Agentic OS: 16 AI Assistants Ranked by Tier (2026)

gstack is Garry Tan's personal Claude Code setup, published as open source (MIT licence, 134,729 stars). Tan is the president and CEO of Y Combinator, the startup accelerator. The pack describes itself as "Twenty-three specialists and eight power tools, all slash commands": a CEO reviewer, a designer, a QA tester, a security officer and so on, each a role Claude Code can play on demand.

It's a brilliant way to see how a heavy user structures their skills. But it's a skill pack for one developer, not an assistant that knows you. It sits on top of Tier 1 and makes it sharper.

Official project: github.com/garrytan/gstack

Best for: developers who want a proven set of review and planning skills to copy.

Daniel Miessler's Personal AI Infrastructure was renamed LifeOS in July 2026 (v6.0.0, MIT, 19,265 stars). It's the closest in spirit to how we build: it's designed as "a single Digital Assistant" built around you, with your identity and goals loaded at the start of every session, a memory of past decisions, and a denylist that "blocks dangerous operations regardless of what any prompt says."

Miessler comes from the security world, and it shows. LifeOS treats safety as part of the design rather than an afterthought.

Official project: github.com/danielmiessler/LifeOS

Best for: people who want a deeply personal assistant for their whole life, not just work, and are happy running it on Claude Code.

NanoClaw (MIT, 30,867 stars) is a smaller, security-first take on the same idea. Every agent runs inside its own container, a sealed-off box on the computer that it cannot see out of, so "Bash access is safe because commands run inside the container." It connects to WhatsApp, Telegram, Slack, Teams and iMessage, keeps a file-based memory, and runs scheduled tasks.

It describes itself as "Built for the individual user." It does have user roles, so a few people can share an agent, but its own docs warn that between people in the same group, "Information will cross-pollinate through agent memory."

Official project: github.com/nanocoai/nanoclaw

Best for: anyone who wants a chat-app assistant and cares about containment.

Hermes Agent from Nous Research (MIT, 250,661 stars) is the one with the most interesting memory. It keeps "bounded, curated memory that persists across sessions" plus a user profile with "your preferences, communication style, expectations", and it writes new skills for itself after complex tasks. It runs from a laptop or a $5 server and talks to you through Telegram, Discord, Slack, WhatsApp and Signal.

Its security policy is refreshingly direct: "Hermes Agent is a single-tenant personal agent." Several people can talk to the same instance, but it doesn't keep them apart.

Official project: github.com/NousResearch/hermes-agent

Best for: people who want an assistant that learns and improves on its own, and who want to see a memory system worth copying.

The Agentic OS from Simon Scrapes is the paid option: a private repository you get through his Agentic Academy community (a monthly or annual membership). It's built on Claude Code for business owners rather than developers, with brand files, a personality file, a curated memory, 26 skills covering marketing, strategy and operations, scheduled jobs that "run autonomously" since version 1.1.4, and a local Command Centre dashboard.

It handles several clients from one install, but that's one operator managing many clients, not many people sharing it. His "Team OS" is in beta, and we haven't been able to verify how it works, so its multi-user and permissions marks are unconfirmed.

Best for: non-developers who want a finished, supported system rather than building their own.

Google announced Gemini Spark at its I/O conference on 19 May. It started with testers and a beta for Google AI Ultra subscribers in the US, and now comes with Google AI Pro ($19.99 a month) and Ultra. Google calls it "Your 24/7 personal AI agent", and it works in the background "even if your phone and laptop are turned off." It connects to Gmail, Calendar, Drive, Docs and other Google apps (those connections are off until you turn them on), and you can save reusable instructions for jobs you repeat.

It's designed to check with you before major actions. But Google's own help page is candid about the limit: if a scheduled task runs while you're offline, you may not be able to stop an unintended action. It isn't offered in the UK, the European Economic Area or Switzerland.

Official page: gemini.google/overview/agent/spark

Best for: people who live in Gmail, Calendar and Docs and want an assistant with no setup.

Meta launched Muse in the US on 8 September, on iOS, Android, muse.ai and WhatsApp, and it's now in the US and Canada. It's free with a usage limit, or $20 a month (Power) and $100 a month (Maximum) for more. It's the most clearly aimed at non-technical people of the six.

