AIOS 判断Thesis Scheduler Sandbox MCP Bus Kernel 怎么用In use 生态Ecosystem
AI 智能体的操作系统层The OS layer for AI agents

AIOS 你缺的不是模型,
是调度它们的系统。
You're not missing a model.
You're missing the OS.

不是又一个聊天框。是装在你 Mac 上的运行时:你说一句话,它跨模型编排、在系统级沙箱里真办事、全程留痕、还能接入整个 MCP 生态Not another chat box. A runtime on your Mac: say one sentence, and it orchestrates across models, does real work inside a system sandbox, keeps every step on the record, and plugs into the whole MCP ecosystem.

⌥ Space 唤起 · 真执行 · 危险操作先问你⌥ Space to summon · real execution · dangerous ops ask first
AIOS⌥Space
把下载文件夹里的发票整理出来tidy up the invoices in my Downloads
找到 7 个发票,已验证存在。沙箱内按月份归类中…Found 7 invoices, existence verified. Filing by month, in-sandbox…
已归类到 发票/2026 — 7 个文件 rollback: trash_path 可一键还原Filed to Invoices/2026 — 7 files rollback: trash_path, one-click undo
继续说…say more…
我们对未来的判断What we're betting on
模型会越来越多。
缺的是调度它们的操作系统。
Models will keep multiplying.
What's missing is the OS that runs them.
01
编排,比模型更重要。Orchestration beats the model.
再强的模型也只是一步。把理解、拆解、选模型、调工具、验证、回滚串成一条可靠的链,价值在链上,不在某个模型里。Even the best model is just one step. The value is in the chain — understand, decompose, route, call tools, verify, roll back — not in any single model.
02
操作系统应当中立The OS stays neutral.
不绑定任何一家模型、任何一个 Agent。DeepSeek、千问、豆包、以及任何讲 MCP 的工具,都只是插进来的部件——AIOS 负责调度,而不是站队。Bound to no single model or vendor. DeepSeek, Qwen, Doubao, and anything that speaks MCP are just parts you plug in. AIOS orchestrates — it doesn't pick sides.
03
可靠,是地板,不是天花板。Reliability is the floor, not the ceiling.
能办事只是起点。沙箱隔离、危险操作先问、全程留痕、出错能自愈、收到更新能升级——先把"靠谱"做到顶级工具的水平,再谈别的。Getting things done is table stakes. Sandboxing, ask-before-danger, full traces, self-healing, one-click updates — match the reliability of top-tier tools first, then differentiate.
04
护城河不在模型里。The moat was never the model.
它在:把话翻译成对的指令、记得住你、中立地编排、傻瓜化到一步能用、以及一道关不掉的安全闸。这些,才是别人抄不走的部分。It's turning intent into the right instructions, remembering you, orchestrating neutrally, staying one-step simple, and a safety gate that can't be switched off. That's the part no one can copy.
01 · SCHEDULER

编排Orchestrate 每一步,都用对的模型。The right model, every step.

AIOS 自己判断这一步该交给谁:便宜快的活走 DeepSeek,要造物走千问,深推理升级到更强模型,某家挂了自动切换。你只说目标,它决定怎么调度。AIOS decides who handles each step: cheap-and-fast goes to DeepSeek, generation to Qwen, deep reasoning escalates to a stronger model, and a cool-down fallback kicks in when a provider fails. You state the goal; it routes the work.

DeepSeekQwen / 千问Doubao / 豆包能力画像路由capability routing动态 fallbackdynamic fallback
route · task实时live
"写一版落地页文案 + 配图建议""write landing copy + image direction"
DeepSeek-V4文档 · 选中writing · picked
Qwen3-Coder造物 · 备选generation
V4-Pro深推理 · 待命deep · standby
02 · SANDBOX + GUARDIAN

守护Guard 能在你机器上干活,碰不到不该碰的。It works on your Mac. It can't touch what it shouldn't.

真执行最怕"放手就闯祸"。AIOS 用系统级 Seatbelt 沙箱把它关进笼子:默认断网、敏感目录(密钥/钱包/浏览器)读不到、越界写直接拒。危险命令先过安全闸——规则能硬拦的(rm -rf / 及各种绕过写法)永远拦死,模糊的交给模型复核,整条链都失败就回退问你。The scary part of real execution is "let go and it breaks things." AIOS cages it in a system Seatbelt sandbox: network off by default, secret dirs (keys/wallets/browser) unreadable, out-of-bounds writes denied. Dangerous commands pass a gate first — what rules can hard-deny (rm -rf / and its bypass variants) is always denied; the ambiguous goes to an LLM review; if the whole chain fails, it falls back to asking you.

Seatbelt默认断网net-off default脱敏留痕redacted trace硬拦不可翻案hard-deny wins
policy · reviewaios.policy_review.v1
$ rm -rf /规则硬拦HARD DENY
~/.ssh/id_rsa敏感 · 拒读BLOCKED
$ curl example.com沙箱 · 断网NO NET
$ mv invoice.pdf …先问你ASK
// hard_deny 规则优先,LLM 永远翻不了案。整链超时/失败 → ask_user。// hard_deny is rule-first; the LLM can never overturn it. Whole-chain timeout/fail → ask_user.
03 · MCP BUS

接入Plug in 把整个生态,插进来。Plug in the whole ecosystem.

