Archie drafts replies and proposes calendar changes. Both wait for your approval, and no setting turns that offSource 1.
Already have an AI agent? Check the agent you have →
Archie is in beta.
Install Archie, pick your AI account, connect your chat. See the five stepsSource 1.
Personal is free at 20 jobs a day on your own AI account; paid plans start at $30 a month plus AI usageSource 1.
Checked October 1, 2026. Monthly US prices; AI usage and limits differ.
Hosted agents
Lindy: inboxes and workflows, $29.99 per userSource 14. alfred_: email, calendars and tasks, $29.99Source 15. Martin: text, calls, email or SlackSource 16.
alfred_ stores your email and memory on its infrastructureSource 15. Lindy and alfred_ also wait for your approval to send emailSource 14Source 15.
AI subscriptions
ChatGPT offers chat, research and agent features, which vary by planSource 21. Claude Pro is $20 a monthSource 17; Google AI Pro, with Gemini, is $19.99Source 18.
None of these covers Archie’s AI usageSource 1.
Scheduling and work planners
Reclaim schedules tasks, habits and meetings on your calendarSource 20. Motion also covers projects, documents, notes and workflowsSource 19.
If scheduling is the job, a planner may fit best.
Open-source agents
OpenClaw runs on your computer, supports local models, has open code and desktop setup, and no software subscriptionSource 3. AI and hosting costs are separate.
Compare its controls with Archie’s or bring an OpenClaw agent over.
Across: whose computer the work happens on. Up: asks first, what it does before a step that matters, or ease, how you get it running.
We aim for the top right. In the Ease view nobody is there, us included. Read September 18, 2026, September 30 for OpenAI Dots, and October 7 for Tab and Eden.
Hermes, OpenClaw and Vellum publish their code, so you can read and change every lineSource 2Source 3Source 4. Archie’s is not, and our Trust page shows you how to watch every connection it makes. Why Archie’s code is not published →
OpenClaw has five permission modes, one that asks before every command, and per-command allow listsSource 3. Archie’s asking is built in: less to set up, and less to change.
OpenAI Dots, a Grok Bot, Symphony’s agents, Tab and Eden keep going with your laptop shutSource 11Source 6Source 8Source 12Source 13. Archie needs your computer awake, and can wake it for a routine set for a time. See what to leave on.
Our Trust page sets out where the work happens, each gate named above, the local model check, everything Archie sends and to whom, and how to watch it yourself with a network monitor. Archie is a download from our install page, the app row. The overview was checked October 1, 2026 against our five-step walkthrough and pricing page.
vellum.ai, read September 16, 2026. “By default, your assistant runs locally.” And: “High-stakes actions like sending emails to new contacts, modifying important files, or spending credits require explicit approval until you mark them as trusted.”
Read September 18: its permissions docs give four modes, from Strict, “Everything: every action requires explicit approval”, to Full access, “your assistant never asks for permission”. Its download page offers “Vellum for Mac: Native Mac app with computer use” and “Vellum for Windows: Desktop app for Windows PCs”, the app row, and lists the App Store and Google Play. Its homepage says Vellum is open source.
Its pricing page: “Vellum is free to download and start.”
openclaw.ai, read September 16, 2026 and October 1: “Mac, Windows, or Linux. Bring hosted, subscription-backed, gateway, or local models. State lives on your machine, not a vendor cloud.” It has no subscription, no hosted tier and no token, and is MIT licensed.
Its docs on tool and agent permissions, read September 16: “Full Access, including Default (Full Access), authorizes permitted changes without an approval prompt; restricted runs require human approval”. So by default it acts on its own, and you can make it ask.
Read September 18: its exec approvals give five modes, “deny”, “allowlist”, “ask”, “auto” and “full”, an “ask: always” that “keeps prompting”, and per-command allow lists: the most control on this board. It starts with a shell command, “Try it in one command”, the terminal row, and also offers “Desktop Apps: Full apps that install everything for you”.
hermes-agent.nousresearch.com, the Nous Research page, read September 16, 2026: a desktop app for macOS and Windows, the app row, plus a terminal install that also covers Linux and “Deploy to the cloud” through Nous Portal. It is “free and open source under the MIT license”.
Its security docs, read the same day: “Before executing any command, Hermes checks it against a curated list of dangerous patterns. If a match is found, the user must explicitly approve it”, in three modes, “smart | manual | off”.
Anthropic’s Cowork guide, read August 20, 2026. “Open Claude on the web at claude.ai, in the Claude Desktop app, or in the Claude mobile app”, choose Cowork, and “describe the task you want Claude to complete”, the sign-in row. Its work “runs on Anthropic’s servers, in an isolated environment, and your sessions and files are saved to your Claude account”.
Re-read September 16: the mode selector “offers Auto and Manual (default)”, and in Manual “Claude pauses and asks for approval for actions”. In Skip mode, “Claude doesn’t pause to ask and nothing checks its actions automatically”.
x.ai/news/introducing-grok-bot and the Grok Bot overview, read August 20, 2026 and September 16. Each Grok Bot “runs on a persistent cloud VM with a browser, filesystem, and terminal”, and “They finish jobs end to end, and only come back when something needs your approval”. That puts it at the top.
“Download the app, create your first teammate, and start handing off work”, with a Download for macOS button: the app row.