It also has the most thought-through safety design. Each person's Muse runs on a "Muse Secure VM", a dedicated virtual machine that holds the agent and the person's data. And "A separate Sentinel agent runs on that same machine, kept apart from Muse at the system level": nothing Muse does reaches the internet unless the Sentinel approves it. Under the hood it does more than it shows: Meta's security page says it "launches swarms of subagents" and builds its own tools. A small-business version, launched on 29 September, connects Muse to a company's Facebook Pages, Meta ad accounts, Shopify and more.

Official announcement: about.fb.com

Best for: non-technical people in the US or Canada who want the strongest guardrails of the ready-made options.

Manus launched Cue on 28 September, a new app for your personal agents on phone and desktop. It goes further than the others in giving an agent a life of its own: "In Cue, each agent has its own email, phone number, wallet, and computer, so it can send messages, pay within the budget you set, and see a task through on its own machine." You choose what it can access and which actions need your approval.

You can also "Put several agents in a group chat with a shared goal, and they hand work to each other." It's in early access and free with an invite code.

Why provisional: Cue has no help or documentation pages yet. Manus hasn't published how its memory works, which model it uses or where it's available, so memory is marked ✗ (not documented) and we'll revisit this placement once it has.

Official announcement: manus.im/blog/introducing-manus-2-0

Best for: early adopters who want an agent that can message and pay on their behalf, inside a budget.

OpenAI announced dots at its developer day on 29 September. In its words: "Powered by GPT‑6 Astra, they have their own cloud computer, learn from feedback over time, and can work towards your goals 24/7." Your first dot "is included in your Pro or Business Premium plan at no extra cost", and ChatGPT Pro starts at $100 a month.

A dot builds on your existing ChatGPT memories, connects to your apps through plugins, and works in ChatGPT, Slack and Teams. Custom Rules decide how far it goes on its own, from acting without asking to handing the task back to you. It's rolling out "in markets excluding the European Economic Area, Switzerland, and the UK."

Official announcement: openai.com/index/introducing-dots

Best for: heavy ChatGPT users already on a Pro or Business plan.

Cowork is Anthropic's agent for office work, with no terminal involved. Anthropic's own guide is direct about what that means: "Claude has access to the local files you grant it permission to access, and can take real actions on your behalf." It launched as a research preview in January, and on 16 September Anthropic merged Cowork and chat into one Claude, rolling out to Pro and Max plans.

Plugins bundle skills, connectors and sub-agents, and Anthropic says "Claude breaks complex work into smaller tasks and coordinates parallel workstreams to complete them." Tasks can run on a schedule, approval modes range from asking every time to acting on its own, and "Project sharing is available on Team and Enterprise plans", with view or edit access per person. Claude Pro costs $20 a month billed monthly. Of the six, it's the only one confirmed for the UK.

Official announcement: claude.com/blog/cowork-is-now-claude

Best for: office workers who want an agent on their own documents, and UK readers who can't get the others yet.

This is the redesign that started all the fuss. Announced on 17 September, it turns a project into a single long-running conversation: "Projects have threads that do the work and a coordinator that directs them." You brief the coordinator like a team lead, and it opens threads to do the work. Each thread is usually a session running on Anthropic's servers, working on its own copy of the code, and they keep going after you close your laptop.

Each cloud thread also starts with the project's knowledge already loaded: its written instructions, the CLAUDE.md files from its code (the instruction files Claude Code reads at the start of every session), their skills, and a shared MEMORY.md index of what the project has learned.

So it has orchestration and unattended work, the Tier 3 features. What it doesn't have yet is the reach. It's in a beta for Pro and Max plans, there are "no organization-level controls", and "A project belongs to one user. You can't share a project."

It's also where most of the excitement gets it wrong. The two features people are most excited about, a search layer over project files and team permissions, belong to the older chat Projects, not this redesign. We cover what actually shipped below.

Official docs: code.claude.com/docs/en/claude-projects

Best for: Claude Code users who want to hand off bigger jobs and let them run in parallel.

OpenClaw is the giant of the category, with 391,182 stars, an MIT licence and, in its own words, "no paid tier, hosted service, or token." Peter Steinberger built it as a personal assistant that runs on your own hardware and lives in your chat apps, with a personality file, Markdown memory, skills, scheduled automations and sub-agents.

