GoldieBench Blog · 7 min read
Hermes Agent SSH + Which AI Models To Trust With A Remote Server, With Real GoldieBench Scores
Every command lands on a real machine, so the model choice counts. Here are ten GoldieBench scores and how to use them for profiles that read or change a server.

The model behind hermes agent ssh matters more than usual, because every command it chooses lands on a real server you own, not in a throwaway container.
The Hermes SSH backend reference explains how commands reach the server, but it doesn't recommend any model for the job.
Below are the GoldieBench scores I read on the live GoldieBench leaderboard on 14 November 2026, and how to use them when you choose a model for a server-facing profile.
If you'd rather my team matched a model to your servers, book a free SEO strategy session.
What The Model Actually Does On An SSH Backend
With terminal.backend set to ssh, Hermes stays on your machine and sends each shell command to the host in TERMINAL_SSH_HOST.
The model decides which commands to run, reads the output and decides the next step.
Hermes keeps its dangerous-command check on SSH, per the security comparison, so risky commands still pause for approval.
That check is a safety net, not a substitute for a model that reads logs carefully and picks conservative commands.
A model that guesses at file paths or misreads an error can still waste approvals or make a harmless-looking change in the wrong folder.
The GoldieBench Scores I Used
| Model | GoldieBench average | Listed as |
|---|---|---|
| Fusion | 8.59 | OpenRouter, best at Games |
| Claude Opus 5 | 8.27 | Anthropic, best at Sims |
| Hermes MoA | 8.17 | Hermes Mixture of Agents, best at Visuals |
| GPT-5.6 Sol | 8.16 | OpenAI, best at Pages |
| Claude Fable 5 | 8.10 | Anthropic, best at Games |
| Qwen 3.8 | 8.10 | Alibaba, best at Sims |
| Grok | 8.09 | xAI, best at Games |
| Kimi K3 | 7.89 | Moonshot AI, best at Games |
| GLM-5.2 | 7.77 | Zhipu / Z.ai, best at Pages |
| Gemini 3.6 Flash | 7.08 | Google, best at Pages |
All averages are out of 10, as shown on the leaderboard on 14 November 2026.
Scores move as new tasks are scored, so recheck the leaderboard before you commit.
What A GoldieBench Score Tells You Here
The methodology page says every model gets the same fixed prompt, one shot, and must return one self-contained HTML file.
So these scores measure one-shot building quality across games, visuals, simulations and pages.
They are not a server-administration test, and no GoldieBench task runs commands on a remote host.
What carries over is general reliability: a model that follows a long prompt exactly and gets it right first time is more likely to follow your server brief exactly too.
Use the scores to shortlist, then test the shortlist on your own server with read-only jobs.
Picking A Model For A Profile That Changes The Server
If the profile edits configs, restarts services or deploys files, start from the top single models on the table.
Claude Opus 5 sits at 8.27 and GPT-5.6 Sol at 8.16.
Hermes MoA, at 8.17, is listed as a mixture of agents: its model page describes a panel of frontier models whose drafts are merged by an aggregator.
That suits careful planning, though a panel adds calls to every step, which you will feel on a chatty maintenance session.
Fusion leads the whole table at 8.59, and is listed under OpenRouter.
Picking A Model For Read-Only Server Reports
Daily status checks, disk reports and log summaries change nothing on the server.
For those, a model a few places lower is a sensible place to test, such as Qwen 3.8 at 8.10, Grok at 8.09 or Kimi K3 at 7.89.
The Kimi K3 page also lists an external reference line, "Terminal Bench the standout", from launch-day community benchmarks.
That line is a reference benchmark the page pulls from elsewhere, not a GoldieBench medal score, but it is the one note on these pages that points at terminal work.
Keep the read-only profile on a narrow login so a mistaken command has little room to do harm.
Split Server Jobs Across Two Profiles
A clean pattern is one profile per risk level.
The reporting profile runs a mid-table model and a login that can only read logs and folders.
The change profile runs a top-table model and a login with the extra rights it needs.
Each profile carries its own model setting and env file, so both can point at the same server with different users.
