Your own AI · on a Mac in your building
Run your own AI: your models, your data, your rules.
An AI worker on a Mac in your building. Send it a quote request and it drafts the priced quote from your price list, the row beside every line; a person checks and sends. It also answers questions from your files and sorts the inbox. First measured job: quotes.
Got it. A person replies with two times for your free hour; if you do not hear back, call 404 493 6248. While you wait: try five real requests from our test runs, and have one job you repeat in mind (for quotes, a price list or three recent requests).
No customer has run this yet, so for quote work you test it on your own requests first, free. Open-weight models on the box; your documents stay on the box in your building unless you turn cloud on with your own key, in writing. Measured so far: quotes from a price list.
Family office or real estate firm? Stated plan: no lease or loan document has been tested yet. Your documents would stay on a box in your building, with cloud off unless you switch it on in writing; the free hour says whether this box can do the job. Ask for the hour for lease and loan questions.
Quote work? Skip the call: paste 3 requests, drafts back within 2 business days
Where it failed
Your price list is scored free before you pay, because on each of three trades it had never seen (electrical, plumbing, lumber; fictional, 55 answers each) the unscored build made wrong prices: 8, 11 and 15, and a fix tuned on two did not carry to the third. That is why every site gets its own build and score. The three-trade resultsWe'd like three of the 24-port PoE switches and eight wifi 6E access points, installed.
| 3 × SW-24P at $1,050.00 | $3,150.00 |
| setup 3 × $150.00 | $450.00 |
| 8 × AP-6E at $389.00 | $3,112.00 |
| setup 8 × $60.00 | $480.00 |
| Total (2 lines) | $7,192.00 |
- company.md
- hardware-price-list.md
- sops/user-offboarding.md
- volume-discounts.md
- sops/quote-a-hardware-line.md
Raw output
Quote action:
- 3 x SW-24P at $1,050.00 = $3,150.00, plus setup 3 x $150.00 = $450.00; line total $3,600.00
- 8 x AP-6E at $389.00 = $3,112.00, plus setup 8 x $60.00 = $480.00; line total $3,592.00
ANSWER: $7,192.00 (2 lines)Your AI, your rules
Four things you keep when the AI runs on a box in your building.
Each one is set at install and written down. Change them later through us or your IT provider; the box is yours after the first paid month.
Your models
Open-weight models run on the box through Ollama. A cloud model on your own key, for jobs you name in writing, is optional. Need a bigger model than a Mac holds? Memory decides, not CPU; we would help source the hardware (stated; none sourced yet). Nobody trains on your data.
Measured on Qwen3 8B, 16 GB Mac mini · 27B on a 24 GB GPU · bigger boxes
Your data
Price lists, procedures, notes (and, as a stated plan, leases) become a plain-text brain on the box that your staff can open and edit. It stays on your network unless you say otherwise.
LAN only by default · not used to train anyone’s model
Your rules
Sending is off. Shell, code, browser and web tools are off. What it may read is agreed in writing at install; a person approves anything that leaves the building.
Nine tool groups switched off by the installer
Your work
After the free hour, quote work goes to the free 3-request test and 30-quote score on your past quotes, run on our Mac mini before anything is installed (the test can be requested with the hour; tested on fictional firms so far; other jobs would get their own test). Every miss becomes a test.
First measured job: quotes from a price list
The software is open source (MIT). What you pay for is the box, the build from your files, the score, and the monthly re-test. How it works, in full.
Who it is for
Offices with files they trust and a job they repeat, starting with these three.
One is measured on two fictional firms. Family offices and real estate firms have first jobs named but no test yet. IT providers are the channel. The free hour is where we check the fit. Two dated write-ups of work for people close to us, neither a paying customer, are under Work.
Start where you are: I send quotes at a shop or distributor · I run an IT provider · I run a family office or a real estate investment firm · I worry about wrong answers · I already run my own AI server
Shops and distributors (and IT providers’ own quote desks)
Quotes drafted from your price list, each line citing its row. You see the score on your own price list before you pay; if it fails there, it costs you nothing. That is why each site is scored before it goes live: on three distributor trades it had never seen, an unscored build made wrong prices (8, 11 and 15; measured on a fictional IT provider and a cabinet shop otherwise).
15 of 15 held-out messy requests right · Mac mini · 6 Oct
Family offices and real estate investment firms
Your documents stay on a box in your building unless you turn on cloud with your own key; a person sends. First jobs: lease abstracts, rent-roll and loan questions, deal-memo and investor-update drafts. No lease or loan document has been tested at all. In the free hour we say whether this box can do your job.
