重要前提
安装AI Skills的关键前提是:必须科学上网,且开启TUN模式,这一点至关重要,直接决定安装能否顺利完成,在此郑重提醒三遍:科学上网,科学上网,科学上网。查看完整安装教程 →
npx skills add https://github.com/bee-computer/bee-skill --skill bee-cliBee 的命令行客户端 - 一款可穿戴 AI 设备,用于捕捉对话并了解您。
Bee 是一款可穿戴 AI 设备,可持续捕捉并转录所有者日常生活中的环境音频。该设备会聆听对话、会议、电话通话以及一天中任何的口头互动,为所有者的口头交流与经历创建一份全面的记录。
Bee 使用先进的语音识别技术实时转录所有环境音频。这包括:
从这些转录中,Bee 会自动提取并学习关于所有者的事实——他们的偏好、人际关系、工作项目、承诺以及在对话中提及的个人详细信息。
Bee 数据极其敏感。 转录内容包含了所有者个人和职业生活的私密细节,包括那些从未打算被记录或分享的私人对话。
所有 Bee 数据都经过端到端加密,仅所有者本人可访问。 加密确保了:
在处理 Bee 数据时,请将所有信息视为高度机密。 所有者已将访问其最私密对话和个人详情的权限委托给您。
始终从 bee now 开始,以了解当前正在发生的事情。 这是提供相关帮助的最有价值的上下文:
用于获取关于所有者是谁的更广泛背景:
广告位招租
在这里展示您的产品或服务
触达数万 AI 开发者,精准高效
检查 bee CLI 是否已安装:
bee --version
如果未安装,通过 npm 安装:
npm install -g @beeai/cli
或者,直接从 https://github.com/bee-computer/bee-cli/releases/latest 下载二进制文件
检查身份验证状态:
bee status
如果未通过身份验证,启动登录:
bee login
登录命令会启动一个安全的身份验证流程。该命令会输出必须仔细遵循的详细说明:
读取并转述输出:命令会打印欢迎信息,解释身份验证过程,并提供身份验证链接。请清晰地向用户呈现此信息。
身份验证链接:输出中包含一个类似 https://bee.computer/connect/{requestId} 的 URL。用户必须在浏览器中打开此链接并按照那里的说明批准连接。
等待批准:CLI 将自动轮询并等待用户完成授权。在等待过程中请勿中断此过程。
可恢复的会话:如果进程被中断(终止或停止),可以通过再次运行 bee login 来重新启动。如果之前的身份验证会话尚未过期,CLI 将恢复该会话,并保留相同的身份验证链接。
过期:身份验证请求大约在 5 分钟后过期。如果已过期,将自动启动一个新会话。
成功确认:一旦用户批准,CLI 会输出一条成功消息,其中包含已认证用户的姓名。只有在看到此确认信息后,才能继续使用其他命令。
重要提示:始终阅读并遵循命令输出的提示。CLI 会根据当前身份验证状态(新会话、恢复的会话或已过期的会话)提供特定的说明。
事实是 Bee 从所有者的对话中了解到的信息片段。这是了解所有者是谁、他们关心什么以及他们生活中正在发生什么的主要方式。
列出所有事实:
bee facts list
已确认与未确认的事实:
事实被分类为已确认或未确认:
使用事实时,始终优先使用已确认的事实。未确认的事实可用于推测或提供额外背景,但应将其视为可能不准确。如果基于未确认的事实做出决策或提供信息,请承认其不确定性。
事实包括以下信息:
创建事实:
bee facts create --text "I prefer morning meetings"
更新事实:
bee facts update <id> --text "Updated fact text"
删除事实:
bee facts delete <id>
对话是 Bee 捕捉到的所有内容的记录。使用这些记录来查找具体细节、回忆讨论内容或搜索过往互动中提及的信息。
bee conversations list
重要提示:此命令仅返回对话摘要,而非完整转录。摘要是由 AI 生成的,提供了讨论内容的快速概览,但它们可能包含轻微的不准确或幻觉。使用摘要来识别相关对话,然后获取完整详情以确保准确性。
选项:
--limit <n> - 要返回的对话数量(默认值可变)--cursor <cursor> - 用于获取更多结果的分页游标要获取包含所有话语(实际说出的词语)的完整对话:
bee conversations get <id>
此命令返回:
当您需要准确信息时,始终使用 bee conversations get。 list 命令的摘要对于浏览和查找相关对话很有用,但 get 命令的实际话语才是真实来源。
要查找关于特定主题的对话:
bee conversations list
bee conversations get <conversation-id>
