荐使用 VPS虚拟私有服务器而非本地部署“It’s going to have things like regular backups, disaster recovery, you know, fire protection, natural disaster protection, and theft protection that you don’t have when you run Open Claw on something like a Mac Mini or your own computer”02:21-02:33VPS 优势定期备份灾难恢复防火/自然灾害/盗窃保护成本低约 $7/月随时可关闭无需投资硬件Hostinger 一键安装演示视频展示了使用 Hostinger KVM2 计划的一键安装流程选择 1 个月或 12/24 个月计划使用优惠码 “techwithtim” 可享额外 10% 折扣取消勾选 “ready to use with AI”手动配置 API Key选择服务器位置输入 API Key演示使用 Anthropic点击 Deploy自动创建 Docker 容器重要提示“make sure that you have credits that you purchase some credits in whatever platform you’re using for the LLM”08:11-08:16作者演示时遇到 API 额度不足的问题需要先在 Anthropic 控制台购买 $10 额度才能正常使用。 多模型配置策略为什么需要多模型视频作者解释了多模型策略的必要性“if you run Opus every single day and you’re doing a lot, you can spend 100, 200, 300, $500 per day. I know someone, one of my friends was spending about $200 per day running this purely with Opus, which is insanity”10:32-10:41成本对比Opus 4.6最强模型但极贵且有速率限制Codex 5.2编码任务足够用几乎无限额度如果使用 ChatGPT $200/月订阅推荐配置策略视频中的配置指令“I want you to always use the Opus 4.6 model by default. But whenever an instruction is provided to you, I want you to determine whether or not you need that level of intelligence or if you can switch to a cheaper model like codex 5.2”16:57-17:11具体策略默认使用 Opus 4.6处理复杂规划任务编码任务自动切换到 Codex 5.2低杠杆任务使用便宜模型每次任务开始时告知使用的模型创建模型切换命令视频演示了创建/model命令/model codex# 输出model set to codex 5.2配置步骤告诉 AI“make a command called /model that allows me to switch between the opus 4.6 and codex 5.2”AI 自动创建命令测试命令是否生效添加 OpenAI Codex 模型视频展示了 OAuth 连接流程请求 AI 提供 OAuth 链接点击链接登录 OpenAI复制重定向 URL粘贴回 OpenClaw 完成配置关键提示“OpenAI has made that available to do and not against their terms of service”09:51-09:55使用现有 ChatGPT 订阅的额度是合规的。 Telegram 集成创建 Bot视频演示了完整的 Telegram Bot 创建流程步骤在 Telegram 搜索BotFather认准蓝色认证标记发送/newbot输入 Bot 名称如 “Tim AI”输入 Bot 用户名必须以_bot结尾如 “TimA123456_bot”复制生成的 Token配置 Bot配置命令# 在 OpenClaw 中运行openclaw telegram connect# 粘贴 Bot Token配对流程点击 AI 提供的配对链接在 Telegram 中启动 Bot发送/start复制配对码粘贴回 OpenClaw 完成配对测试在 Telegram 发送 “hi”Bot 回复 “Hey, Telegram is connected. What do you want me to do next?”19:59-20:03创建主题分组视频演示了如何创建不同主题的分组来组织对话步骤在 Telegram 创建新群组如 “startup ideas”添加 Bot 到群组将 Bot 提升为管理员赋予消息访问权限在群组中 Bot 名称进行测试配置指令“The name of the group chat or the name of the group channel is the topic that I want to discuss in that channel. Don’t talk about anything else and only reply or send me messages in the appropriate channels based on the name of them”24:49-25:02这样 Bot 会根据群组名称理解对话主题保持上下文专注。