OpenAI dots, Grok Bot, Meta Muse and Gemini Spark on one side, OpenClaw and Hermes on the other. What they are, how they differ, and how much is actually known.
OpenClaw and Hermes are probably familiar names by now. You may also be hearing about OpenAI's "dots" and Grok's "Bots". The field includes versions from Google, Meta, Microsoft and Anthropic too. This article covers what personal agents do, the differences between hosted and open-source options, their intended users, and how they work. It also compares what each offers. I haven't tried the hosted products. My sources are vendor documentation, project documentation and press coverage. Where the evidence is weak, I've said so. Names change quickly here. The ones used in this article are current as of 5 October 2026.
These products share five components. An AI model, an LLM such as GPT, Gemini or Claude, interprets your request and decides what to do next. Tools and connectors give it access to things like mail, calendars and a browser. Memory carries information across sessions. A scheduler starts work without waiting for you to ask. The agent returns its results through a channel, most often a chat window.
What separates this setup from a chatbot is the combination of a scheduler and a computer that keeps running. Work can continue after you close your laptop. You review the results later.
With a vendor-hosted agent, the company supplies the model and runs the system on a cloud computer it controls. These are the current options:
For a self-hosted agent, you keep an open-source program running on your own computer or a rented server. The main choices are OpenClaw, which a foundation now runs, and Nous Research's Hermes Agent. Both use the MIT license and allow you to pick the model. Each runs a gateway process to connect with chat apps, including Telegram, WhatsApp and Slack. There is no software charge, but you cover the machine and model usage.
Both are extended through skills, which are add-ons that give the agent a new ability, such as reading a calendar or working with a particular website. OpenClaw has a public registry called ClawHub where anyone can publish skills. Hermes has a Skills Hub that installs from several sources, including GitHub repositories, Vercel's skills.sh directory and ClawHub itself. Hermes can also write its own skills. This community library is a real advantage of the self-hosted side. It is also a risk, because a skill is code and instructions written by a stranger. Security researchers reported hundreds of malicious skills on ClawHub earlier this year. Both projects now scan skills when you install them, but I found no data on how well that scanning works.
OpenClaw was the first to arrive. It went viral early this year, and the hosted agents released since then resemble it closely. In June, Microsoft described Scout as "powered by OpenClaw open-source technology". The other vendors haven't stated such a connection. For them, the comparison comes from press coverage and the timing of their releases.
| Self-hosted (OpenClaw, Hermes) | Vendor-hosted (dots, Spark, Muse, Grok Bot) | |
|---|---|---|
| Control | You own the machine, the schedule and the memory files | The vendor owns the runtime and sets the limits |
| Privacy | Files stay on your machine, but prompts still go to a model provider unless you run a local model | Data lives in the vendor's VM under the vendor's policy |
| Model | Any provider, or a local model | The vendor's own model in most products |
| Extensions | Community skills from public registries, which you vet yourself | The connectors and plugins the vendor offers |
| Safety defaults | Mostly opt-in, you harden it yourself | Isolation and approval checks are on from the start |
| Convenience | You install, patch and monitor it | Sign in and connect your apps |
| Cost | Free software, a small server, and per-token model spend with no ceiling | A flat subscription with an allowance that is mostly unpublished |
Privacy and cost are easy to misread in this comparison. I got both wrong myself.
Keeping the agent on your machine means its memory and your files stay there. It doesn't mean their contents never reach another company. The model provider still receives every prompt, email excerpt and tool result you send to its model. To avoid that, you need a local model.
Low costs also depend on the model. Self-hosting is cheap only if the model is cheap or local. OpenClaw's heartbeat wakes the agent every 30 minutes by default. Every wake-up uses tokens. Published monthly estimates run from a few dollars to several hundred. Most come from businesses selling hosting or API routing, so I wouldn't rely on any one estimate. Using a frontier model can make a self-hosted agent more expensive than a $20 hosted plan. Spending is uncapped unless you impose a limit.
The benefits of self-hosting are more specific than those broad privacy and cost claims suggest. You can read and change OpenClaw's memory because it uses plain Markdown files. Its documentation states that "there is no hidden state". Model choice and chat app choice are yours, and access doesn't depend on your country.