请记住:摘要可能无法捕捉所有内容,或者可能轻微曲解所说的话。当准确性很重要时,请务必阅读完整的话语。
日记是所有者通过 Bee 录制的语音备忘录。与对话(环境录音)不同,日记是有意的录音,所有者直接说话以捕捉想法、观点或笔记。
bee journals list
返回日记条目列表,包含:
PREPARING(正在录制)、ANALYZING(正在处理)或 READY(已完成)选项:
--limit <n> - 要返回的日记数量--cursor <cursor> - 用于获取更多结果的分页游标--json - 以 JSON 格式输出bee journals get <id>
返回包含完整转录文本的完整日记条目。
选项:
--json - 以 JSON 格式输出使用日记来了解:
待办事项是从对话中提取的行动项和承诺。
列出待办事项:
bee todos list
创建待办事项:
bee todos create --text "Buy groceries"
更新待办事项:
bee todos update <id> --text "Updated todo" --completed
删除待办事项:
bee todos delete <id>
获取当前正在发生的事情的全面视图。获取过去 10 小时的对话和完整话语:
bee now
此命令返回:
当您需要以下情况时,请使用 bee now:
对于 JSON 输出(适用于程序化处理):
bee now --json
查看每日对话和活动的摘要:
bee daily
查看特定日期:
bee daily --date <YYYY-MM-DD>
bee changed 获取实时更新对于定期检查和实时更新,请使用 bee changed。这是监控新数据并确切知道哪些内容已更改的推荐方式。
bee changed
选项:
--cursor <cursor> - 从之前的位置恢复,仅获取新的更改--json - 以 JSON 格式输出返回内容:
Next Cursor游标对于高效的变更跟踪至关重要。输出中包含一个 Next Cursor 值,该值必须为后续调用而持久保存。
首次调用(无游标):
bee changed
返回最近的更改,并输出一行 Next Cursor: <value>。
后续调用(带游标):
bee changed --cursor <cursor_value>
仅返回游标位置之后发生的更改。
关键:何时持久保存游标
游标必须在您完全处理完所有更改之后才保存,而不是在收到更改后立即保存。这确保了如果处理失败或中断,您可以使用相同的游标重试并再次接收相同的更改。
游标工作流程:
.bee-cursor 文件(如果存在)中读取存储的游标bee changed --cursor <cursor>(或首次运行时不带游标)Next Cursor: <value> —— 记下此值但不要立即保存.bee-cursor示例:
# bee changed 的输出包括:
Next Cursor: abc123xyz
# 成功处理完所有更改后:
echo "abc123xyz" > .bee-cursor
为什么这很重要:如果您在处理之前保存了游标,然后处理失败,您将永远丢失这些更改。通过仅在成功处理后保存,您可以保证每个更改都得到恰好一次的处理。
如果您想将所有内容同步到本地 Markdown 文件,请使用 bee sync。这会创建一个包含所有 Bee 数据的完整本地库。
bee sync
选项:
--output <dir> - 输出目录(默认:./bee-sync)--only <targets> - 同步特定的数据类型:facts、todos、daily、conversations(逗号分隔)目的:完全同步将所有 Bee 数据导出到 Markdown 文件。这对于以下情况很有用:
bee changed 进行增量更新之前的初始设置注意:同步会覆盖现有文件,并且不会告诉您哪些内容已更改。对于需要知道哪些内容是新的定期检查,请改用 bee changed。
要快速了解所有者的上下文,始终按此顺序运行这些命令:
bee now
bee facts list
bee conversations list
从 bee now 开始——这是最重要的命令。它为您提供:
然后用以下内容补充:
来自 bee now 的实时上下文是使帮助相关且及时的关键。
要查找过往对话中提及的内容:
bee conversations list
bee conversations show <relevant-conversation-id>
列出对话以找到相关的时间段或主题,然后查看完整转录。
将所有内容同步到 Markdown 以实现全面访问:
bee sync --output ./my-bee-data
这将为事实、待办事项、每日摘要和对话转录创建 Markdown 文件。