First job named in the free hour; we would score it on your past work before it goes live. No such test yet.
IT providers
You install, patch and join the monthly review. A flat monthly price for the Mac mini, box included. Run the free 30-quote score on a client’s past quotes before you commit. Partner terms are on request.
First partner conversations under way; the first named partner is still open
The test results
15 of 15 on a small held-out set, for a fictional firm we built the brain for.
6 Oct, messy set 2 (unseen on 2 Oct); when unsure it asks instead of guessing. A wrong total that reaches a customer is the costly failure, so the test counts it separately. On 5 October, 6 drafts had a wrong total, every one on a request that also asked about an item not stocked. The 6 October fix flags that item instead, and wrong totals went to 0 on these fictional sets (not on the three unseen trades; see “Where it failed”). On 6 Oct it declined 10 requests it could have priced; a person sees each one. Step by step: how the AI quoting software works.
Try a real request
Five requests from the 6 October runs on our 16 GB Mac mini (Qwen3 8B with the notes and the quote action, fictional companies Fieldstone IT and Peachtree Cabinet Works). Request, output and seconds are copied from the results files; the grade is the test's own.
How the same steps handle a forwarded RFQ: RFQ automation from your own price list.
▸ clients
▸ policies
▾ sops
quote-a-hardware-line
route-an-email
mfa-lockout
hardware-price-list
volume-discounts
labor-rates
SOP: price a hardware line
- Find the SKU in hardware-price-list. Not there? Stop: special order for the service director.
- Hardware = quantity × unit price.
- Discount from volume-discounts, hardware only.
- Setup = quantity × setup fee (never discounted).
- Over $10,000 needs the service director (approvals).
ANSWER service_desk
ANSWER NOT IN VAULT: MON-34 is not on the price list: special or custom order, not priced here. Ask the person who owns pricing (see people).
Every answer, misses included: the Staffbox scorecard on Hugging Face, and the raw results files on GitHub for Fieldstone IT and the messy 15.
That was a made-up company. See it on your own requests. Quote work? Send 3 requests, free Call us: 404 493 6248
Where to start
It starts on one job you repeat every week, using only files you approve. Quotes are measured; answers and routing are tested only as items inside the 40-test sets.
Each answer names the file it came from.
Quote drafts
Your price list and discount rules become a priced draft, each line traceable to its row. An item you don't stock is flagged and asked about instead of substituted (by design; 0 wrong totals in 475 test answers).
15 of 15 held-out messy requests right · Mac mini · 6 Oct
Answers from your files
"What do we charge for after-hours work?" answered from your own notes, naming the note: $225 an hour, 2-hour minimum, from labor-rates.
graded right · 4.8 s · Mac mini · 6 Oct
Inbox routing
Each incoming email goes to the right queue: a jammed second-floor printer goes to the service desk, not to the person who quotes.
graded right · 2.6 s · Mac mini · 2 Oct
Works with today: your mailbox and shared files; drafts land in your drafts folder and a person sends them. Not today: writing into your ERP, CRM or estimating system; connectors are set up per site and still planned. Name your system and we say plainly whether it connects. Paperwork, reconciliation and shop-floor knowledge are next on the list: see the planned scope.
How it works
Your files go in, a sourced draft comes out, and a person sends it.
It fits the way your office already works. Nobody has to become an AI company.
Point it at what you already trust
Your price list, procedures and a quotes mailbox, agreed in writing at install. They become a plain-Markdown brain on the box that your staff can open and edit.
It prepares the next move
A priced quote, an answer or a routed email lands in your drafts folder, usually within minutes, with the price-list row and its date beside each line.
A person checks and sends
Someone on your team edits if needed and sends. Sending is off in the agent; at install we state in writing which limits your mail platform also enforces.
Inside the box, the four open layers, and how every miss becomes a test: how Staffbox works, in full.
Safe by default
It cannot send, and your data stays in the building unless you say otherwise.
For a small company, trust is practical: the data stays in the building, every number has a source you recognise, and a real person has the last word.
- Drafts onlySending is off in the agent; at install we state in writing which limits your mail platform also enforces. Delete rights are off at install.
- Every priced line shows its sourceThe price-list row and that list's date, beside the number.
- Not on your list? It asks.An unstocked item is flagged instead of priced as the nearest thing you stock.
- Cloud AI is offOn only for jobs you choose, if you ask in writing, on your own key.
- Risky tools are offThe installer turns off nine tool groups plus image generation: terminal, code execution, browser, computer use, web, connections, delegation, cron jobs and text-to-speech.
