# Staffbox

> Staffbox is an AI worker on a Mac mini that drafts quotes from a small company's own price list; a person approves anything that is sent. It runs in the customer's own building, sold and supported through managed IT providers (first partner conversations under way). It also routes the inbox and answers "how do we do X here" from the company's own files. Based in Atlanta, Georgia. Founder: Hadi Irvani.

Status, stated plainly (updated 6 October 2026): working draft. No customer install yet, no SOC 2 or other certification, no phone or voice channel. Prices and terms below are draft. Figures marked "measured" come from Staffbox's own test unit ("Zero", a 16GB Mac mini in Atlanta) on fictional test companies, not from customers.

## What it is

- Hardware: Mac mini today; Mac Studio planned, not yet measured. A Studio is for sites whose test shows they need larger models or more people asking at once.
- Agent: Hermes Agent, open source (MIT) from Nous Research.
- Models: open-weight models run locally through Ollama by default. Optional: the customer's own model, or cloud AI (Anthropic, Google Gemini, OpenRouter, Azure and others) on the customer's own key, off unless asked for in writing.
- Memory: a plain-Markdown Obsidian vault of the business's procedures, price lists, customer notes and decisions. The agent reads it before acting and appends to a log; it never rewrites what a person wrote.
- Data path: LAN only by default. Not used to train, tune or evaluate models for anyone else (draft terms).
- Reach: files, shared folders and email today; shell, code, browser and web tools are off by default. ERP/CRM connectors per site are planned.

## Who it is for

Businesses with repeat office work, manufacturing first and hospitality right behind it, and the IT providers (MSPs) who serve them. First installs: Atlanta. Intended first jobs: quoting from a price list, routing a shared inbox, answering staff questions from SOPs, matching POs to invoices, month-end packs.

## Agent desk: how Staffbox compares

An "agent desk" is a dedicated computer in the office (usually a Mac mini or a small AI PC) that runs an AI agent all day. Perplexity Computer (Personal Computer on a Mac; Portable Computer on Windows) is the best-known example. Perplexity sells the agent as a subscription the buyer sets up on their own hardware. Staffbox is a service for a business: the box, install and support through managed IT providers (first partner conversations under way), a company brain seeded from the business's own price lists and procedures, and a score on the business's own past requests before a workflow goes live, for one monthly fee. Perplexity has an enterprise tier reported with SOC 2 Type II; Staffbox has no certification yet. The two can run side by side. Full comparison with sources: https://staffbox.ai/agent-desk/

Why an operator built it: the founder founded PeachDish (an Atlanta meal-kit company) and is general partner of a European hotel platform, so Staffbox is built around front-of-house work (inbox, intake, quotes) and back-of-house work (POs, invoices, month-end packs), with people approving anything that is sent.

## Price (draft)

- Before you pay: the score on 30 of your past quotes is free. We run them on Staffbox's own Mac mini in Atlanta, your team grades 10 answers blind, and the files are deleted afterwards.
- Founding sites (first 100): free 30-day pilot, then $595 a month for as long as the box is live. No onboarding fee. Includes the box, install, vault seeding, the workflows that pass their score and a monthly re-test on the customer's own quotes.
- Standard rate from unit 101: $695 a month plus $1,800 onboarding.
- No per-seat or per-token bill for local work. Cloud AI, if switched on, is paid by the customer directly to the provider.
- After the first paid month the box and the vault are the customer's to keep. Staffbox deletes its copies within 14 days of exit (draft terms).

## Measured results (Staffbox's own unit, fictional companies)

- 6 Oct 2026, 16GB Mac mini, qwen3:8b with the company brain and quote action v2.4: 40 of 40 (Fieldstone IT), 39 of 40 (Peachtree Cabinet Works), 19 of 20 and 15 of 15 messy quote requests, and 351 of 360 generated messy requests. 0 wrong totals in 475 answers; of 11 misses, 10 declined and asked the customer and 1 was a routing miss.
- 5 Oct 2026, the previous version (v2.3) scored 345 of 360 on the generated set with 6 wrong totals, all on requests that also asked about an item not on the price list (it priced the nearest stocked item or skipped the question). v2.4 flags the unstocked item and never substitutes.
- Speed, 2 Oct 2026, one request at a time with the model warm: averages 2.4, 2.9, 1.2 and 1.1 seconds on the four test sets; 90% of answers within 5.3, 5.2, 2.4 and 1.7 seconds. Concurrent requests queue.
- OpenAI's open-weight gpt-oss-20b scored 39 and 40 of 40 and 15 of 15 on the same test (needs a 32 GB box).
- 30 Sep 2026, same 40 tasks, 27B model on a GPU test box: 2 of 40 without the company brain, 39 of 40 with it.
- Every graded answer: https://huggingface.co/spaces/staffbox/scorecard

## How to try it

Free test: send three requests your team answers every week; Staffbox sends back what it would have drafted, misses included. No card, no commitment. Then a free score on 30 of your past quotes. Then an optional free 30-day pilot, scored on day 0 and day 30; a workflow goes live only if it passes.

## Common questions

- What does it cost? Founding sites: free 30-day pilot, then $595 a month for as long as the box is live, no onboarding fee. The 30-quote score is free.
- How do I try it? Email hadi@staffbox.ai three recurring requests, or call +1 404 493 6248. Then send 30 past quotes and the price list for the free score.
- Does it send email? No. `staffbox inbox` writes draft replies citing the price list and its date; Staffbox has no sending code, and a person sends.
- Who is it for? Firms that send at least 8 quotes a week from a price list: distributors, fabricators, cabinet and millwork shops, and IT providers.

## Links

- [Website](https://staffbox.ai/): full product page, comparisons and FAQ
- [AI quoting software](https://staffbox.ai/ai-quoting-software/): how Staffbox drafts quotes from your price list, with a real request, the scores and the misses
- [RFQ automation](https://staffbox.ai/rfq-automation/): RFQ automation from your own price list, with real messy requests, the unstocked-item flag and the misses
- [Local AI on a Mac mini](https://staffbox.ai/local-ai-mac-mini/): what a 16 GB Mac mini runs for a small business, measured scores, speeds and Apple's power rating
- [Agent desk comparison](https://staffbox.ai/agent-desk/): Staffbox vs Perplexity Computer and other agent desks
- [Models](https://staffbox.ai/models/): every model option and its score on the same test; cloud providers on the customer's own key (Anthropic, Gemini, Azure, OpenRouter, Salad, Baseten, Crusoe)
- [For IT providers](https://staffbox.ai/partners/): how managed IT providers install and support Staffbox for their clients
- [Government](https://staffbox.ai/government/): public offices; non-Chinese model option; Georgia purchasing paths with sources
- [Trust Center](https://trust.staffbox.ai/): security and privacy controls with honest statuses
- [System status](https://status.staffbox.ai/)
- [Source code](https://github.com/Staffbox-ai/staffbox): installer, company brain, quote tool, test results
- [Hardware tiers](https://github.com/Staffbox-ai/staffbox/blob/main/docs/hardware.md)
- [Staffbox on Hugging Face](https://huggingface.co/staffbox)
- [Scorecard: every graded answer](https://huggingface.co/spaces/staffbox/scorecard)
- [Test tasks on Hugging Face](https://huggingface.co/datasets/staffbox/staffbox-evals)

## Contact

Hadi Irvani, founder. hadi@staffbox.ai · +1 404 493 6248 · [LinkedIn](https://www.linkedin.com/in/hadiirvani)