一个操作系统,要能装别人写的程序。AIOS 现在是 MCP 客户端:任何讲 Model Context Protocol 的服务,写进一行配置就接进来,工具自动注入、按信任分级(直接跑 / 先问 / 屏蔽)。会话复用不冷启动、子进程崩了自动重建、一个挂了不拖累其他。这是"成为 AI 的操作系统"的第一块地基。An operating system has to run programs other people wrote. AIOS is now an MCP client: any service that speaks the Model Context Protocol drops in via one config line, its tools auto-injected and governed by trust level (run / ask / deny). Sessions are reused (no cold start), crashed children self-heal, and one bad server never blocks the rest. This is the first foundation stone of becoming the OS for AI.

MCP clientstdio动态注入dynamic tools信任分级trust levels会话自愈self-heal
~/.aios/mcp.json已接入connected
AIOS core · 40+ 内建工具built-in tools
总线注入bus inject
mcp__github__*ask
mcp__filesystem__*trusted
mcp__everything__echotrusted
mcp__yours__…ask
KERNEL · 系统内核the rest of the kernel
一个操作系统该有的,它都在长Everything an OS needs, taking shape

编排、沙箱、生态之外,内核里还跑着这些子系统。Beyond orchestration, sandbox and ecosystem, these subsystems run in the kernel.

Brain / 任务内核task kernel

把一句话变成可执行任务图:理解 → 检查 → 规划 → 执行 → 观察 → 修复 → 验证。多数 Agent 死在缺"观察+验证",这条循环是脊柱。Turns one sentence into an executable task graph: understand → inspect → plan → act → observe → repair → verify. Most agents die without observe+verify; this loop is the spine.

Black Box / 黑匣子trace

每个工具调用都落 trace.tool.v1:用了什么、多久、在不在沙箱、过没过闸——密钥和路径自动脱敏。可审计、可回滚。Every tool call writes trace.tool.v1: what ran, how long, sandboxed or not, which policy — keys and paths auto-redacted. Auditable, reversible.

Memory / 长期记忆memory

记得住你的习惯和常用文件,存在 ~/.aios/memory.md,你能随时看、随时改。越用越懂你。Remembers your habits and frequent files in ~/.aios/memory.md — readable and editable by you anytime. The more you use it, the better it knows you.

Vision / 截图问screen ask

框选屏幕任意区域,原生 OCR 读出文字(中英),接着问:解读、翻译、总结。看不懂的界面,指一下就行。Box any region of your screen; native OCR reads the text (EN/中文) and you ask on top of it — explain, translate, summarize. Point at any UI you don't get.

Shell / 轻量躯壳the shell

Tauri 2 + Rust,几 MB 常驻不卡。⌥Space 唤起、用完收回、回复逐字流式蹦出。平时不打扰。Tauri 2 + Rust, a few MB, always-on and light. ⌥Space to summon, dismiss when done, replies stream token by token. Out of your way otherwise.

Guardian / 自我维护self-upkeep

崩溃能恢复、连接超时会降级、锁中毒能自愈;收到更新提示可一键升级。会照顾自己的系统,才敢交给你日常用。Recovers from crashes, degrades on timeout, heals poisoned locks; one-click update when a new build lands. A system that looks after itself is one you can rely on daily.

怎么用In use
说人话,它真办事Say it plainly. It actually does it.

不用翻菜单、找设置、记路径——说出来就行。No menus, no settings hunt, no paths to remember — just say it.

放首毛不易的歌play a song by Mao Buyi
它做it does抓真实视频直接播,"换一首"在同一标签切歌,绝不编死链。Grabs a real video and plays it; "next" swaps in place — never a made-up link.
把下载文件夹整理一下tidy up my Downloads
它做it does先看清真实有哪些 → 弹确认卡 → 按类型归类。危险动作先问你。Sees what's really there → confirm card → sorts by type. Dangerous moves ask first.
10 分钟后提醒我喝水remind me to drink water in 10 min
它做it does自动算时间,写进提醒事项,带定时。"明天9点开会"也认。Computes the time, files it into Reminders with an alarm. "9am tomorrow" works too.
我的简历在哪where's my résumé
它做it does搜本地文件;教过一次的固定文件(如早报)以后秒开。Searches local files; anything you taught it once (like your daily brief) opens instantly.
记住我爱用 B 站看视频remember I prefer Bilibili
它做it does长期记住偏好,以后据此个性化。存在你能看到的记忆文件里。Remembers the preference long-term and personalizes from it — in a memory file you can see.
帮我做一份开店方案build me a store launch plan
它做it does拆任务、跨模型协作、交付市场/财务/选址/营销一整套,不止"建议"。Decomposes, orchestrates across models, delivers market/finance/location/marketing as a set — not just "suggestions."
未来生态The ecosystem
从一个 app,长成一层操作系统From one app into a layer

MCP 落地的那一刻,"开放生态"不再是 PPT,而是今天就能接的东西。The moment MCP shipped, "open ecosystem" stopped being a slide and became something you can connect today.

01

开发者把 Agent / 工具接进来Developers plug in agents & tools今天就能,via MCPavailable today, via MCP

02

用户挑选、组合需要的能力Users pick and compose skills信任分级,敢用trust-graded, safe to run

03

企业定制自己的 AI WorkforceEnterprises build an AI workforce沙箱 + 留痕,可审计sandboxed + traced, auditable

04

AIOS 负责中立编排、安全与交付AIOS handles neutral orchestration, safety & delivery操作系统,不站队the OS, taking no sides

别再学怎么用 AI。
装一个会替你用的操作系统。
Stop learning how to use AI.
Install the OS that uses it for you.