Hermes Agent hub

Hermes Agent for business: what it gives and what it still needs

Short answer: Hermes Agent is free; with company notes and a quote action, a 16 GB Mac mini got 15 of 15 unseen quote requests right (6 Oct 2026). The companies were fictional.

Hermes Agent is a capable open-source agent, and plenty of teams are trying it for real work. Out of the box it knows nothing about your prices, and it can be handed a shell, a browser and an email gateway that answers by itself. Give it one narrow job, your own notes, risky tools off and a person approving every draft, then score it before it goes live. Tell us what you are building in the form below; a person on our team replies.

Updated 14 October 2026 · built on Hermes Agent, not affiliated with Nous Research

Staffbox is built on Hermes Agent (Nous Research, MIT license). Staffbox is not affiliated with Nous Research.

Hermes Agent is a free, general base that leaves your business rules to you

Hermes Agent is an open-source AI agent from Nous Research, released under the MIT license. Its own site describes a self-improving agent with persistent memory, skills it writes for itself, and a messaging gateway that connects it to chat and email. As of 13 October 2026 its GitHub repository shows about 253,000 stars and a latest release of v0.21.7 (as reported by star-history.com and the repository release list). It runs on your own machine or a rented server, against a local model or a cloud one.

General is the point, and also the gap. The agent does not know your price list or which customers may ask for a quote, and it can be given tools that touch files, the web and a shell. A small firm that sends quotes needs the opposite: one narrow job, done on its own prices, with a person checking each result.

A finished install adds five things on top: notes, a quote action, tools off, drafts only and a test set

Staffbox is our version of that finished install: Hermes Agent tuned for one job, on a Mac mini in the customer's building. The diagram shows the layers; the table splits the work piece by piece.

What Hermes Agent gives and what Staffbox addsWhat Staffbox addsBrain: your notes, a vault you ownQuote action: prices in code, cites row + dateSafe default: shell, code, browser, web offDrafts only: no sending code, a person sendsTest set + scorecard, re-run before updatesWhat Hermes Agent gives (MIT)Agent loop, memory, skills, toolsMessaging gateway for chat and emailProfiles and model settings (config.yaml)Talks to a local or a cloud modelWhat both run onMac mini in your buildingOllama on 127.0.0.1, open-weight model
Staffbox is a layer on top of a Hermes Agent install. Hermes items are from its own site and docs, as reported; Staffbox items are in the open repository. The table below has the same split, piece by piece.
PieceHermes Agent gives youWhat Staffbox adds
Agent loop, memory, messagingYes, open sourceUsed as is
Company knowledgeGeneric memoryA "brain": a Markdown/Obsidian vault the customer owns
QuotingNo quote logicA quote action that prices from the customer's own price list and cites the row and the list's date
ToolsTerminal, code, browser, web, delegation are availableThe installer turns them off
EmailGateway can reply by itself (per its docs)Writes draft replies only; Staffbox has no sending code
ModelYour choiceLocal open-weight models through Ollama; cloud only on the customer's own key, off by default
ProofNone for your businessA scorecard on 30 of your past quotes, free, before you pay fact

The pattern behind this: score first, then fix in a fixed order (the brain, then the tool, then the model), go live only on workflows that pass, and re-test every month. Corrections from your staff become new tests. See how Staffbox works for the full loop.

Notes plus a quote action took an 8B model from 1/40 to 40/40, and it got 15 of 15 unseen requests (6 Oct)

No customer has run Staffbox yet, so every score here comes from fictional companies: Fieldstone IT, an IT provider, and Peachtree Cabinet Works, a cabinet shop. Answers were graded by machine against known answers. The honest test is the set nobody saw while fixing: on a 16 GB Mac mini with Qwen3 8B, company notes and the quote action, the unit got 15 of 15 unseen messy requests right (6 October 2026). On the clean sets it scored 40/40 on Fieldstone IT and 39/40 on Peachtree Cabinet Works. A generated regression set (requests written by a model, answers checked by code) came back 351 of 360. On a new trade it had never seen, an electrical distributor (14 Oct, Dell test box), the default Qwen3 8B setup scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55, because per-foot units, price levels and trade shorthand are not handled yet; that is why we score your own list first (the electrical result).

Right answers by test set, 16 GB Mac mini, 6 October 2026 Messy set 2 (unseen) 15/15 Messy set 1 (seen while fixing) 19/20 Fieldstone IT (clean set) 40/40 Peachtree Cabinet Works (clean set) 39/40 Generated regression set (model-written) 351/360 0%100% of each set
Right answers by test set; bars start at zero and the grey track is the whole set. 0 wrong totals in 475 answers; of the 11 misses, 10 declined and asked the customer and 1 was a routing miss. Fictional companies, graded by machine; 16 GB Mac mini M4, Qwen3 8B, 6 Oct 2026. Source: results files on GitHub.

The notes did most of the lift and the quote action closed the rest. On Fieldstone IT (Dell RTX 3090, 30 September 2026), Qwen3 8B scored 1/40 alone, 33/40 with the notes, and 40/40 with the notes plus the quote action; Qwen3 27B scored 2/40 alone. The local models page has the full comparison.

Fifteen or forty requests is a small sample, and fictional companies are easier than real ones. The free 30-quote test on your own past quotes is how you find out. Every graded answer is public on our scorecard Space.

Run it yourself when mistakes are cheap; use a scored, installed unit when a wrong price costs money

If you have someone comfortable with a terminal, a spare machine and a job that tolerates mistakes, the open-source project is free and well documented. Our managed versus DIY page sets out the trade. If the job is sending quotes to customers, a wrong price costs money, and you will want a score, a log and a person approving each reply. That is the gap we built Staffbox to fill.

Price: $595 a month at founding sites (first 100) for as long as the box is live (draft pricing), free 30-day pilot, no onboarding fee. $695 a month plus $1,800 onboarding from unit 101.

Every Hermes Agent guide, one question per page

You areStart hereWhen you're ready
Running Hermes Agent yourself for a businessInstall Hermes Agent on a Mac miniTalk Hermes with us
An IT provider adding Hermes Agent for clientsHermes Agent for MSPsGet the draft partner terms
A business that wants it run for youManaged or DIY Hermes AgentSend 3 requests, free

Related pages elsewhere on this site: how it works, AI quoting software, local AI on a Mac mini, models, trust.

Questions

Is Hermes Agent free?

The Hermes Agent software is open source under the MIT license, so the code is free. You pay for the machine, the model if it is a cloud model, and your own time to set it up and keep it running. Staffbox is a paid service built on it, priced at $595 a month at founding sites (first 100) for as long as the box is live.

What is the difference between Hermes Agent and Staffbox?

Hermes Agent is the general agent. Staffbox is a Hermes Agent install configured for one job, drafting quotes from your price list, with risky tools switched off, drafts only, a scorecard and monthly re-tests.

Can Hermes Agent run without the cloud?

Yes. Hermes can use a local model through Ollama. Staffbox runs this way on a Mac mini, with the model server listening on 127.0.0.1 only and no cloud provider configured.

Does Staffbox have customers?

Not yet. No customer has run Staffbox. All scores on this site come from fictional demo companies.

Sources

Or call us: 404 493 6248

Ask usTextCall