Personal Agent Review
Persistent cloud agents that act on your behalf: 11 Personal Agents (Cue(Manus), Meta Muse, Today AI and more) compared on spec, USD price and model. Updated daily.
Data updated: 2026-11-05 09:38Personal Agents compared (11)
| Agent | Identity stability | Cloud computer spec | VPS | Price (USD) | Channels / model |
|---|---|---|---|---|---|
| Cue(Manus)Manus · 2026-09-28 | Public web tasks only Not for sensitive accountsTested: Datacenter·dynamic | 2 vCPU / 1 GB / 35 GBmeasured / reported | Yes | Free tier / paid from $9.99official | Web, Desktop, Mobile (iOS awaiting App Store review)Not announced |
| LightVela (Tencent Lighthouse)Tencent Cloud Lighthouse · 2026-09-26 | Public web tasks only Not for sensitive accountsTested: Datacenter·static | 2 vCPU / 8 GB / 50 GBofficial | Yes | Free tier / paid from ≈$3.44 (¥23)media report | WeChat, QQ, WeComPre-installed Hermes Agent and DeepSeek… |
| Meta MuseMeta · 2026-09-08 | Public web tasks only Not for sensitive accountsTested: Datacenter·dynamic | 2 vCPU / 8 GB / 100 GBmeasured / reported | No | Free tier / paid from $20media report | iOS, Android, WebMuse Spark 1.3 |
| Claude DuetAnthropic · 2026-10-21 | - | - | - | Free tier / paid from $17official | iOS, Android, WebOpus 5.5 |
| Today AIToday AI · 2026-08-27 | - | 3 vCPU / 5.9 GB / 24 GBagent self-report | - | From $20/moofficial | App (iOS/Android/desktop, see official download page)Not announced (official website states … |
| Simular AI SaiSimular AI | - | - | - | From $50/moofficial | Web, iMessage, TelegramAgent S |
| Vellum CloudVellum | - | 1 vCPU / 2 GB / 6 GBofficial | Yes | From $30/moofficial | Mac, iOS, AndroidOpenAI, Anthropic, Gemini, or open weig… |
| Zentor MateoZentor | - | - | - | From $20/moofficial | Discord, Gmail, NotionNot announced |
| Zo ComputerZo | - | 4 vCPU / 32 GB / 100 GBofficial | Yes | Free tier / paid from $18official | Web, macOS, WindowsNot announced |
| elie222 Rakazoelie222 | - | 2 vCPU / 2 GBofficial | - | From $10/moofficial | Web, macOS, LinuxClaude Pro/Max (Sonnet or Opus), ChatGP… |
| OpenAI DotsOpenAI · 2026-09-29 | - | - | - | Not published | ChatGPT Desktop, Web, iOS/AndroidGPT-6 Astra |
Work Agents (11): digital employees, not personal proxies
| Agent | Identity stability | Cloud computer spec | VPS | Price (USD) | Channels / model |
|---|---|---|---|---|---|
| Grok BotxAI / Cursor · 2026-08-11 | OK for ordinary logins Not for sensitive accountsTested: Datacenter·dynamic · tested ≠ official | 8 vCPU / 16 GB / 128 GBmeasured / reported | Yes | From $20/moofficial | macOS, Windows, LinuxNot announced (depends on subscription … |
| Manus 2.0Manus · 2026-09-28 | OK for ordinary logins Not for sensitive accountsOfficial: Datacenter | 2 vCPU / 1 GB / 35 GBmeasured / reported | Yespaid Cloud Computer | Free tier / paid from $20official | Web, macOS, WindowsNot announced |
| DevinCognition · 2024-12-10 | OK for ordinary logins Not for sensitive accountsOfficial: Datacenter·shared static | - | No | Free tier / paid from $20official | Web, Devin Desktop, CLISWE-2 (self-developed) + optional OpenA… |
| Replit Agent 4Replit | Public web tasks only Not for sensitive accountsThird-party: Datacenter·dynamic | 0.5 vCPU / 1 GB / 2 GBmeasured / reported | - | From $18/moofficial | WebNot announced |
| Cursor ProjectsCursor | - | - | - | Free tier / paid from $8official | macOS, CLI, SlackGrok 4.7, GPT-5.6 Sol, Fable 5.1 Max, O… |
| MarblismMarblism | - | - | - | From $24/moofficial | WebNot announced |
