Hermes Agent for business: questions people ask

Hermes Agent is open source under the MIT licence from Nous Research, so the software costs nothing to download. You still pay for the computer or server it runs on, the model if you use a cloud one, and the time to set it up. Staffbox is a separate paid service built on it: $595 a month at founding sites (first 100) for as long as the box is live (draft pricing), with a free 30-day pilot.

Yes. Hermes Agent can use local models through Ollama. A Staffbox unit runs local open-weight models on a Mac mini, the model server listens on 127.0.0.1 only, and no cloud provider is configured. Cloud AI runs only on the customer's own key and is off by default.

For the Staffbox quote desk we scored a 16 GB Mac mini M4 with Qwen3 8B. Pilots run on a 32 GB unit, measured on the same tests before day 0, and the non-Qwen option gpt-oss-20b needs 32 GB. Mac Studio is planned but its speed has not been measured.

It depends on the job, so we score models on the job. On our fictional quoting tests, Qwen3 8B on a 16 GB Mac mini scored 40 of 40 on Fieldstone IT and 39 of 40 on Peachtree Cabinet Works on 6 October 2026. gpt-oss-20b scored 39 of 40 and 40 of 40 on a Dell. No customer has run it yet.

As reported by other writers and a GitHub issue, Hermes Agent expects at least 64K tokens of context, which is why models with smaller windows give errors. Hermes refuses local model windows below 64K, so the Staffbox installer sets OLLAMA_CONTEXT_LENGTH=65536 on the Ollama server and context_length and ollama_num_ctx: 65536 in the Hermes profile, and all Mac mini scores were measured at that setting. Prompts longer than Qwen3 8B's native 40,960 tokens have not been tested; the test prompts are far shorter. Changing the size costs about 100 seconds on a 16 GB Mac, so we keep one size per test run.

It depends on how it is set up. Security researchers have reported problems with Hermes Agent, and we link them as reported, not as verdicts. A Staffbox unit turns off nine tool groups by default (terminal, code execution, browser, computer use, web, connections, delegation, cron jobs and text-to-speech) plus image generation, and has no sending code. It has no SOC 2 or other certification yet.

It can if you connect a mailbox. In Staffbox, the inbox command reads requests from an agreed folder or mailbox and writes draft replies that cite the price-list row and its date. The access is agreed in writing at install and limited to the folders and mailbox named.

On a Staffbox unit, no. Staffbox has no sending code and the unit has no permission to send or delete anything. A person checks each draft and sends it. Other Hermes setups may behave differently, so check the settings of whatever you run.

You can do it yourself from the official docs, rent a hosted version, or have your IT provider install a Staffbox unit. Staffbox is sold and supported through managed IT providers: they install the Mac and give first-line support, and Staffbox builds the brain and gives second-line support.

You find out before you pay. Send 30 past quote requests and your price list. We run them on our own Mac mini in Atlanta, you grade 10 held-back answers blind, and we delete the files within 14 days of sending your results (per our draft data policy). The score is free. You can also start with a free test of three requests.

No. Staffbox is not affiliated with Nous Research. We build on Hermes Agent, which is MIT-licensed open source, and we link the official repository and documentation.

Not yet. There is no customer install and no paying customer. The first sites are founding sites with the best price. All our measured results use fictional demo companies, graded by machine.

Yes. The default model is qwen3, which comes from Alibaba. For buyers who cannot use Alibaba models, such as Georgia state entities under GTA standard SS-22-002, there is a non-Qwen option, gpt-oss-20b, which needs 32 GB. Its speed on Apple hardware is not measured yet.