- Locked down on your networkModels answer on 127.0.0.1 only. Firewall, stealth mode and key-only SSH are set at install by your IT provider from our checklist (not yet automated in the installer); verified on our demo unit, 1 Oct.
- Nobody trains on your dataNot us, not a model maker. The brain is plain Markdown and always yours.
Approval rule, pilot agreement and what it can touch
Approval rule: by agreement, the worker drafts and a person at your company approves anything that goes to a customer, supplier or regulator; nothing is to be deleted or sent on its own. The default install turns off nine tool groups plus image generation, including the shell, web access and outside connections; replies go out only as drafts through the mailbox door. Sending is off in the agent; at install we state in writing which limits your mail platform also enforces. The brain is plain Markdown, so you can keep it under version control. A site can turn tools on, in writing.
In the pilot agreement: data stays on site; cloud models stay off unless you ask in writing with your own key; no training on your data for anyone else; every remote session logged; our copies deleted within 14 days of exit.
What it can touch: what the worker can read and change is agreed per site, in writing, before install. By agreement it drafts and a person approves anything sent; every remote session is logged. The access spec is not published yet; ask for the current draft.
Want these rules written down for your own office? Request my free hour Call us: 404 493 6248
How we work with you
For quote work: from hello to a scored pilot in 30 days; it goes live only if it passes. Other jobs are scoped in the free hour.
A person on the Staffbox team replies, with the price in the first message and two call times to choose from.
You hear from a person
You get a reply from a real person, with the price in it and two call times to pick from.
A free hour; for quote work, also three requests and 30 quotes, free
In the free hour we look at one job you repeat and say what would run on your own box, what it costs, and whether it fits. For quote work, send three recurring requests and we send back what Staffbox would have drafted, misses included. Then we score 30 of your past quotes on our Mac mini, free; your team grades 10 answers blind; we delete the files within 14 days of sending your results. For non-quote jobs there is no test yet; the hour says whether the box can do the job (per our draft data policy).
Install and baseline
If the score passes, you sign a short pilot agreement. The box goes on your network, and you see the day-0 score on it before anything goes live.
Fix and re-run
Each miss is fixed and the test runs again, until a pass comes back clean or we tell you what it can't do yet.
You decide
With the day-30 score in hand, continue at the founding rate or stop. If you stop, we collect the box and delete our copies.
The box comes with the monthly fee, and after the first paid month the box and the brain are yours to keep.
The founding-site pilot costs nothing and has no onboarding fee.
It stays on site and is not used to train, tune or evaluate models for anyone else. If you leave, we delete our copies within 14 days.
Price · draft
$595 a month at founding sites (first 100) for as long as the box is live (draft pricing), box included, re-tested on your quotes monthly.
After the free hour and, for quote work, a passing score you start with a free 30-day pilot, then pay $595 a month at founding sites (first 100) for as long as the box is live. From unit 101 the same service is $695 a month plus $1,800 onboarding. What the AI quoting software does for that price, with a real request and its draft.
$595 / month while the box is live, first 100 sites
No onboarding fee. Hardware, install, vault seeding, the workflows that pass their score and the monthly review, at the founding rate for as long as the box is live. From unit 101, onboarding is $1,800.
$695 / month + $1,800 onboarding
Box included, and yours after the first paid month. A monthly re-test on your own quotes, with updates applied by your IT provider after the tests pass. One flat monthly price, no per-seat bill. Cancel any time.
| Compared with | Monthly | Notes |
|---|---|---|
| Staffbox | $595 founding · $695 standard | Measured 2 Oct 2026 on the 16GB Mac mini, 8B model kept warm, one request at a time: averages of 2.4, 2.9, 1.2 and 1.1 seconds on the four test sets, and 90% of answers within 5.3, 5.2, 2.4 and 1.7 seconds. Concurrent requests queue. measured Good for drafted quotes and answers; with several people asking at once, expect a wait. opinion Power: Apple rates the M4 Mac mini at 4 W idle and 65 W at most (Apple), and nothing runs in a cloud data center unless you turn it on. fact Team sizing per box is published once more units are measured. planned Sites whose test calls for a Mac Studio, or for enterprise hardware sized by memory for a larger model, are quoted per site; none sourced yet. stated |
Scroll sideways for more columns
Run IT for small businesses? Read how IT providers install and support Staffbox for their clients. Partner terms for MSPs are on request. Ask partners@staffbox.ai for the current draft.
Not sure it pays for itself? Bring the numbers to the free hour. Request my free hour Call us: 404 493 6248
Why now
Harper Carroll says the days of “tokenmaxxing” might be over.
