How it works
The box, the brain, and the test it has to pass every month.
A Mac mini in your building drafts from your notes and goes live only if it passes a test on 30 of your past quotes.
Your team answers the same requests every week from the same files. A tool you cannot inspect, test or take back is hard to trust with them. Staffbox is open parts on one box (an MIT-licensed agent, open-weight models, a plain-text brain), scored on your past work before it goes live and every month after. Qwen3 8B ships by default; buyers who cannot use a Chinese-made model get gpt-oss-20b (OpenAI, open weights), which needs a 32 GB box. Our numbers come from fictional test companies; no customer has run it yet. Send up to three requests; drafts come back within 2 business days.
Agent, memory and reach
What is in the box
Four open layers sit on one Mac, and you can replace any of them.
Agent, models and memory format are open source or open weight. Staffbox is sold and supported through managed IT providers.
Mac mini today; Mac Studio planned
Silent, small, sits on a shelf. Mac mini today; Mac Studio planned, not yet measured. A Studio is for workflows whose test shows they need larger models or many people asking at once. In our view Apple silicon is the cheapest reliable way to run local models at this size today.
Hermes Agent
Open-source agent framework from Nous Research, MIT licensed. Builds reusable skills from tasks it completes, keeps memory across sessions, connects to tools over MCP. The same software we run on our own test unit. Our Hermes Agent guides explain how we set it up for a business.
Local first, cloud when you choose
Open-weight models run on the box through Ollama, so routine work has no per-token bill. You can also run a model your company already has, or send chosen jobs to a cloud AI provider such as Anthropic, Google Gemini, OpenRouter, Azure, Salad, Baseten or Crusoe on your own account and key. The default is local.
Obsidian brain
A plain-text vault of your procedures, price lists, customer notes and decisions. Human-readable, ready for version control, yours. Opens in Obsidian or any Markdown editor. The agent reads it before it acts and appends to a log afterwards; it never rewrites what a person wrote.
Local by default
Inference and memory stay on the LAN. Connected tools like email and your CRM behave as they do today. Cloud AI providers stay off until you ask for them in writing, on your own key.
Readable memory
The vault is plain Markdown. Open it, edit it, delete it. No hidden embeddings store.
Supported through your MSP
Staffbox is sold and supported through managed IT providers, so the box can be patched and monitored like the rest of your fleet, on terms agreed between you and them.
No lock-in
MIT-licensed agent, open-weight models, plain-text memory. Every part can be replaced or taken in-house.
Intended scope for the first installs
First installs take one repeat workflow at a time, and only workflows that pass their score run.
Staffbox takes the repeat work off your people, one workflow at a time, so they spend the day on customers and judgement calls. It drafts; they decide.
Answers and intake
Read the shared inbox, draft replies from your vault, log the request in the CRM, flag what needs a human. Quote from your price list, in your voice.
Paperwork and reconciliation (planned)
Match POs to invoices, chase missing certs, prepare month-end packs, turn meeting notes into tasks. Work inside the Docker services your MSP already runs.
Ops knowledge
Answer "how do we do X here" from your SOPs, draft work instructions, watch a folder for drawings and file them. Manufacturing first, hospitality right behind it: at a hotel the same three jobs are front-desk intake, back-office reconciliation and staff SOP answers.
Staff talk to it in chat, email or the terminal. Teams already using cloud agents will be able to point them at Staffbox as a tool (planned). Small teams with none can start without a long procurement cycle; larger buyers follow their own process, and we fit it. Public offices: read how a public office can evaluate and buy Staffbox. Each install starts with the workflows built with your team that pass their score; nothing else runs on day one. Unvalidated outside our own unit so far.
Start small, grow by the test
The test decides the hardware: the Mac mini is measured, a Mac Studio is planned, and cloud is opt-in.
The software is the same on every box. Your scorecard decides what you need: we move a workflow to a bigger box or, if you allow it, a cloud model only when the test shows the smaller one can't pass it. What the 16 GB box runs, how fast and on how much power: local AI on a Mac mini for a small business.
One site, the first workflows
Small local models. Our box: 40 of 40 on the Fieldstone test with the quote action. Need a model that is not from a Chinese vendor? OpenAI's open-weight gpt-oss-20b scored 39 and 40 of 40, and 15 of 15 messy requests, with the same notes and quote action on our test server; it needs a 32 GB box. Bigger open models without buying a Mac Studio: US-made open-weight models such as OpenAI's gpt-oss-120b can run on Baseten on your own account and key, for the jobs you choose. Every model option and its score. On three trades it had never seen (14 and 15 Oct, Dell test box), the default Qwen3 8B setup made 8, 11 and 15 wrong prices in 55 answers; that is why we build and score on your own list first (the electrical result).
Bigger models, more people
Larger local models and several departments on one box. Not yet measured on a Studio. A 27B-class model scored 40 and 39 of 40 with the quote action on our Dell RTX 3090 test box.