| Metaview fillmoreMetaview | - | - | - | Free tier / paid from $100official | Slack, Email, LinkedInNot announced |
| Rio Cobre Holdings MORE WorkforceRio Cobre Holdings | - | - | - | From $149/moofficial | Web, Phone, Google CalendarNot announced |
| SoWorkSoWork | - | - | - | Free tier / paid from $5.4official | Web, iOS, AndroidClaude (integrable) |
| Sofy AI Web Testing AgentSofy (on Microsoft Marketplace) | - | - | - | From $49/moofficial | Web, Mobile, APINot announced |
| Copilot AutopilotMicrosoft · 2026-09-25 | - | - | - | Not publishedmedia report | Copilot applications (desktop/mobile/web), Teams, OutlookBased on OpenClaw framework (media repo… |
Hosting platforms (9): runtimes, not ready-made agents
| Agent | Identity stability | Cloud computer spec | VPS | Price (USD) | Channels / model |
|---|---|---|---|---|---|
| Hermes Agent(Nous Research)Nous Research · 2026-09-07 | - | 4 vCPU / 2 GBofficial | - | Free tier / paid from $20official | Web, macOS, Windows200+ cutting-edge models (via Nous Port… |
| DigitalOcean Managed Agents (Harness Runtime)DigitalOcean | - | 1 vCPU / 1 GBofficial | - | Not published | Web, CLIClaude Code, Codex CLI, OpenCode, Herme… |
| Letta AgentLetta | - | 1 vCPU / 2 GBofficial | - | Free tier / paid from $20official | CLINot announced |
| PPIO Agent Sandbox (supports integration with OpenAI Agents API)PPIO | - | 1 vCPUofficial | - | Not published | CLI, SDKNot announced |
| PrizedPrized | - | 2 vCPU / 2 GBofficial | - | Free tier / paid from $10official | Web, iOS, AndroidNot announced |
| Vercel DrivesVercel | - | 1 vCPU / 1 GBofficial | - | Free tier / paid from $20official | API, CLINot announced |
| Tencent Cloud Intelligent Agent Development Platform ADP (Agent Harness)Tencent Cloud | - | 2 vCPU / 4 GB / 60 GBmeasured / reported | - | Free tier / paid from ≈$28.12 (¥188)official | WeCom Bot, WeCom App, DingTalkLLM+RAG, Workflow, Multi-agent |
| Alibaba Cloud Agent Studio enterprise-grade Agent full-stack service platformAlibaba Cloud | - | 2 vCPU / 2 GB / 50 GBofficial | - | Not published | -Not announced |
| Alibaba Cloud Wuying Agentic ComputerAlibaba Cloud | - | 4 vCPU / 8 GBofficial | - | Not published | ClientNot announced |
Definitions
- Personal Agent (11, plus 1 unverified): one per person, acts on your behalf, with long-term memory and an always-on cloud computer
- Work Agent (11, plus 7 unverified): digital employees for teams; accounts and permissions follow the organization
- Hosting platform (9, plus 7 unverified): runtime only, not a ready-made agent
Spectrum: office tools → Work Agent → Personal Agent. Borderline products are classified one by one; reasons are on each product page. How identity stability and VPS are tested: test method; each agent’s own runtime: runtime survey.
VPS vs VM (sandbox)
- VM (sandbox): A temporary execution environment provided by the platform; its lifecycle and state may not be stable. Everyday chat and simple tasks run in a VM; every agent, chat assistants included, has one, so it is not listed.
- VPS (cloud computer): A long-lived environment whose state persists, which the user or agent can enter again and again. A VPS always includes a VM, but not the other way round; we only track whether a product has a VPS.
Radar: unverified (15)
Candidates found automatically and re-checked against their official sites daily; once prices or specs match the official page 3 days in a row they move into the tables above. Products with neither published stay here.
