xAI Grok BotxAI Grok Bot

模型与开放条件Model & openness

官方没点名底层模型，也没说能不能换模型。

（已查 x.ai 的「Introducing Grok Bot」与「Grok Bot is now included with more plans」两页。）

站内平行证据：同期 Introducing Grok 4.6 页写：

官方原文："a particular focus on long-running agents"

这与 Grok Bot 的常驻定位吻合。

但两页没有互相点名。

所以本站不写「Grok Bot 跑在 4.6 上」（R5：不推断）。

与另两家对比：

• OpenAI Dot —— 明确写了 GPT‑6 Astra
• Meta Muse —— 明确写了 Muse Spark
• Grok Bot —— 留白

是否可换模型、可换哪些，官方未说明。

The official pages name no underlying model and never say whether the model can be swapped.

(x.ai's "Introducing Grok Bot" and "Grok Bot is now included with more plans" pages both checked.)

Parallel in-site evidence: the same-period Introducing Grok 4.6 page writes:

Official wording: "a particular focus on long-running agents"

This matches Grok Bot's always-on positioning.

But the two pages never name each other.

So this site does not write "Grok Bot runs on 4.6" (R5: no inference).

Compared with the other two:

• OpenAI Dot — explicitly names GPT‑6 Astra
• Meta Muse — explicitly names Muse Spark
• Grok Bot — blank

Whether the model can be swapped, and to which ones, is not stated officially.

运行位置Where it runs

形态与另两家同构：一台自己的云端电脑，一直开着。

官方原文：

They have their own computer, work inside tools and apps like you do, and keep working 24/7

另一篇写：

They have their own computer in the cloud. It's always on, with browser and terminal access

有终端，是本条与另两家一个可辨的差异。

官方明确写了 browser and terminal access，而 Dot 与 Muse 官方都只提浏览器。

官方强调的差异点是「工作在真正的工具里，而不是聊天草稿」：

They sign into your apps and work happens in the actual tool instead of a chat draft

可用端：官方写 desktop and iOS。

The form factor is isomorphic with the other two: its own cloud computer, always on.

Official wording:

They have their own computer, work inside tools and apps like you do, and keep working 24/7

Another post writes:

They have their own computer in the cloud. It's always on, with browser and terminal access

Having a terminal is one distinguishable difference between this entry and the other two.

The official wording says browser and terminal access, while both Dot and Muse officially mention only the browser.

The difference the official pages stress is "work happens in the actual tool instead of a chat draft":

They sign into your apps and work happens in the actual tool instead of a chat draft

Available surfaces: official wording says desktop and iOS.

本地文件Local files

官方未提及本地文件系统，工作面在它自己的云端机器。

官方原文用的是单引号，与另一处引用同一句：

They have their own computer. It's always on, with browser and terminal access

可间接看到的工作面是已登录的 app —— 官方例子里有 audit email, Drive, and paid subscriptions，以及：

Works across apps and inboxes

这说明它确实在动邮箱与云盘。

权限粒度与是否全量同步，官方未说明。

（已查两篇发布页与 x.ai/bot 治理规范。）

The official pages never mention the local file system; its workspace is its own cloud machine.

The official wording uses single quotes, citing the same sentence as elsewhere:

They have their own computer. It's always on, with browser and terminal access

The workspace visible indirectly is the signed-in apps — official examples include audit email, Drive, and paid subscriptions, plus:

Works across apps and inboxes

That shows it really does touch mail and cloud drive.

Permission granularity and whether there is full sync are not stated officially.

(Two launch pages and the x.ai/bot governance spec checked.)

关机后的任务Tasks after shutdown

官方直接给了时长承诺：24/7。

首发页原文：

keep working 24/7

另一篇原文：

Bots work across apps and inboxes, keep going when you step away, and only pull you in for judgment calls

这一维度里最清楚的一句是「only pull you in for judgment calls」。

它明确了回调条件是「需要判断的时候」，而不是定时汇报。

还有一个官方给的固化机制：

Ask a Bot to follow along the next time you do the job, so it can run on its own after that

也就是教一次之后自己跑。

并明确它可并行：

Stand up a researcher, writer, and chief of staff. Put them in a group chat so they pass work between themselves

后台自主时工具的权限级别（只读还是读写），官方未说明。

（已查两篇发布页与 x.ai/bot 治理规范 —— 后者只列必审动作，不述权限级别。）

The official pages give a duration promise directly: 24/7.

Launch page wording:

keep working 24/7

Another post's wording:

Bots work across apps and inboxes, keep going when you step away, and only pull you in for judgment calls

The clearest sentence in this dimension is "only pull you in for judgment calls".

It pins the callback condition to "when judgment is needed", not periodic reporting.

There is also an official hardening mechanism:

Ask a Bot to follow along the next time you do the job, so it can run on its own after that

i.e. teach it once, then it runs by itself.

And it is explicitly parallelizable:

Stand up a researcher, writer, and chief of staff. Put them in a group chat so they pass work between themselves

The permission level of tools during autonomous background work (read-only or read-write) is not stated officially.