Your key, for the jobs you choose
Anthropic, Google Gemini, OpenRouter, Azure, Salad and others, on your own account. You pay the provider; Staffbox does not bill for tokens. Off unless you ask in writing; with it on, that job's data goes to your provider under their terms.
Bring one you already have.
Run any open-weight model Ollama supports, a model your team already has, or your company's private deployment in Azure or another cloud.
Fine-tuning on your past requests is experimental and not part of a pilot. Nobody else trains on your data or gets a copy.
A common mix, in our view: local for the daily repeat work, a cloud model for the few jobs your test shows the box can't do yet. Hardware tiers and Apple list prices: docs/hardware.md.
How it gets better
Every miss gets a cause and a fix: your notes plus the quote action took the same 8B model from 1 to 40 of 40.
Each pilot is set up to be tested on your own past work before it goes live, and again every month. Each miss gets a named cause and a fix. The score is yours to keep.
- 01 · Your test30 past requestsyou already answered, with the right answer
- 02 · RunStaffbox answerson the box, from your vault
- 03 · GradeGrade every missscore it, name the cause of each miss
- 04 · Fix, cheapest firstvault → skill → tool → bigger model or boxonly if the rest fail; then re-run the test until a pass comes back clean
- 05 · LiveClean pass → go livea person approves
- 06 · Every monthStaff corrections become new test casesthe test gets harder, and the loop starts again at 01
Add it to the vault. Minutes, and you can read the change.
Edit the skill or the instructions for that workflow.
Connect the tool: inbox, ERP, a shared folder.
Move that task to a larger local model, on a Mac Studio if the box needs to grow (planned, not yet measured), or to cloud burst if you allow it.
We run the loop on ourselves first
Zero is our own unit. The same model, first alone, then with the company notes, then with the notes plus the quote action. Fieldstone IT, a fictional IT provider, 40 tasks, on our GPU test box.
| Model | Alone | With notes | Notes plus quote action |
|---|---|---|---|
| Qwen3 8B | 1 of 40 | 33 of 40 | 40 of 40 |
| Qwen3 27B | 2 of 40 | 39 of 40 | 40 of 40 |
Scroll sideways for more columns
The notes do most of the lift; the quote action does the prices in code and closes the rest. Dell RTX 3090, 30 September 2026: results on GitHub.
On the Mac mini itself, 6 October
The 16 GB Mac mini we install, Qwen3 8B, your notes plus the quote action. Two fictional companies, 40 tasks each, plus two sets of messy quote requests (typos, forwarded threads, two items in one email, "no setup"): 15 it had never seen, and 20 we also used while fixing it. The last two rows are a generated regression set (requests written by a model, answers checked by code).
| Test set, 6 Oct | Right | Average answer | 90% within |
|---|---|---|---|
| Messy requests it had never seen | 15 of 15 | 1.1 s | 1.7 s |
| Fieldstone IT (an IT provider) | 40 of 40 | 2.3 s | 5.3 s |
| Peachtree Cabinet Works | 39 of 40 | 2.5 s | 4.4 s |
| Messy requests, also used while fixing | 19 of 20 | 1.2 s | 2.3 s |
| Generated regression set, Fieldstone | 201 of 204 | 1.6 s | 2.1 s |
| Generated regression set, Peachtree | 150 of 156 | 3.2 s | 3.4 s |
Scroll sideways for more columns
One question at a time with the model kept warm; concurrent requests queue. Of the 11 misses, 10 declined and asked the customer and 1 sent a request to the wrong queue. The 5 Oct version made 6 wrong totals on the generated set, every one on a request that also asked about an item not stocked; the 6 Oct version (after the fix) flags unstocked items instead of substituting (by design; 0 wrong totals in 475 test answers). A person checks every draft.
How we measured: 6 Oct 2026, our own 16 GB Mac mini M4, fictional companies, graded by machine with no human edits. Every graded answer is on Hugging Face.
What a founding site's first month looks like
Score before you pay
Send 30 past quotes and your price list. We score them on our Mac mini, free; your team grades 10 answers blind; we delete the files within 14 days of sending your results (per our draft data policy).
Day-0 score on site
The box goes on your network, with your IT provider where you have one, and Staffbox runs the same test on it. You see the number before anything goes live.
Fix and re-run
Grade, fix, re-run, until a pass comes back clean or we tell you plainly what it can't do yet. This per-site build of the notes and rules is the step that matters: out of the box the worker made wrong prices on three trades it had never seen (the three-trade results).
Live, then monthly
A person approves go-live. A monthly re-test runs the test set again before any update goes live.
Start here · free, before you pay
Send 3 requests free. Then get the free 30-quote score.
First, three requests your team answers every week: we send back what Staffbox would have drafted, misses included. Then 30 past quotes and your price list: we run them on our Mac mini in Atlanta, your team grades 10 answers blind, and we delete the files within 14 days of sending your results (per our draft data policy). Nothing is installed at your site for either step.
Or call us: 404 493 6248