You run a small business, or you are the business. You have read about AI agents, maybe watched a Grok Bot demo where a bot signs into a CRM and works through the night, and one part of you thought "I want that for my invoices" while the other part thought "I am not handing a cloud company my inbox, my Drive and my client files." Both reactions are reasonable. This post is for the second one.
I built SAMBOT because I had the same split. It is a Windows 11 desktop app that runs a team of AI workers on your own PC, on whatever model you already pay for or a free local one, and it costs $24.99 once after a launch window that is free until September 25, 2026, 00:00 IST. Below is a straight comparison with Grok Bot, three use cases written the way they play out on screen, and the parts I would want to know before installing anything built by one person.
Two weeks ago I published Grok Bot: what xAI's own docs admit that the launch posts skip, and nothing here contradicts it. Grok Bot is a real product doing real work for people. The question this post answers is where you want the work to happen.
The comparison
| Grok Bot | SAMBOT | |
|---|---|---|
| Where it runs | xAI's cloud, one persistent Linux computer per account, shared by all your bots | Your Windows 11 PC |
| Who owns the model | xAI, running Grok | You: an OpenAI, Anthropic, Gemini, Groq or OpenRouter key, or a free local model via Ollama or LM Studio |
| Price model | Bundled with SuperGrok and Cursor subscriptions, weekly usage allowance, overage billed at token cost | Free until Sep 25 2026 00:00 IST, then $24.99 one-time, no subscription |
| Where your data lives | On the cloud computer, including browser sessions and logins | On your disk; keys and files never leave the machine |
| What you need to start | A paid SuperGrok or Cursor plan | Windows 11 plus an API key or a local model |
| Who it is for | People who want always-on agents in the cloud and already pay for the bundle | Operators who want the team, the files and the cost on their own box |
The first difference is the computer. Grok Bot gives every account one cloud machine that all your bots share, with a browser signed in as you, which is what lets it drive any web app without an API. SAMBOT has no cloud machine. The workers run on your PC, read and write files in folders you choose, and reach Gmail, Drive, Calendar, Sheets and LinkedIn through an OAuth app you create in your own account. There is no server of mine in the loop.
The second is the meter. Grok Bot is included in a subscription with a weekly allowance, and past it the overage bills at token cost. SAMBOT has no credits and no meter of its own. Use a cloud key and you pay that provider for exactly the tokens you use, under their spend caps. Use Ollama or LM Studio on your own CPU and you pay nothing per run and wait a few minutes for each team run instead.
The third is the model. Grok Bot runs Grok. SAMBOT runs whatever you point it at, so the team you build today can run on a small local model for rough drafts and a frontier key for the run that matters, and you can change providers next month without rebuilding anything.
The fourth is what you can see. Every message, tool call and file the team touches shows up in a live activity feed while it runs, and any sensitive action, sending mail, posting, deleting files, running code, stops and waits for you to click approve. That is not a claim about Grok Bot's safety model, which the earlier post covers. It is what will be on your screen.
Use case one: ten call notes into a weekly summary and six follow-ups
Priya is a solo operations consultant with six retainer clients. Every call ends with notes dumped into a Drive folder called Call Notes, one file per call, and every Friday she owes each client a recap and a follow-up email. Ten calls this week, written up by hand at 9 pm on Fridays for two years.
In the Ask SAMBOT chat she types: "I need a team that reads my call notes for the week, summarises them per client, and drafts a follow-up email for each client that I approve before it goes out." SAMBOT proposes three workers: a Reader that pulls every file from the Drive folder, a Summariser that writes a weekly summary grouped by client, and a Drafter that turns each client's section into an email. She accepts, and the three land on the canvas already connected, Reader to Summariser to Drafter, with her Drive and Gmail connectors attached to the workers that need them.
Then the objective: "Read every file in Drive folder Call Notes / Week 38. Write weekly-summary.md grouped by client, with dates, commitments I made, and open questions. Then draft one follow-up email per client and send each one from my Gmail."
The activity feed starts scrolling. Reader lists the folder, finds ten files, opens each one and hands the text down the line. Summariser writes weekly-summary.md and the feed shows the file being created. Drafter starts composing, and the run pauses. A box appears: Drafter wants to send an email to a client via Gmail, subject and body shown in full, approve or reject. She reads it and approves. The next one appears. She approves four more and rejects one because the tone is wrong for that client, and rewrites it herself from the draft. The deliverable opens: weekly-summary.md, six headings, one per client, the commitments she made on each call listed under it, in a folder on her own disk next to the notes it came from.