Show 15 unverified products
- sundae_bar Sundae · sundae_bar · Personal — Sundae is a personal, always-on AI assistant, accessible vi… site
- Momentic · Momentic · Work — Momentic offers AI QA engineer Mo and a testing platform to… site
- Kognitos AI Employees · Kognitos · Work — Kognitos offers an enterprise-grade deterministic Agentic A… site
- Meta Muse for Small Business · Meta · Work — AI assists small businesses with 24/7 customer service, pro… site
- Vannor (Agentic Marketing Model) · Vannor LLC · Work — Provides fully automated intelligent marketing infrastructu… site
- Shizai AI Shizai Agent · Shizai AI · Work — Shizai Agent is an enterprise-grade AI intelligent agent pl… site
- Google Cloud Gemini (agent for work) · Google Cloud · Work — Google Cloud's enterprise-grade general agent handles knowl… site
- Alibaba Cloud QoderWake CN (Digital Employee) · Alibaba Cloud · Work — 24/7 online autonomous AI digital employee, assisting with … site
- Microsoft Execution Containers · Microsoft · Platform — Microsoft Execution Container (MXC) provides a policy-drive… site
- Huawei Cloud Managed Agents (Zhiguo AgentArts Platform) · Huawei Cloud · Platform — Huawei Cloud provides a one-stop enterprise-grade agent bui…
- PPIO Agent hosting (PPClaw / PPHermes) · PPIO · Platform — PPIO provides cloud hosting services for AI Agents, support… site
- Huawei Cloud AI Asset Marketplace · Huawei Cloud · Platform — Huawei Cloud AI Asset Marketplace: A one-stop cloud AI plat… site
- Volcano Engine Seedance Managed Agent (Seedance MA) · Volcano Engine · Platform — Volcano Engine's fully managed AI Agent service, accelerati… site
- Alibaba Cloud Agent Sandbox · Alibaba Cloud · Platform — Provides AI Agents with a secure and isolated cloud executi…
- Alibaba Qwen Go! · Alibaba · Platform — Alibaba Cloud provides a large model and Agent capability h… site
FAQ
What is a Personal Agent?
A Personal Agent is a persistent agent that is yours: one per person, acting on your behalf. It is built for individual users, keeps long-term memory across sessions, handles email, bookings, shopping and payments under your identity, runs on an always-on cloud computer (24/7) and takes tasks through an app or chat/messaging apps. Tracked here: Cue(Manus), LightVela (Tencent Lighthouse), Meta Muse, Claude Duet, Today AI, Simular AI Sai, Vellum Cloud, Zentor Mateo, Zo Computer, elie222 Rakazo, OpenAI Dots.
How is a Personal Agent different from a Work Agent?
A Personal Agent gives each person a digital proxy, with memory and identity that belong to the individual. A Work Agent gives a team a digital employee that plugs into office apps and shared workspaces, with accounts and permissions that follow the organization. Classified as Work Agents here: Grok Bot, Manus 2.0, Devin, Replit Agent 4, Cursor Projects, Marblism, Metaview fillmore, Rio Cobre Holdings MORE Workforce, SoWork, Sofy AI Web Testing Agent, Copilot Autopilot. Think of a spectrum: office tools (add intelligence to a workflow) → Work Agent (give a digital employee a computer) → Personal Agent (give every person a digital proxy).
What cloud computer spec do Personal Agents run on?
Published or measured specs: Cue(Manus) 2 vCPU / 1 GB RAM / 35 GB storage; LightVela (Tencent Lighthouse) 2 vCPU / 8 GB RAM / 50 GB storage; Meta Muse 2 vCPU / 8 GB RAM / 100 GB storage; Today AI 3 vCPU / 5.9 GB RAM / 24 GB storage; Vellum Cloud 1 vCPU / 2 GB RAM / 6 GB storage; Zo Computer 4 vCPU / 32 GB RAM / 100 GB storage; elie222 Rakazo 2 vCPU / 2 GB RAM; Grok Bot 8 vCPU / 16 GB RAM / 128 GB storage; Manus 2.0 2 vCPU / 1 GB RAM / 35 GB storage; Replit Agent 4 0.5 vCPU / 1 GB RAM / 2 GB storage. The de-facto standard entry form is 2 vCPU / 8 GB RAM, set by Meta Muse.
What is agent identity stability, and how is it tested?