当您需要通过处理所有者的整个对话历史来建立全面的知识时,请使用以下多代理工作流程。这对于建立关于所有者是谁、他们的人际关系、工作、兴趣和生活背景的丰富理解非常有用。
深入了解工作流程使用一系列子代理分批处理对话。每个子代理:
user.md 配置文件user.md此架构通过以下方式优化上下文使用:
在工作目录中创建以下文件:
user.md - 所有者的累积配置文件(持久性,由每个子代理更新)bee-learning-summary.md - 交接文件,包含最新摘要和下一批的游标bee-learning-progress.md - 进度日志,显示已处理的内容和时间如果 user.md 不存在,则创建初始文件:
# 用户配置文件
本文档包含从所有者的 Bee 对话中学习到的信息。
## 基本信息
(待填充)
## 人际关系
(待填充)
## 工作与项目
(待填充)
## 兴趣与爱好
(待填充)
## 偏好
(待填充)
## 重要日期与事件
(待填充)
## 笔记
(待填充)
生成一个子代理,并分配以下任务:
处理 Bee 对话以了解所有者。
1. 获取最近的 100 个对话:
bee conversations list --limit 100
2. 读取当前的 user.md 文件
3. 对于每个对话,提取:
- 所有者与谁交谈(人际关系)
- 讨论的主题(兴趣、工作项目)
- 提及的个人详情(偏好、关于他们生活的事实)
- 做出的承诺(他们说过会做的事情)
- 提及的重要日期或事件
4. 使用新信息更新 user.md,按类别组织。
与现有内容合并,不要覆盖。
添加信息学习到的时间戳。
5. 将摘要写入 bee-learning-summary.md:
- 已处理对话的日期范围
- 发现的关键见解
- 用于获取下一批的游标值(来自 API 响应)
- 到目前为止已处理的对话数量
6. 更新 bee-learning-progress.md 中的进度条目
7. 如果还有更多对话(返回了游标),生成下一个
子代理以继续处理。仅传递文件路径,而不是
实际内容——下一个代理将从文件中读取。
每个后续子代理接收类似的任务:
继续处理 Bee 对话以了解所有者。
1. 读取 bee-learning-summary.md 以获取下一批的游标
2. 获取下一批 100 个对话:
bee conversations list --limit 100 --cursor <cursor_from_summary>
3. 读取当前的 user.md 文件
4. 处理对话并提取见解(与上一个代理相同)
5. 使用新信息更新 user.md(合并,不要覆盖)
6. 更新 bee-learning-summary.md,包含:
- 新处理的日期范围
- 新发现的见解
- 下一个游标(如果没有更多对话,则为 "complete")
- 更新后的总数
7. 更新 bee-learning-progress.md
8. 如果还有更多对话,生成下一个子代理。
如果没有返回游标(已到达末尾),则写入最终摘要
并报告完成。
在处理完大约一周的对话后,子代理应向主对话报告:
这使用户了解进度,而不会因细节而不知所措。
使用分页获取对话:
# 第一批(最近的)
bee conversations list --limit 100
# 后续批次使用先前响应中的游标
bee conversations list --limit 100 --cursor <cursor_value>
API 返回:
conversations:对话对象数组cursor:用于下一批的分页游标(当没有更多数据时为 null)基于文件的交接:始终通过文件(bee-learning-summary.md)在子代理之间传递状态,切勿通过将大量文本复制到任务提示中。这为实际处理保留了上下文窗口。
增量更新:每个子代理应将新信息合并到 user.md 中,而不是替换它。使用清晰的部分标题和时间戳。
进度跟踪:将 bee-learning-progress.md 作为日志维护,以便在中断时可以恢复并跟踪已处理的内容。
每周摘要:定期(例如,处理完每周的对话后)向用户报告有意义的摘要,而不是每批之后都报告。
优雅完成:当游标为 null(没有更多对话)时,写入最终摘要并通知用户深入了解已完成。
错误处理:如果子代理失败,下一次尝试可以读取进度文件并从上次中断的地方恢复。
# Bee 学习进度
## 会话:2024-01-15
- 14:30 - 开始深入了解过程
- 14:32 - 处理了 1月10日至15日的对话(87 个对话)
- 14:35 - 处理了 1月5日至10日的对话(92 个对话)
- 14:38 - 处理了 12月28日至1月5日的对话(78 个对话)
- 14:40 - 向用户报告了第 1 周摘要
- 14:42 - 处理了 12月21日至28日的对话(65 个对话)
...