With an Anthropic or OpenAI key this run takes a couple of minutes. On a CPU-only local model, plan for longer. Either way, no email left without her click.
Use case two: the Monday competitor price check
Dev runs a small online store with about forty products. He keeps a Google Sheet called Competitor Watch: one row per product, his price, and two competitor URLs. Checking it by hand eats a Monday morning, so it happens about once a month.
His team is three workers he described in plain English: a Researcher that "reads each row of a sheet and fetches the current price at each competitor URL," an Analyst that "compares competitor prices against ours and flags anything more than 10 percent off," and a Writer that "turns the analysis into a short report with a table." He drew the dependencies himself: Analyst waits for Researcher, Writer waits for Analyst.
Monday, he types: "Open the sheet Competitor Watch. For every row, fetch the current price at both competitor URLs. Compare each with our price. Write price-check-2026-09-21.md with a full table and a short list of the products where we are more than 10 percent above or below the cheapest competitor."
The feed shows Researcher reading the sheet through the Sheets connector, then working through the URLs. Most pages come back with a price. A few do not, and the feed shows the Researcher noting "no readable price on this page" and moving on rather than guessing. Analyst runs the comparison. Writer builds the report. Nothing pops for approval this time, because reading a sheet, fetching pages and writing a file on his own disk are not sensitive actions. Had he asked the team to email the report to his supplier, the send would have stopped for a click.
The deliverable is one markdown file: a table with product, his price, both competitor prices and the gap, a list of seven products worth a look, and at the bottom the three URLs the Researcher could not read, which he checks by hand. Next Monday he types the same objective with a new date. This is the run where I would use a cloud key without hesitation: eighty page fetches and a lot of extracted text is the workload that makes a CPU-only local model crawl.
Use case three: one job post, one resume, one honest application
Maya is a freelance content designer between contracts. Her resume lives on her PC as resume.md. A LinkedIn post for a six-month contract has caught her eye, and she wants an application that maps to the role, not a generic cover letter.
Her team: a Job Reader with the LinkedIn connector attached, a Matcher that "compares job requirements against my resume file and lists what fits, what is missing and what to emphasise, without inventing anything," and a Writer that "drafts a cover letter under 300 words in my voice using only facts from the resume."
The objective: "Read the LinkedIn job post at this link. Compare it against C:\Users\maya\resume.md and write fit-notes.md. Then draft application.md as a cover letter under 300 words. When both files are written, email the text of both to me from my Gmail so I have them on my phone."
The feed shows Job Reader pulling the post text, Matcher opening the resume, and the two outputs meeting in a list of matches and gaps. Writer drafts. Then the pause: the Writer wants to send an email to Maya's own address with both drafts pasted in, approve or reject. She approves. Two files open on her disk. fit-notes.md says plainly that the post asks for three years of design-system work and her resume shows one, and suggests leading with the accessibility audit she ran last year instead. application.md is 280 words and every claim in it is something she can back up.
This one is small, two short files and one page read, and a local model through Ollama handles it in a few minutes. If you have no cloud key and want to try SAMBOT for free end to end, start with a job this shape.
What it will not do
I would rather you read this here than find it after installing.
SAMBOT is Windows 11 only. No Mac build, no phone app. Grok Bot ships desktop, iOS and Android apps, and if you need to reach your agents from a phone, that is a real point in its favour.
The installer is unsigned, because a code-signing certificate is a real cost for a one-person shop. Windows SmartScreen will warn you. Click More info, then Run anyway.
Local models on a CPU are slow. A team run that takes two minutes on a cloud key can take a good while on Ollama with no GPU. It works, it costs nothing per run, and you wait.
It is built and supported by one person, me. Updates ship when I ship them. If you want a vendor with a support desk and a status page, that is a fair reason to pick the cloud product. All sales are final.
The window
SAMBOT is free until September 25, 2026, 00:00 IST. After that it is $24.99 one-time, no subscription. An account created during the free window keeps a free plan of up to 5 agents for good, and Pro removes that limit and includes every 0.x update.
The full demo is 3 minutes 10 seconds of the real app, connecting Ollama, creating an agent by chatting, building a three-agent team, running an objective with an approval, and opening the deliverable. It is embedded on the download page.
Download SAMBOT at clskillshub.com/sambot. Enter your email and the link lands in your inbox. If you are still deciding whether agents belong in your work at all, the free 75-page Claude guide covers the prompt structure that makes any of these tools safer to hand real work to.