Checkpoints and /rollback still work for both on SSH, according to the checkpoints docs.
Settings That Matter Whichever Model You Pick
- Persistent shell. Only the working folder and exported variables survive between commands; tell the model to export anything it needs later.
- Skill variables. The server drops forwarded values unless
sshd_confighas a matchingAcceptEnvline. - Provider keys. Hermes never forwards keys like
OPENAI_API_KEYto the server, so the model's credentials stay on your machine. - Sudo. The tools reference says
SUDO_PASSWORDgoes to the backend running the command, so on SSH it is the server's password.
Why Approval Prompts Are A Model Test Too
On the SSH backend, the dangerous-command check turns every risky step into a prompt you have to answer.
That makes the number of prompts a useful, free measure of a model's judgement on your server.
A careful model reads the folder first, then proposes one targeted command, so it triggers few prompts and each one makes sense.
A careless model reaches for broad commands early and fills your screen with approvals you have to read and refuse.
Track that count for a week per model, alongside whether the reports were correct.
Container backends skip this check entirely, so the measure only works on SSH and local.
It is a simple signal that sits alongside the GoldieBench averages and tells you something the leaderboard cannot.
Remote Installs Through Hermes Desktop
If the server runs its own full Hermes install, Desktop can drive it through an SSH gateway connection, added under Settings, Gateways with the host as user@host:22.
In that setup the model is chosen on the remote install, not on your laptop.
The v0.23.0 release keeps every registered gateway connected, so you can compare a remote profile and a local one side by side.
The Hermes multi-gateway connections guide covers that layout.
Testing Your Pick On A Real Server
- Give the shortlisted model a read-only brief: hostname, disk space, last 50 log lines.
- Check that it used the right paths and summarised the errors correctly.
- Ask for one reversible edit in a scratch file, then confirm
/rollbackrestores it. - Count how many approval prompts it triggered and whether each one was needed.
- Repeat with the next model on the shortlist and keep the one with cleaner runs.
Members of the AI Profit Boardroom share which models they run on their own servers on the weekly calls.
My Verdict On Models For Hermes Agent SSH
Put a top-table single model on any profile that changes a live server, with Claude Opus 5 at 8.27 and GPT-5.6 Sol at 8.16 as the starting shortlist.
Test mid-table models like Qwen 3.8 and Kimi K3 on read-only reporting work, where a slip costs far less.
Treat every GoldieBench number as a shortlist tool, then let a week of real server runs make the final call.
Also On Our Network
- Agent Operatorsthe six-step setup for pointing a Hermes agent's terminal at a remote server
- AI Income Deskwhether a small business should let its Hermes agent work on its own server
- AI Tool Verdicta nine-point scorecard for the Hermes SSH terminal backend, rated from the docs
- agentos.guidefive copy-paste prompts to point an Agent OS profile at a remote server
Shortlist two models from the leaderboard, give each a read-only server brief, keep the cleaner one, then ask my team on a free strategy call or compare notes with members of the AI Profit Boardroom while you tune your hermes agent ssh profiles.
FAQ
Which model should run a Hermes profile that changes my server?
Start with a top single model on GoldieBench: Claude Opus 5 (8.27) and GPT-5.6 Sol (8.16) were the leading single models on the leaderboard read on 14 November 2026.
Do GoldieBench scores measure SSH or server work?
No. The methodology is one fixed prompt, one shot, one self-contained HTML file. Use the scores to shortlist, then test on read-only server jobs.
Can a cheaper model handle read-only server reports?
That is the safest place to test a mid-table model such as Qwen 3.8 (8.10) or Kimi K3 (7.89), on a login that can only read.
Does Hermes recommend a model for the SSH backend?
The Hermes SSH documentation covers connection settings, persistence and passthrough, and names no model for the job.
Does my model's API key get sent to the server?
It stays at home. Provider keys like OPENAI_API_KEY are excluded from SSH passthrough by design, so the model's credentials remain on the Hermes machine.
Should I use Hermes MoA on a server profile?
It scores 8.17 and merges drafts from a panel of frontier models, which suits careful planning, but each step makes several model calls.