Vendors rarely publish this, so TANDIU tests it in real use: egress IP lookup, comparison with your own device, switching devices, toggling your own egress, restarting the instance and retesting the next day. Each item uses Tested > Official > Third-party; conflicts count the more conservative one. Full steps, verdict rules and terms: tandiu.cn/en/agents/method. Current verdicts: Grok Bot — OK for ordinary logins · Not for sensitive accounts (logins persist and egress can be pinned to your own proxy; if the device providing your egress goes offline, traffic silently falls back to the cloud egress); Cue(Manus) — Public web tasks only · Not for sensitive accounts (dynamic egress IP and no bring-your-own proxy option found; datacenter IPs are easily flagged by risk engines); LightVela (Tencent Lighthouse) — Public web tasks only · Not for sensitive accounts (browser logins are not kept (lost on restart); you have to log in again; datacenter IPs are easily flagged by risk engines); Meta Muse — Public web tasks only · Not for sensitive accounts (dynamic egress IP and no bring-your-own proxy option found; datacenter IPs are easily flagged by risk engines); Manus 2.0 — OK for ordinary logins · Not for sensitive accounts (logins persist and you can switch to your own egress; the vendor itself advises keeping sensitive accounts off the cloud browser); Devin — OK for ordinary logins · Not for sensitive accounts (logins persist and the egress IP is stable; datacenter IPs are easily flagged by risk engines); Replit Agent 4 — Public web tasks only · Not for sensitive accounts (dynamic egress IP and no bring-your-own proxy option found; datacenter IPs are easily flagged by risk engines).
More questions (6)
What is the difference between a VPS and a VM (sandbox)? Can you get in yourself?
A VM (sandbox) is a temporary execution environment provided by the platform whose lifecycle and state may not be stable; every agent, chat assistants included, has one. A VPS (cloud computer) is long-lived, keeps its state, and the user or agent can enter it again and again. We only track whether a product has a VPS: if reopening lands on the same machine with files and logins intact, it does. A VPS does not mean a stable egress IP; that is under identity stability. Current: Grok Bot: Yes; Cue(Manus): Yes; LightVela (Tencent Lighthouse): Yes; Meta Muse: No; Manus 2.0: Yes (paid Cloud Computer); Devin: No; Vellum Cloud: Yes; Zo Computer: Yes. Can you get in yourself (desktop / terminal / SSH): Grok Bot: Yes; Cue(Manus): Yes; LightVela (Tencent Lighthouse): Yes; Meta Muse: No; Manus 2.0: Yes (paid Cloud Computer only, web terminal or SSH, no graphical desktop); Simular AI Sai: No.
Why does the agent's egress IP matter?
The egress IP is the address websites see when the agent's cloud computer goes online. Datacenter IPs are easily flagged as bot traffic and dynamic IPs change; when an agent uses your own accounts, both tend to trigger CAPTCHAs and suspicious-login alerts. Each product's result is in the Identity stability column above.
How much do Personal Agents cost?
Cue(Manus): Free tier / paid from $9.99; LightVela (Tencent Lighthouse): Free tier / paid from ≈$3.44 (¥23); Meta Muse: Free tier / paid from $20; Claude Duet: Free tier / paid from $17; Today AI: From $20/mo; Simular AI Sai: From $50/mo; Vellum Cloud: From $30/mo; Zentor Mateo: From $20/mo; Zo Computer: Free tier / paid from $18; elie222 Rakazo: From $10/mo; OpenAI Dots: Not published. Prices are monthly, shown in USD (CNY prices converted at ¥6.68 per USD) and checked against official pages daily.
Which Personal Agents launched most recently?
Claude Duet (2026-10-21); OpenAI Dots (2026-09-29); Cue(Manus) (2026-09-28); LightVela (Tencent Lighthouse) (2026-09-26). New products are discovered, classified and added automatically every day.
How often is this data updated?
Prices and specs are re-checked against official pages every day, and new launches are searched for daily and classified against the definitions above. Every price carries a confidence label: verifiable on official page, media reports, or not published.
What is covered, and how are the numbers produced?
Prices are monthly in USD (CNY converted at ¥6.68/USD) with a confidence label; official-page prices are checked daily; new launches are discovered, added and classified automatically. Chat assistants and coding tools (ChatGPT, Claude, Codex, Cursor) are out of scope. Items labeled "Tested" (such as egress IP) were measured by TANDIU in real use; no benchmarks or self-made rankings. Products found automatically start in the Radar section at the bottom, are re-checked against official sites daily, and move into the tables after matching 3 days in a row.