(Two launch pages and the x.ai/bot governance spec checked — the latter only lists must-approve actions, not permission levels.)

工具与扩展Tools & extensions

官方用「例行工作」清单代替了能力表，这份清单本身就是选型依据。

官方原文列的 jobs：

Bots doing today

清单（九个岗位）：

• Sales prospector —— 销售线索
• Website builder —— 网站搭建
• Digital declutterer —— 数字清理
• Customer support —— 客服
• Game artist —— 游戏素材
• Office manager —— 办公排程
• Inbox manager —— 收件箱
• Meeting stand-in —— 会议代听
• Refunds manager —— 退款跟进

官方点名的具体外部服务：payments provider、Gmail、Slack、ServiceTitan、 Quo、client portal、Drive。（官方点名）

它是三条里唯一没提支付结账的 —— 官方没在 Grok Bot 发布页写它能直接结账。

Meta Muse 明确写了 Stripe Link，Dot 未提。本站不推断，仅记录官方未提。

「Website builder」那条官方写了会 purchases the domain and deploys the final version，涉及采购，审批归属未单独说明。

The official pages replace a capability table with a list of "routine jobs", and that list is itself the selection basis.

The jobs listed in official wording:

Bots doing today

The list (nine roles):

• Sales prospector — sales leads
• Website builder — site building
• Digital declutterer — digital cleanup
• Customer support — support
• Game artist — game art
• Office manager — office scheduling
• Inbox manager — inbox
• Meeting stand-in — meeting proxy
• Refunds manager — refund chasing

Specific external services named officially: payments provider, Gmail, Slack, ServiceTitan, Quo, client portal, Drive. (named officially)

It is the only one of the three that never mentions payment checkout — the official Grok Bot launch page does not say it can check out directly.

Meta Muse explicitly names Stripe Link; Dot does not mention it. This site does not infer; it only records that the official pages do not mention it.

The "Website builder" entry officially writes purchases the domain and deploys the final version — procurement is involved, and approval ownership is not separately stated.

上下文与记忆Context & memory

官方表述偏交互记忆，不是上下文窗口。

官方原文：

They remember conversations, learn how you like things done, and get sharper the more work together

接入口是一条文字线程：

A text-thread UX. Message a Bot the way you would message someone on your team, from mobile or desktop, so you can pick up the same thread on either — with nothing to set up first

即跨端续同一条对话线程。

多 Bot 协作有独立描述：在 group chat 里互相传活，用户不在中间。

上下文窗口大小、记忆的存储与删除机制，官方未说明。

（已查两篇发布页与 x.ai/bot 治理规范。对比 Meta Muse 官方给了删除与 forget skill —— 这是本条明显的信息缺口。）

The official framing is interaction memory, not context window.

Official wording:

They remember conversations, learn how you like things done, and get sharper the more work together

The entry point is a single text thread:

A text-thread UX. Message a Bot the way you would message someone on your team, from mobile or desktop, so you can pick up the same thread on either — with nothing to set up first

i.e. the same conversation thread continues across devices.

Multi-Bot collaboration has its own description: they pass work to each other in a group chat, with the user not in the middle.

Context-window size, and how memory is stored and deleted, are not stated officially.

(Two launch pages and the x.ai/bot governance spec checked. By contrast Meta Muse officially provides deletion and a forget skill — this is a conspicuous information gap in this entry.)

权限与限制Permissions & limits

审批只有一层写在发布页上：「回调即审批」。另有一份官方域内的 Bot 治理规范，补上了必审动作清单。

补核验日期 2026-10-03，那份规范的性质见正文「那份治理规范怎么读」一节。

发布页已核验三处：

1. 回调即审批 —— 「only come back when something needs your approval」、 「Leaves every send for you to approve in your inbox or navigator」 2. 例行任务自带保守动词 —— declutterer「Only discards or unsubscribes if you say so」；refunds manager「Comes back with the refund or discount it recovered」，不自行执行退款，把结果拿回来 3. 一个值得记的设计 —— Website builder 官方写 「Helps configure plugins, delivers a live URL, and adds redirect rules」， 它连改域名解析都做，但官网未把它列为需审批项

官方域内 x.ai/bot/… 的 Bot 治理规范给出了明确的审批清单。

原文前缀：

Explicit approval required before…

必审动作：

• 创建 / 编辑 / 复制 / 隐藏 / 删除 Bot
• 保存或实质修改 Skill
• 创建 / 启用 / 编辑 / 暂停 Routine
• 创建或变更群聊
• 安装或认证连接器
• 扩张权限
• 写入外部源
• 发送或发布
• 联系外部人员
• 采购、转账、改定价、商业承诺
• 删除或覆写数据
• 改动生产环境
• 接受法律条款
• 敏感雇佣 / 医疗 / 法律 / 财务决策
• 创建公开分享链接或市场提交