Her caption: token-heavy agent loops are the most talked-about way to use AI, and she has published guides on fine-tuning open models for systems that use AI selectively. her view, as reported
Staffbox takes a different route, with no fine-tuning: a small open model on a Mac in your building, a brain made from your files, and code (not the model) doing the arithmetic. On the fictional Fieldstone set (Dell test box, 30 Sep), the model alone got 1 of 40 quote requests right; with the notes and the quote action it got 40 of 40, on a set we built the brain for (see “Where it failed” above). measured
Staffbox’s position: a flat monthly fee (draft price; onboarding applies from unit 101), and no per-seat or per-token bill for local work. The price · The test results
If you are comparing
Keep your office suite’s assistant; Staffbox does the work outside it.
An office-suite assistant seat
- Works across your suite: mail, files, shared drives and chat
- Billed per user, per month
- Runs on the models the suite vendor offers, in its cloud
- Good at email, documents and meetings inside the suite. Keep it.
Staffbox, for the work outside the suite
- Built for the shared drives and SOPs that sit outside the suite; line-of-business connectors per site are planned
- One flat monthly fee per box, no per-seat or per-token bill
- Your choice of model: local, your cloud provider on your key, or your own
- Each workflow is set to go live only after a score on your own past requests (quote work today; other jobs would get their own test), with thinking and memory on your network by default
Our read of each category, not a benchmark. Your IT provider can run both side by side. opinion
| Option | What it gets right | Where Staffbox differs |
|---|---|---|
| Do it yourself: a Mac mini or a home server with a one-click agent app | Same open-source software, no subscription. If someone on your team will set it up, test it on your work and keep it patched, do that. | The software is free either way. The monthly fee pays for the vault built from your business, the test on your own past requests before it goes live, and the monthly re-test. opinion |
| AI appliances and deskside AI workstations | More compute and bigger models, with the hardware maker's support. | Usually bought as hardware plus services, for teams with their own IT. In our tests, price-list quotes scored 40/40 and 39/40 on an 8B local model for two fictional firms, and on three trades it had not seen it made wrong prices, so each site is scored on its own quotes; when a task needs more, it moves to a Mac Studio (planned, not yet measured), enterprise hardware sized by memory (stated, none sourced yet) or your own cloud key. One monthly fee, Mac mini included and larger hardware quoted per site, installed by the MSP you already use (planned; none named yet). opinion |
| Boxes sold with ready-made virtual employees | A receptionist or sales assistant on day one; some answer the phone. | Staffbox works on email and files, not calls. It starts with workflows built from your own requests, each scored before it goes live. stated Until there is a customer to show you, the free hour says whether it fits, and the free 3-request test shows it on your own quotes. |
| Cloud agent platforms | Live in minutes, no hardware on site. | Per-seat or per-credit bills that grow with use, and your data leaves the building. Staffbox is a flat fee and local by default. opinion |
| An office-suite assistant through your IT provider | Works inside the suite you already pay for. | A per-seat licence answering from what is in the suite. Staffbox works with the shared drives and SOPs that never went into it, with line-of-business connections planned per site, and keeps the model and your files on your LAN. Different ground; your IT provider can run both. opinion |
| Guest-messaging AI for hotels | Talks to guests directly: booking questions, requests, complaints. | Staffbox has no guest-facing channel yet. It works on email and files (vendor mail, staff SOPs today; reconciliation planned) alongside a guest-messaging tool, not instead of one. opinion |
| AI built into your ERP or hotel system | Reads and writes that system directly. Use it. | It covers one system. Staffbox works across the inbox, shared files and SOPs, alongside it. opinion |
Looking at an agent desk such as Perplexity Computer? Read how Staffbox compares, with sources.
Compare it on your own work, not ours. Request my free hour Call us: 404 493 6248
Team
Six people, staff and contractors, across product, sales, engineering, design, marketing and research, build and test it; your IT provider installs it.
The Staffbox team
Six people, staff and contractors, across product, sales, engineering, design, marketing and research. From the first call to a scored go-live; your IT provider installs on site. How we test and run it is written up in our guides to running AI on a quote desk.
Built by an operator: the founder started PeachDish (an Atlanta meal-kit company) and is a general partner in a European hotel platform. We run our own test unit. Who does what.
What you get from us
- A reply with the price in it.
- Two call times to pick from.
- Your day-0 and day-30 scores, misses included.
- A straight "it can't do that yet" when that is the answer.
Zero
The product, running on our own desk
The Staffbox that Staffbox runs for itself, on a 16GB Mac mini in Atlanta. It reads the company vault, drafts, logs its own work and reports its own limits. It goes through the same test-and-fix loop we run for customers (see its results). Every Mac mini result on this page was run on Zero, our own Mac mini; other rows name their machine.