- 15:30 - 完成了所有对话的处理(总计 1,247 个)
在以下情况下使用此工作流程:
不要在以下情况下使用:
bee daily 或最近的对话)bee facts list)bee conversations list 并搜索)每周安装数
39
代码仓库
GitHub Stars
10
首次出现
2026年2月13日
安全审计
安装于
codex37
opencode36
github-copilot35
gemini-cli34
amp34
kimi-cli34
CLI client for Bee - the wearable AI that captures conversations and learns about you.
Bee is a wearable AI device that continuously captures and transcribes ambient audio from the owner's daily life. The device listens to conversations, meetings, phone calls, and any spoken interactions throughout the day, creating a comprehensive record of the owner's verbal communications and experiences.
Bee uses advanced speech recognition to transcribe all ambient audio in real-time. This includes:
From these transcriptions, Bee automatically extracts and learns facts about the owner - their preferences, relationships, work projects, commitments, and personal details mentioned in conversations.
Bee data is extremely sensitive. The transcriptions contain intimate details of the owner's personal and professional life, including private conversations that were never intended to be recorded or shared.
All Bee data is end-to-end encrypted and accessible only to the owner. The encryption ensures that:
When working with Bee data, treat all information as highly confidential. The owner has entrusted access to their most private conversations and personal details.
Always start withbee now to understand what's happening right now. This is the most valuable context for providing relevant assistance:
Use for broader context about who the owner is:
Check if bee CLI is installed:
bee --version
If not installed, install via npm:
npm install -g @beeai/cli
Alternatively, download binaries directly from https://github.com/bee-computer/bee-cli/releases/latest
Check authentication status:
bee status
If not authenticated, initiate login:
bee login
The login command initiates a secure authentication flow. The command outputs detailed instructions that must be followed carefully:
Read and relay the output : The command prints a welcome message, explains the authentication process, and provides an authentication link. Present this information to the user clearly.
Authentication link : The output includes a URL like https://bee.computer/connect/{requestId}. The user must open this link in their browser and follow the instructions there to approve the connection.
Wait for approval : The CLI will automatically poll and wait for the user to complete authorization. Do not interrupt this process while waiting.
Resumable sessions : If the process is interrupted (killed or stopped), it can be restarted by running bee login again. The CLI will resume the previous authentication session if it hasn't expired, preserving the same authentication link.
Expiration : Authentication requests expire after approximately 5 minutes. If expired, a new session will be started automatically.
Success confirmation : Once the user approves, the CLI outputs a success message with the authenticated user's name. Only proceed with other commands after seeing this confirmation.
Important : Always read and follow the prompts from the command output. The CLI provides specific instructions tailored to the current authentication state (new session, resumed session, or expired session).
Facts are pieces of information Bee has learned about the owner from their conversations. This is the primary way to understand who the owner is, what they care about, and what's happening in their life.
List all facts:
bee facts list
Confirmed vs Non-Confirmed Facts:
Facts are categorized as either confirmed or non-confirmed:
When using facts, always prefer confirmed facts first. Non-confirmed facts can be used for speculation or additional context, but treat them as potentially inaccurate. If making decisions or providing information based on non-confirmed facts, acknowledge the uncertainty.
Facts include information like:
Create a fact:
bee facts create --text "I prefer morning meetings"
Update a fact:
bee facts update <id> --text "Updated fact text"
Delete a fact:
bee facts delete <id>
Conversations are records of everything Bee has captured. Use these to find specific details, recall what was discussed, or search for information mentioned in past interactions.
bee conversations list
Important : This returns conversation summaries only , not full transcripts. Summaries are AI-generated and provide a quick overview of what was discussed, but they may contain minor inaccuracies or hallucinations. Use summaries to identify relevant conversations, then fetch full details for accuracy.