️ 语音模式配置启用语音转录视频演示了语音功能的配置过程初始尝试“enable speech to text so you can transcribe my audio messages”20:13-20:17遇到问题AI 最初要求提供 OpenAI API Key但作者发现可以使用本地 Whisper“install whisper locally to convert speech to text”22:21-22:24本地 Whisper 配置配置选项选择准确性优先还是速度优先作者选择 “balanced”平衡模式测试流程从手机发送语音消息电脑麦克风可能被 OBS 占用Bot 自动转录并回复验证转录准确性最终效果“Awesome. And finally, it looks like it’s working. We have voice mode enabled. I can now send audio messages for it to transcribe”22:58-23:05️ 技能系统详解什么是 Skill视频作者解释了技能的本质“a skill is effectively just a markdown file… it’s just a bunch of formatted text that explains what the bot should be able to do”27:35-27:42技能结构skills/ ├── skill-name/ │ ├── SKILL.md # 技能说明文件 │ └── scripts/ # 可选的脚本文件 │ └── task.py # Python 脚本创建自定义技能视频演示了创建 “hello world” 技能指令“Make a simple skill that says hello world in five languages. And let’s just go five random languages”30:15-30:20生成的技能文件--- name: hello-world-languages description: Say hello in 5 languages --- # Hello World Languages Say hello in: 1. English 2. Spanish 3. French 4. German 5. Japanese使用方式/skill hello-world-languages内置技能视频提到了一些有用的内置技能GitHub代码版本控制Weather天气查询Spotify音乐播放Voice calling语音通话Create skill创建新技能安装技能方式一通过 Gateway 界面打开 Gateway Dashboard进入 Skills 页面点击安装所需技能方式二通过对话“enable this skill” 复制技能配置技能市场Clawhub视频提到了 Clawhub 技能市场“Claw Hub has a list of pre-built skills that you can use and just build into your bot”27:07-27:12建议“before you start just going and adding all of these random skills, I would suggest that you understand what a skill is”27:14-27:20先理解技能原理再有选择地安装。 记忆系统配置记忆文件类型视频介绍了两种记忆文件1. MEMORY.md长期记忆“this is persistent long-term memory that it will always read before it’s doing some kind of action”34:56-35:00用途存储需要永久记住的信息2. Daily Memory每日记忆“it has memories for each day… Open Claw is only going to read two days of previous memory”35:13-35:27用途记录当天发生的事件增强记忆的配置视频演示了两个关键配置配置 1Memory Flushcompaction.memory_flush.enabled: true“when the conversation gets too long, OpenClaw will trigger a memory flush prompt. So important context gets written to memory files before compaction drops history”36:14-36:22配置 2Session Memorymemory.search.experimental_session_memory: true“lets memory search include recent session transcripts, not just files in memory, improving recall of what just happened”36:23-36:29向量记忆搜索高级视频展示了使用 QMD 后端的向量搜索配置“enable vector memory search using a QMD backend, which is significantly better at searching through the memory than the default backend service”37:09-37:16配置步骤复制官方文档中的配置内容告诉 AI“enable this”AI 自动安装依赖并配置效果记忆搜索更准确、更高效 核心配置文件USER.md用户信息用途存储关于用户的信息视频演示的采访方式“give me an interview and ask me questions on what you want”38:43-38:46示例问题你的名字是什么怎么称呼你你的时区工作时间你的目标示例回答“My name is Tim. You can call me Master Tim. My pronouns are the normal ones, he, him. My time zone is Asia, Dubai. I usually work from 10:00 a.m. to 1:00 in the morning. My goal is to grow my YouTube channel”39:40-39:55IDENTITY.mdAI 身份用途定义 AI 的行为方式、语气、边界默认内容包含Core Truths核心原则Boundaries边界Vibe风格Continuity连续性说明SOUL.md灵魂文件用途更深层的行为指导内容示例“each session, you wake up fresh. These files are your memory. Read them. Update