GPT-5.6 Sol + Hermes Agent: Full Setup With No API Bill

I connected GPT-5.6 Sol to Hermes Agent through the same $20 ChatGPT plan I already pay for. It now runs on a VPS, stays online when my laptop is closed, and builds websites I can open from anywhere.

The setup works without an OpenAI API key, so there is no separate per-token API bill. The trick is choosing the Codex sign-in path inside Hermes instead of pasting an API key into the setup screen.

Why GPT-5.6 Sol belongs inside an agent#

The raw intelligence is impressive, but efficiency matters more when a model works for hours instead of answering one prompt.

In OpenAI's GPT-5.6 launch data, Sol nearly matches Claude Fable 5 on its overall intelligence benchmark while finishing the work in 61% less time at about half the cost. On Agents' Last Exam, a benchmark for long tasks that require planning and tool use, Sol scores above the other models shown in the comparison.

That is the fit. Hermes is built for multi-step work. GPT-5.6 Sol is strongest when it can plan, use tools, inspect the result, and keep going.

I still use Codex and Claude Code on my laptop when I want to build something hands-on. Hermes handles the work I don't want to babysit. A VPS keeps it running 24/7, and the agent only gets access to that server instead of my whole computer. I can also reach it from my phone through Telegram.

If you want a broader tour before doing this setup, my Hermes Workspace guide explains the dashboard, agent teams, and the no-code deployment.

The setup has three real steps#

First, deploy the one-click Hermes Workspace app on a Hostinger VPS. KVM 2 is enough for most people. KVM 4 makes more sense if you plan to run several sessions at once. During setup, leave the OpenAI API key field blank.

Second, open the correct terminal. This is where most people get stuck.

The Hostinger install creates separate containers for the Hermes agent and the web dashboard. Open Docker Manager and choose the terminal on the container card that does not show the :3000 port. That is the agent terminal. If you use the Terminal tab inside the dashboard and see hermes: command not found, nothing is broken. You are simply in the wrong container.

Third, run:

hermes model

Choose OpenAI, then select the Codex sign-in option instead of the direct API option. Hermes gives you a device code and sends you to https://auth.openai.com/codex/device. Sign in with the OpenAI account tied to your ChatGPT subscription, approve access, and pick gpt-5.6-sol from the model list.

Verify the connection with:

hermes auth list

You want the result to show openai-codex and oauth. That confirms Hermes is using your ChatGPT login rather than an API key.

The free GPT-5.6 Sol and Hermes setup guide includes every command, prompt, and click from the video. Use it while you set this up so you do not have to copy commands from the screen.

The website problem nobody shows you#

Once GPT-5.6 Sol is connected, Hermes can write a full website from one plain-English prompt. But the first time I tested it, Hermes gave me a localhost link that my browser could not open.

The code was fine. The site was trapped inside the server.

I fixed that by running a small nginx preview server on port 8080 and giving Hermes one saved preference: put each new website in /opt/data/previews/<name>/index.html, make the files readable, then return the public VPS link.

There are two permission details that matter. The preview folder must be writable by the agent, and every site folder needs read and execute access for nginx. Miss the second one and the page returns a 403 even though index.html exists.

You do this once. After that, every site Hermes builds gets a live link in the same reply.

What GPT-5.6 Sol built from four short prompts#

I tested the setup with four different builds:

• Nightshift, a premium cold-brew subscription page with a dark, expensive look
• Sofia's, a warm neighborhood pizza site with a menu, hours, and pickup button
• A travel photographer portfolio with stock images and hover effects
• A responsive browser game with falling blocks and a working score

Each started from a one-sentence prompt. The pages were not production-ready businesses, but they were real, responsive sites that I could open in a browser and inspect from my phone.

The useful part is the loop. Ask for the first version, open the link, then say, "make the hero dark with a gold accent" and refresh. You are no longer copying code out of a chatbot or downloading a new file after every change.

For larger builds, I would split the job across a frontend agent, backend agent, and reviewer. I show that full workflow in my video on building autopilot agent teams with Hermes. It is the same principle I use in my AI business systems: one agent produces the work, another checks it, and you keep the final decision.

If websites are not your priority, start with one of the 11 Hermes workflows I run in my business. The same always-on setup can handle inbox triage, lead research, content briefs, and scheduled monitoring.

The honest limit of the ChatGPT route#

"No API bill" does not mean unlimited use.

ChatGPT plans have Codex usage limits, including a weekly cap. Normal building and testing can fit comfortably, but a busy agent running heavy jobs all day can hit the limit before it resets. If you need guaranteed high-volume work, move the heavy jobs to an API key or a higher ChatGPT tier.

OAuth tokens can also expire. If Hermes stops authenticating, run hermes model and repeat the device login. And if you ever see an API key attached to the OpenAI provider, remove it before continuing so you do not get surprise token charges.

If you would rather spread tasks across cheaper API models, my $8 Hermes Agent setup explains the OpenRouter cost cascade. Hostinger sponsored the video this post recaps, and the Hermes Workspace link above is an affiliate link. I used that same VPS setup for every test in the video.

The model will change. The useful system will not: put a strong model inside an isolated agent, keep it online, and make its work easy to review.

Watch the full walkthrough on YouTube: https://youtu.be/eo4X_BqDozA