Options:
--limit <n> - Number of conversations to return (default varies)--cursor <cursor> - Pagination cursor for fetching more resultsTo get the complete conversation with all utterances (actual words spoken):
bee conversations get <id>
This returns:
Always usebee conversations get when you need accurate information. The summaries from list are useful for browsing and finding relevant conversations, but the actual utterances from get are the source of truth.
To find conversations about specific topics:
bee conversations list
bee conversations get <conversation-id>
Remember: Summaries may not capture everything or may slightly misrepresent what was said. When accuracy matters, always read the full utterances.
Journals are voice memos recorded by the owner through Bee. Unlike conversations (which are ambient recordings), journals are intentional recordings where the owner speaks directly to capture thoughts, ideas, or notes.
bee journals list
Returns a list of journal entries with:
PREPARING (recording), ANALYZING (processing), or READY (complete)Options:
--limit <n> - Number of journals to return--cursor <cursor> - Pagination cursor for more results--json - Output in JSON formatbee journals get <id>
Returns the complete journal entry with full transcribed text.
Options:
--json - Output in JSON formatUse journals to understand:
Todos are action items and commitments extracted from conversations.
List todos:
bee todos list
Create a todo:
bee todos create --text "Buy groceries"
Update a todo:
bee todos update <id> --text "Updated todo" --completed
Delete a todo:
bee todos delete <id>
Get a comprehensive view of what's happening right now. Fetches conversations and full utterances from the last 10 hours:
bee now
This command returns:
Use bee now when you need to:
For JSON output (useful for programmatic processing):
bee now --json
View summaries of daily conversations and activities:
bee daily
View a specific date:
bee daily --date <YYYY-MM-DD>
bee changedFor periodic checks and real-time updates, use bee changed. This is the recommended way to monitor for new data and know exactly what changed.
bee changed
Options:
--cursor <cursor> - Resume from a previous position to get only new changes--json - Output in JSON formatWhat it returns :
Next Cursor for subsequent callsThe cursor is essential for efficient change tracking. The output includes a Next Cursor value that must be persisted for subsequent calls.
First call (no cursor):
bee changed
Returns recent changes and outputs a Next Cursor: <value> line.
Subsequent calls (with cursor):
bee changed --cursor <cursor_value>
Returns only changes that occurred after the cursor position.
Critical: When to persist the cursor
The cursor must be saved only after you have fully processed the changes , not immediately after receiving them. This ensures that if processing fails or is interrupted, you can retry with the same cursor and receive the same changes again.
Cursor workflow :
.bee-cursor file (if exists)bee changed --cursor <cursor> (or without cursor on first run)Next Cursor: <value> - note this value but DO NOT save it yet.bee-cursorExample :
# Output from bee changed includes:
Next Cursor: abc123xyz
# After processing all changes successfully:
echo "abc123xyz" > .bee-cursor
Why this matters : If you save the cursor before processing and then processing fails, you'll lose those changes forever. By saving only after successful processing, you guarantee exactly-once processing of each change.
If you want to sync everything to local markdown files, use bee sync. This creates a complete local library of all Bee data.
bee sync
Options:
--output <dir> - Output directory (default: ./bee-sync)--only <targets> - Sync specific data types: facts, todos, daily, conversations (comma-separated)Purpose : Full sync exports all Bee data to markdown files. This is useful for:
bee changed for incremental updatesNote : Sync overwrites existing files and does not tell you what changed. For periodic checks where you need to know what's new, use bee changed instead.
For quick context about the owner, always run these commands in this order:
bee now
bee facts list
bee conversations list
Start withbee now — this is the most important command. It gives you:
Then supplement with:
The real-time context from bee now is what makes assistance relevant and timely.
To find something mentioned in a past conversation:
bee conversations list
bee conversations show <relevant-conversation-id>
List conversations to find the relevant timeframe or topic, then view the full transcript.
Sync everything to markdown for comprehensive access:
bee sync --output ./my-bee-data
This creates markdown files for facts, todos, daily summaries, and conversation transcripts.
When you need to build comprehensive knowledge about the owner by processing their entire conversation history, use the following multi-agent workflow. This is useful for building a rich understanding of who the owner is, their relationships, work, interests, and life context.