Your IT provider
Install · support · monthly review
Staffbox is sold and supported through managed IT providers (first partner conversations under way). The plan: the MSP that already runs your network keeps the box patched and joins the monthly review. We are in early talks with IT providers in Atlanta; none is named yet.
Questions buyers ask
You can stop at day 30, and your data stays on site unless you say otherwise.
What happens in the free hour?
You bring one job your team repeats (quotes; a shared inbox; lease questions, a stated plan). We tell you what would run on a box in your building, what it would cost, and whether it is a fit. If it is not a fit, we say so in the hour. No card, no commitment.
We are a family office or a real estate investment firm, not a shop. Does this apply?
Yes, as a stated plan rather than a measured result: we have not installed at one yet, and no lease or loan document has been tested at all. The first jobs we would scope are lease abstracts, rent-roll and loan questions answered from your own files, and first drafts of deal memos and investor updates. Your documents stay on the box in your building unless you turn on cloud with your own key, and a person sends anything that goes out. Bring one of those jobs to the free hour.
What does it cost after the pilot?
$595 a month at founding sites (first 100) for as long as the box is live. That covers the box, install, vault seeding, the workflows that pass their score and a monthly re-test on your own quotes. No per-seat or per-token bill for local work. From unit 101 it is $695 a month plus $1,800 onboarding.
What if it gets something wrong?
It will, sometimes. That is why it drafts and a person approves anything that leaves the building, and why every miss becomes a test case that is re-run before the next change goes live. Our own misses are on this page.
Does our data leave the building?
Not by default. The model and the memory run on the box on your network. Cloud models stay off unless you ask for them in writing and use your own account. Your data is not used to train, tune or evaluate models for anyone else.
What happens if we cancel?
If you stop at the end of the free pilot, we collect the box and can wipe it in front of you. Once you have paid for the first month, the box and the vault are yours to keep. Either way we delete our copies within 14 days and confirm it in writing.
Do we need an IT provider?
It works best with one: Staffbox is sold and supported through managed IT providers, who patch it like the rest of your network. Here is what an IT provider does for a Staffbox install. If you do not have one, say so. We are in early talks with IT providers in Atlanta; none is named yet.
How much of our time does the pilot take?
For quote work: one person who knows your quotes (to grade 10 answers and check drafts), plus the files for the first jobs (price lists, procedures, an inventory export). Your IT provider handles the shelf and the network port.
Can we use cloud AI such as Claude or Gemini?
Yes, for the jobs you choose. Ask in writing, add your own key for Anthropic, Google Gemini, OpenRouter, Azure, Salad, Baseten, Crusoe or another supported provider, and we point those jobs at it. You pay the provider directly at their rates; Staffbox does not bill for tokens. With cloud on, that job's data goes to your provider under their terms; everything else stays local. Scoring cloud models on your test is planned.
Can we use our own model?
Yes. Use any open-weight model Ollama supports, a model your team already has, or your company's private cloud deployment. Fine-tuning on your past requests is experimental and not part of a pilot. Nobody else trains on your data or gets a copy.
What if one box is not enough, or we want a bigger model?
Start with the box your test says you need. Mac mini today; Mac Studio planned, not yet measured. Memory, not CPU, decides what model fits: an 8B model runs on our 16 GB Mac mini, gpt-oss-20b needs 32 GB, and the 27B model we measured ran on a test box with a 24 GB GPU. If your job calls for more than a Mac can hold, we help you source enterprise-grade hardware (a workstation or server) sized by memory, quoted per site. stated No such sourcing has been done yet.
Does it work with our software (ERP, CRM, hotel or estimating system)?
Files, shared folders and email today; shell, code, browser and web tools are off by default. Connections to a specific system are set up per site and are still planned. Name your system when you send the free 3-request test, and we will tell you plainly whether it connects today, later, or not at all.
Can we see it work before we sign?
Yes. Send three requests your team answers every week and we send back what Staffbox would have drafted, misses included, free. There is no customer install to visit yet, so the test on your own requests is what we can show today.
What can't it do yet?
It works on email and files: no calls and no guest-facing chat. There is no certification such as SOC 2 and no customer install to show you yet. Those lines on this page change when they do.
Still unsure? Ask a person. Call 404 493 6248 Or request a free hour
Request my free hour.
Two required fields. A person replies with two times to pick from. In the hour: what would run on your box, what it costs, or a plain “not a fit”. No card, nothing installed at your site.