them.”40:48-40:51TOOLS.md工具配置用途存储环境特定配置示例内容SSH 主机和别名首选语音TTS扬声器/房间名称设备昵称 心跳机制Heartbeat什么是 Heartbeat“you can have the Claudebot triggered to wake up effectively every set number of minutes”41:28-41:32默认频率30 分钟需要手动开启配置 Heartbeat视频中的配置指令“update your heartbeat file such that anytime you wake up you review any of the mistakes that you’ve made previously and start making improvements and fixing those mistakes”41:51-41:58生成的 HEARTBEAT.mdOn every heartbeat review recent mistakes, issues, and proposed fixes. When doing this review spawn multiple sub agents and run the tasks in parallel.启用 Heartbeat配置命令“turn on the heartbeat and enable it every 30 minutes”43:58-44:02Heartbeat 适用场景检查邮件查看日历自我改进审查定期检查通知不适用场景长时间运行任务精确时间要求用 Cron⏰ Cron 定时任务Cron vs HeartbeatCron 特点“a cron is a task that can run at a set time at a set frequency”44:34-44:37示例每天早上 9:00 提醒每天备份数据库特定时间发送消息创建 Cron视频演示“in five minutes remind me that I need to finish recording this video”44:58-45:03结果自动创建一次性 Cron4 分钟后执行可在 Gateway 查看历史个人 Cron 示例视频作者分享了自己的 Cron 配置每天早上 9:00 回顾昨天工作每天 3 次定期更新各种提醒任务 Docker 容器访问为什么需要 Docker 命令“if you just type openclaw nothing’s going to happen because this is running inside of a docker container”09:46-09:50访问 Docker 容器的命令步骤 1查看容器dockerps步骤 2进入容器dockerexec-it[容器 ID]/bin/bash步骤 3在容器内运行 OpenClawopenclawVS Code 远程连接视频演示了使用 VS Code 远程连接服务器步骤CtrlShiftP→ “Add New SSH Host”输入 SSH 命令如ssh rootIPCtrlShiftP→ “Connect to Host”选择刚添加的主机输入密码打开文件夹/docker/openclaw-data/openclaw查看文件结构openclaw/ ├── credentials/ ├── cron/ └── workspace/ ├── memory/ │ └── YYYY-MM-DD.md ├── MEMORY.md ├── USER.md ├── IDENTITY.md ├── SOUL.md ├── TOOLS.md └── HEARTBEAT.md⚠️ 注意事项成本控制问题“one of my friends was spending about $200 per day running this purely with Opus”10:38-10:41解决方案设置月度预算限制如 $100配置消费通知邮件使用多模型策略API 额度常见问题“my credit balance is too low to access the anthropic API”07:58-08:02解决在对应平台购买额度先购买 $10 测试语音转录问题遇到的问题电脑麦克风被 OBS 占用语音消息空白解决从手机发送语音消息确保麦克风可用Bot 群组权限必要设置将 Bot 提升为管理员赋予消息访问权限在群组中需要 Bot 才能触发回复可复现的工作流完整配置清单# 1. 安全配置security:-运行在 VPS 上-使用独立账户-不连接敏感资源# 2. 模型配置models:primary:opus-4.6fallback:codex-5.2strategy:复杂任务用 Opus编码用 Codex# 3. Telegram 集成telegram:-创建 BotBotFather-配置 Token-完成配对-创建主题群组# 4. 语音模式voice:-安装 Whisper 本地版-选择平衡模式-测试转录# 5. 记忆优化memory:compaction.memory_flush.enabled:truememory.search.experimental_session_memory:truevector_search:qmd-backend# 6. 心跳机制heartbeat:enabled:trueinterval:30mtask:审查错误并改进# 7. 身份配置files:-USER.md采访方式填写-IDENTITY.md-SOUL.md-TOOLS.md关键设置说明多模型策略节省成本的核心记忆优化提升长期表现心跳机制实现自动化改进主题分组保持对话专注总结这期 46 分钟的视频是 OpenClaw 最全面的配置教程之一。视频作者基于 60-70 小时的实际配置经验系统地讲解了从安全部署到高级功能的全流程。视频的核心价值安全优先始终将 OpenClaw 视为不受信任的助手成本意识多模型策略可节省数百美元/天实用导向每个配置都有实际应用场景可扩展性理解原理后可自行定制适合人群刚安装 OpenClaw 需要完整配置指南的用户想优化现有配置提升效率的用户关心成本控制和安全的用户不适合只需要简单 10 分钟快速上手的用户视频较详细不打算长期使用 OpenClaw 的用户配置模板我整理好了需要的朋友评论区扣「模板」或者直接私信我标签#AI #效率工具 #自动化 #OpenClaw #配置教程