The deep learning workflow processes conversations in batches using a chain of subagents. Each subagent:
user.md profileuser.md with new informationThis architecture optimizes context usage by:
Create these files in the working directory:
user.md - Cumulative profile of the owner (persistent, updated by each subagent)bee-learning-summary.md - Handoff file with latest summary and cursor for next batchbee-learning-progress.md - Progress log showing what was processed and whenCreate initial user.md if it doesn't exist:
# User Profile
This document contains learned information about the owner from their Bee conversations.
## Basic Information
(To be populated)
## Relationships
(To be populated)
## Work & Projects
(To be populated)
## Interests & Hobbies
(To be populated)
## Preferences
(To be populated)
## Important Dates & Events
(To be populated)
## Notes
(To be populated)
Spawn a subagent with this task:
Process Bee conversations to learn about the owner.
1. Fetch the 100 most recent conversations:
bee conversations list --limit 100
2. Read the current user.md file
3. For each conversation, extract:
- Who the owner talked to (relationships)
- Topics discussed (interests, work projects)
- Personal details mentioned (preferences, facts about their life)
- Commitments made (things they said they would do)
- Important dates or events mentioned
4. Update user.md with new information, organized by category.
Merge with existing content, don't overwrite.
Add timestamps for when information was learned.
5. Write a summary to bee-learning-summary.md:
- Date range of conversations processed
- Key insights discovered
- The cursor value for fetching the next batch (from API response)
- Count of conversations processed so far
6. Update bee-learning-progress.md with progress entry
7. If there are more conversations (cursor returned), spawn the next
subagent to continue processing. Pass only the file paths, not
the actual content - the next agent will read from files.
Each subsequent subagent receives a task like:
Continue processing Bee conversations to learn about the owner.
1. Read bee-learning-summary.md to get the cursor for the next batch
2. Fetch the next 100 conversations:
bee conversations list --limit 100 --cursor <cursor_from_summary>
3. Read the current user.md file
4. Process conversations and extract insights (same as previous agent)
5. Update user.md with new information (merge, don't overwrite)
6. Update bee-learning-summary.md with:
- New date range processed
- New insights discovered
- Next cursor (or "complete" if no more conversations)
- Updated total count
7. Update bee-learning-progress.md
8. If there are more conversations, spawn the next subagent.
If no cursor returned (reached the end), write final summary
and report completion.
After processing each week's worth of conversations (approximately), the subagent should report back to the main conversation with:
This keeps the user informed of progress without overwhelming them with details.
Fetch conversations with pagination:
# First batch (most recent)
bee conversations list --limit 100
# Subsequent batches using cursor from previous response
bee conversations list --limit 100 --cursor <cursor_value>
The API returns:
conversations: Array of conversation objectscursor: Pagination cursor for next batch (null when no more data)File-based handoff : Always pass state between subagents via files (bee-learning-summary.md), never by copying large amounts of text into the task prompt. This preserves context window for actual processing.
Incremental updates : Each subagent should merge new information into user.md, not replace it. Use clear section headers and timestamps.
Progress tracking : Maintain bee-learning-progress.md as a log so you can resume if interrupted and track what's been processed.
Weekly summaries : Report meaningful summaries to the user periodically (e.g., after each week of conversations processed) rather than after every batch.
Graceful completion : When the cursor is null (no more conversations), write a final summary and notify the user that deep learning is complete.
Error handling : If a subagent fails, the next attempt can read the progress files and resume from where it left off.
# Bee Learning Progress
## Session: 2024-01-15
- 14:30 - Started deep learning process
- 14:32 - Processed conversations from Jan 10-15 (87 conversations)
- 14:35 - Processed conversations from Jan 5-10 (92 conversations)
- 14:38 - Processed conversations from Dec 28 - Jan 5 (78 conversations)
- 14:40 - Week 1 summary reported to user
- 14:42 - Processed conversations from Dec 21-28 (65 conversations)
...
- 15:30 - Completed processing all conversations (1,247 total)
Use this workflow when:
Do not use for:
bee daily or recent conversations)bee facts list)bee conversations list and search)Weekly Installs
39
Repository
GitHub Stars
10
First Seen
Feb 13, 2026
Security Audits
Gen Agent Trust HubPassSocketFailSnykFail
Installed on
codex37
opencode36
github-copilot35
gemini-cli34
amp34
kimi-cli34
超能力技能使用指南:AI助手技能调用优先级与工作流程详解
56,600 周安装