Meta’s announcement, read September 15, 2026: Muse “runs on its own dedicated computer in the cloud, contained so no one else’s agent can reach it”. It is “free for most of what people need, with subscription plans for people who want to do more”, and “rolling out in the US on iOS, Android, and muse.ai”, the sign-in row.
Its permissions, read September 18 and 28, are one of two settings: “Always ask”, or “Ask for some actions”, which asks “before every write action and important read actions”. Its prompt offers “Always allow”, which lets it take “this type of action for this Connector in the future without asking again”, so Muse sits a short way below the top, beside Cowork.
wix.com/symphony, read September 11, 2026. Symphony is used with a Wix account, in a browser or its iOS and Android apps, and Wix publishes no program you run on your own computer: the sign-in row. Wix’s overview says it “checks with you first before your agents act, so nothing important happens without your approval”.
Re-read September 18: “Your agents connect to the suite of tools you already use and can work on their own, 24/7”, and “your agents’ background work uses AI credits”. Wix publishes no sentence saying where that work runs.
Because the agents work while you are away and Wix ships only a phone app and a browser, we place Symphony on the their-servers side. That is our reading, and if Wix publishes the sentence either way we will cite it.
Its pricing page, read September 11: a free plan, “no credit card needed”, whose credits “reset monthly”.
instinct.com, read September 16, 2026: “Text Instinct to get started”, so nothing is installed, the sign-in row. It is “currently available to a private access group as we’re scaling up compute”. Its privacy policy, read the same day, names “third-party hosting”, and says its features “can perform tasks or take actions independently on your behalf, based on the permissions you grant”.
us.norton.com/products/family-assistant, read September 16, 2026: “Asks first, acts after”. Its support FAQ, read September 18, says where it runs: “your family graph, connected-account tokens, agent context, and chat history all live inside your own dedicated instance in the cloud, isolated from other users”.
That puts it on the their-servers side. It “works on any PCs that run in browsers such as Norton Neo, Chrome, and Firefox” or as an iPhone app, the sign-in row.
OpenAI’s announcement, read September 30, 2026: each dot “works on its own cloud computer, while your computer and its contents stay separate unless you choose to connect it”. Its help center: access to your own computer “is optional and starts turned off”. ChatGPT’s guide: a dot “can keep working even when your computer is off”. So it sits on the their-servers side.
Dots “start with built-in rules for when to act independently and when to ask for approval”. One rule you can add, on its controls page, is “Take action without asking”, so it sits a short way below the top, beside Cowork and Muse.
“Create your dot in the ChatGPT desktop app or in ChatGPT on desktop web”: the sign-in row. It comes with ChatGPT Pro and Business Premium. OpenAI’s pages refuse scripts, so we read them through a service that returns a page’s own text.
tab.bot, read October 7, 2026: “Behind the messages, Tab has its own computer and browser”, so it sits on the their-servers side. “If the job takes a while, you can leave. Tab keeps going.”
Its homepage says “Nothing is paid until you approve it”. Its terms add that Tab “may pause and ask for clarification or approval” when a request is “ambiguous, sensitive, irreversible, unusually risky”, or differs from what you approved. We read that as asking first, as with Grok Bot, so it sits at the top.
“Tab starts in Messages”, and joining takes “Your number, then one required text”: the sign-in row. It is in private beta, from a waitlist, and publishes no price.
Eden’s privacy policy, read October 7, 2026: “Every customer runs on a dedicated, isolated instance rather than in a shared multi-tenant pool”. It adds: “Your data is hosted in the European Union”. So it sits on the their-servers side.
“Where an action would commit you to something, it asks you to approve it first.” That puts it at the top, the reading that put Norton there.
meeteden.ai offers “Chat on iMessage” and “Chat on Telegram”: the sign-in row. The same page says “eden is in closed beta”.
Lindy’s pricing page, read October 1, 2026: Plus at $29.99 per user a month, inbox management, scheduled routines and thousands of integrations. Its FAQ says actions with outside impact, including sending email, wait for approval.
Its plan documentation, read October 1, 2026: one $29.99 monthly plan including email, calendars and tasks. Its privacy documentation names email content, memory and conversation history stored on US infrastructure, including Supabase. Its FAQ says drafts wait for review and Send.
Martin’s homepage, read October 1, 2026: text, calls, email and Slack, with calendar scheduling and email drafts. We print no Martin price.
Anthropic’s pricing page, read October 1, 2026: Claude Pro at $20 billed monthly. The annual discount is a different billing choice.
Google’s US plan page, read October 1, 2026: Google AI Pro at $19.99 a month, with expanded Gemini access. Regional availability and limits vary.
Motion’s homepage, read October 1, 2026: projects, tasks, calendars, meetings, documents, notes, reports and workflows.
Reclaim’s homepage, read October 1, 2026: an AI calendar scheduling tasks, habits, focus time and meetings.
OpenAI’s pricing page, read October 1, 2026: chat and deep research, with its always-on Dot agent on Pro. Agent features differ by plan.
If you already have an agent
Where to look for each answer, on any agent.
The head-to-head
Whose computer each works on, and what it costs.
The one to read first
The same AI companies behind both. Archie puts your agent on your computer.
If you already automate
A diagram you build, or a sentence you say.
Open source
Where open code wins, and where Archie does.
Against a person
The office work from an agent, the phone and the judgment calls from a person.
Think a comparison is unfair to the other side? Tell us and we’ll fix it.
For One Person
For a Team