三条与本档案直接相关的行为规约：

1. 沉默不算同意 —— 官方原文「Do not infer approval from silence, vague agreement, a previous approval, another version, or an example」 2. 审批不外溢 —— 官方原文「Approval applies only to the exact action, target, scope, version, configuration, connection, permission, and schedule」 3. 删除要人工 —— 官方原文「Never delete a Bot automatically」 加「Agents cannot be deleted by another agent; the user deletes from the sidebar」

另有两条凭据规则：

Never store credentials

Never request passwords, passkeys, 2FA, private keys, recovery codes, payment confirmations, or API secrets in chat

产品级机制，官方未说明。

官方没有说明的有四项。

可自定义的权限规则界面、自动审核器、审计日志。以及凭据的技术保管方式。

另有一项未说明：那份规范是产品级强制，还是单个 Bot 的约束。

补到的是一份 Bot 行为规范，不是产品机制文档。

（已查两篇发布页与 x.ai/bot 规范。）

Only one approval layer is written on the launch pages: "callback equals approval". A Bot governance spec on the official domain adds the must-approve action list.

Verification supplemented 2026-10-03; the nature of that spec is discussed in the body section "How to read that governance spec".

Three points verified on the launch pages:

1. Callback equals approval — "only come back when something needs your approval", "Leaves every send for you to approve in your inbox or navigator" 2. Routine jobs carry conservative verbs — declutterer "Only discards or unsubscribes if you say so"; refunds manager "Comes back with the refund or discount it recovered" — it does not execute refunds itself, it brings the result back 3. One design worth noting — Website builder officially writes "Helps configure plugins, delivers a live URL, and adds redirect rules", it even changes DNS redirects, yet the official site does not list it as an approval-required item

The Bot governance spec on the official domain x.ai/bot/… gives an explicit approval list.

Original prefix:

Explicit approval required before…

Must-approve actions:

• Creating / editing / duplicating / hiding / deleting a Bot
• Saving or materially modifying a Skill
• Creating / enabling / editing / pausing a Routine
• Creating or changing group chats
• Installing or authenticating connectors
• Expanding permissions
• Writing to external sources
• Sending or publishing
• Contacting external people
• Purchases, transfers, price changes, commercial commitments
• Deleting or overwriting data
• Changing production environments
• Accepting legal terms
• Sensitive hiring / medical / legal / financial decisions
• Creating public share links or marketplace submissions

Three behavioral rules directly relevant to this dossier:

1. Silence is not consent — official wording "Do not infer approval from silence, vague agreement, a previous approval, another version, or an example" 2. Approval does not spill over — official wording "Approval applies only to the exact action, target, scope, version, configuration, connection, permission, and schedule" 3. Deletion requires a human — official wording "Never delete a Bot automatically" plus "Agents cannot be deleted by another agent; the user deletes from the sidebar"

There are also two credential rules:

Never store credentials

Never request passwords, passkeys, 2FA, private keys, recovery codes, payment confirmations, or API secrets in chat

Product-level mechanisms are not stated officially.

Four things the official pages do not state.

A customizable permission-rule UI, an auto-reviewer, audit logs. And the technical custody method for credentials.

One more unstated item: whether that spec is product-level enforcement or a constraint on individual Bots.

What was found is a Bot behavior spec, not a product-mechanism document.

(Two launch pages and the x.ai/bot spec checked.)

适合什么任务Best for

适合按岗位选人的人。

官方给的是一份岗位化的 jobs 清单（见 tools 维），可以当清单用：销售线索、 网站搭建、邮箱清理、客服、游戏素材、办公排程、收件箱、会议代听、退款跟进。

不适合四类人：

• 需要权限机制可查可配置的人 —— 本条最明显的短板
• 要本地文件读写的人
• 需要确认底层模型的人 —— 官方未公布
• 非英语地区用户 —— 官方只写桌面与 iOS，地区表述只到端、不含国家清单

Fits people who pick agents by job role.

The official pages give a role-based jobs list (see the tools axis), usable as a checklist: sales leads, site building, inbox cleanup, customer support, game art, office scheduling, inbox, meeting stand-in, refund chasing.

Does not fit four kinds of people:

• People who need the permission mechanism inspectable and configurable — this entry's most obvious shortcoming
• People who need local file read/write
• People who need the underlying model confirmed — not published officially
• Users outside English-speaking regions — official wording covers only desktop and iOS; the regional statement goes only to surfaces, with no country list

Digital declutterer — Audits email, Drive, and paid subscriptions around the clock.
Only discards or unsubscribes if you say so.

Refunds manager — Finds the claim, files it, and follows it through.
Comes back with the refund or discount it recovered.

Website builder — Builds the site, purchases the domain, and deploys
the final version. Helps configure plugins, delivers a live URL,
and adds redirect rules.

Digital declutterer — Audits email, Drive, and paid subscriptions around the clock.
Only discards or unsubscribes if you say so.

Refunds manager — Finds the claim, files it, and follows it through.
Comes back with the refund or discount it recovered.

Website builder — Builds the site, purchases the domain, and deploys
the final version. Helps configure plugins, delivers a live URL,
and adds redirect rules.

本页由This page is generated from the ai-handbook 数据层生成（CC BY 4.0）。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.