Hosted agents have an advantage in their default safety settings. According to OpenAI's documentation, a dot must pass an approval check before sharing information or doing anything that affects your accounts. Rules you write cannot override its built-in rules. Google says Spark requests permission before sending email or spending money. OpenClaw's README, meanwhile, says its tools execute directly on the host unless you set up sandboxing. The security findings fit that difference. A Cloud Security Alliance note lists nine OpenClaw CVEs over four days in March 2026. An April audit found 4 critical and 9 high severity issues in the default configuration of Hermes Agent.
The hosted products don't all impose the same controls. Reviewers report that a user's Grok Bots all use one cloud computer, including the same logins. Anthropic's Claude Code routines run without stopping for approval at each step. The claim that "hosted is safer" concerns architecture and defaults. I could not find measurements showing how well any of these agents withstands prompt injection.
Vendors don't publish who actually uses their agents. Their examples give the clearest indication of whom they want to reach.
The documented tasks almost always finish with a list, a draft or an approval request. Each of xAI's eight roles includes an explicit stopping rule, such as "Do not change budgets or send the message".
A weekly list of events is one of Google's examples. This walkthrough draws on Google's documentation and TechCrunch's hands-on test of that task.
You pay the $20 subscription. Google doesn't disclose how much of its allowance a weekly task like this consumes. The reviewer considered Spark useful but also noted mistakes: a promo code that didn't work, four items instead of the five requested, and redirect links that were broken.
A morning briefing delivered through a chat app is a common starting project in the OpenClaw community.
HEARTBEAT.md. If you want a specific delivery time, such as 7 a.m., the documentation directs you to a scheduled job. That job starts the agent and sends its output to a chat channel.Your costs are the server and the tokens consumed each time it runs. For this task, community estimates begin well under a dollar a month with a cheap model. Those figures assume very small prompts. I'd use them as a floor.
Setting up the same kind of task takes less work in Hermes Agent. Send it "Every morning at 9am, check Hacker News for AI news and send me a summary on Telegram", and it creates the scheduled job itself. Running the server and securing it are still your responsibility.
The failure with the most documentation involves a self-hosted agent. Summer Yue works on AI alignment at Meta. She gave OpenClaw access to her real inbox with instructions to suggest actions and wait for her approval. The large inbox led to the conversation being compacted to fit the model's context. That removed the instruction. OpenClaw proceeded to trash and archive hundreds of emails before she stopped it.
Her experience shows why the location of an approval rule matters. The "confirm first" instruction was just part of the conversation text. Hosted vendors put approval enforcement outside the model, so shortening the conversation cannot erase it.
Hosted agents run into other limits. Payments, two-factor codes and some sites that block cloud browsers can stop them. Sometimes a website prevents the agent from continuing. Sometimes the vendor has decided it should wait for you. Either way, the user encounters an interruption.
I don't have enough usage information to rank these products. I don't think anyone does yet.
Among hosted agents, only Meta Muse has public figures. TechCrunch reported estimates on 25 September: Sensor Tower put downloads at 3.4 million, Apptopia at 4.3 million and Appfigures at about 2.3 million. Those figures count downloads of a free app with heavy promotion from Meta. They don't tell us how many people actively use it. My search turned up no figures for dots, Spark, Grok Bot or the other products.
GitHub stars provide a number for the self-hosted projects. OpenClaw had about 391,000 on 5 October, while Hermes Agent had about 251,000. That tells us about attention, rather than how many people use them.
These are the developments and gaps I'm watching.
Choosing between hosted and self-hosted agents mostly comes down to responsibility when something goes wrong. With hosting, the vendor places its approval system between the model and your accounts. The subscription, lock-in and interruptions are part of that arrangement. Running an agent yourself removes that approval layer and its restrictions. It also makes security your job.
The usual description of self-hosting as private and cheap needs qualification. Privacy is partial when you use a cloud model, and a frontier model makes it expensive. The more concrete benefits are memory you can read, freedom to choose any model and chat app, and access without regional restrictions.
Claude Code carried out the searches and read the pages for this research. That means the Anthropic sections rely on an Anthropic model reading Anthropic's documentation. I've tried to apply the same standard to those sections as to the others. I'd like to hear what holds up from people using these agents for real work.
First published on ringarc.ai on 5 October 2026. Comments and corrections: vikas@probearc.ai.