It has also gone further than most toward teams. Its multi-user mode "lets several trusted people operate the same OpenClaw agent", with shared sessions "the whole team can open, steer, and take over" and named roles. That's the Tier 4 idea. But the same page says "A gateway is one trust domain." Everyone on the team is trusted with everything, which is fine for three co-founders and risky for a company. We come back to why in the security section.

Official project: github.com/openclaw/openclaw

Best for: a single power user, or a small team that fully trusts each other.

xAI launched Grok Bot in beta on 11 August. It comes with SuperGrok ($30 a month) or a paid Cursor plan, and you reach it from desktop and mobile apps. Each named Bot keeps its own memory, files and preferences, can pick up a skill by watching you do a task once, and runs routines on a schedule or when something happens.

The catch is in xAI's own security documentation: "All of your Bots share one cloud computer assigned to your user account." The files, browser sessions and logins on it are available to every Bot, so separate Bots don't keep work apart.

Grok Bot is also the only ready-made option built to be shared. "A Team Bot is one Bot that your whole team talks to." Only its owner can change its setup, admins choose which groups can use it, and xAI says "A Team Bot never lends one person's access to another." Bots can also run in parallel and pass work to each other.

Official announcement: x.ai/news/introducing-grok-bot

Best for: power users and developers already paying for SuperGrok or Cursor.

HAL is the AI infrastructure I built to run StationX, and it's our worked example of Tier 3. It runs on Claude Code, but over the years it has been plugged into real systems: it manages cloud infrastructure across 6 providers and controls more than 108 tool integrations. Around 80% of a company serving half a million customers now runs through it. It scans our servers and code for vulnerabilities every week, ranks what it finds, reports into Slack, publishes content, and runs our sales and billing workflows.

Take one job it runs end to end. I point it at the open-source WordPress plugins that power much of the web, and it works through them hunting for security holes, across a pool of tens of thousands of plugins. When it finds one, it doesn't just flag it. It proves the hole is real by exploiting it on a safe test site, then produces a short video that shows the attack happening step by step. Finding the bug, proving it and packaging the proof used to be three separate jobs for a security researcher. HAL does the whole chain, and it has turned up real zero-day vulnerabilities: flaws the software's own makers didn't know were there.

It does the other side of the business too. Most of what you're reading right now came through HAL: the research behind our YouTube videos and articles, the thumbnails, the social posts, and the course and training material. One system runs the security work and the content work, because both are just workflows I've taught it.

The governance is what makes that safe enough to do. Hooks block dangerous commands before they run, every change to production needs a named approval, and a panel of AI reviewers from different model families checks code before it ships.

HAL is mine, which makes it a single-user system. It's the most powerful thing on this list for one person, and the next tier is about what happens when it isn't just one person. You can read how HAL works.

Paperclip (MIT, 95,962 stars) sits a level above the personal assistants. Its README puts it neatly: "If OpenClaw is an employee, Paperclip is the company." It "orchestrates a team of AI agents to run a business": agents get job titles and a boss, recurring tasks run on schedules, and humans approve work with budget hard-stops.

Switch on its authenticated mode and it supports "Multiple Human Users" with roles and permissions, which is Tier 4 territory. Out of the box, though, it's "Optimized for single-operator local use." Its own memory layer is still on the roadmap.

Official project: github.com/paperclipai/paperclip

Best for: founders who want to run a company of agents with a human board on top.

Hosted HAL is how we took HAL from one person to a team. It runs on our server, and each of our 20 to 30 users gets their own isolated workspace running Claude Code, with the shared StationX company context loaded by default: our standards, our commands, our skills and how we do things.

The shared part is curated. When someone builds a useful skill, it goes through a review and approval step and is then distributed to everyone. And the separation is deliberate: users can't see each other's data, and staff and customers see separate catalogues.

Content is the clearest example. A team member producing a video or an article doesn't need to read a document about how we do it, or ask the person who wrote the last one. They run a command, and the shared context already knows our research method, our thumbnail style and our house voice. The work compounds across the team instead of being trapped on one person's laptop.

Best for: this one isn't for sale. It's the shape we think every AI-native business will need.