Advanced

Advanced setups

This section is for readers who have finished a basic setup and want more from their agent, whether that means a lower bill, more privacy, or an agent that keeps working while they are away. Each page explains what a setup gives you, what it can cost you, and the steps for the routes that are officially documented. You decide what risk is worth taking, and the pages give you the facts to decide with.

Applies to: Agent Zero, Atomic Agent, Hermes Agent, OpenClaw, Goose, NanoClaw, QwenPaw, Eigent, Everfern, and OpenManus. Each page says near the top which of these it covers.

Start with your agent's own guide. These pages assume the agent already works with a cloud model, a spend limit, and its safe defaults, because every setup here builds on that.

How to read the risk labels

Tables in this section mark the main catch of each choice with one of five labels. A row can carry more than one.

Label What it means
Ban risk The company that sells the model or plan may suspend the account, because this route breaks its rules or sits close to them.
Surprise bill Costs can climb without a clear warning, such as extra charges after a plan's allowance runs out or an unattended run on a paid API key.
Privacy risk Someone besides you and the model maker can see your prompts and files, or the agent becomes reachable from outside your own devices.
Hardware dependent Whether it works well depends on your computer's memory and graphics chip, so results vary a great deal between machines.
Safe official The company documents and allows this route. It says nothing about what the agent itself can do, so the agent's own safety settings still matter.

The seven pages

Read them in any order. Pages 03 and 04 work best together, since an always-on machine usually means unattended runs.

Page What it covers Main risk
Local models in depth Running the model on your own computer with Ollama or LM Studio, which computer runs what, and connecting the model server to your agent Hardware dependent
Mix a cheap model with a strong one Sending routine steps to a cheap model and hard ones to a frontier model, and where each agent keeps that switch Surprise bill
Host access and unattended runs, done carefully Letting the agent work on real files or run while you are away, with backups, a separate account, allow lists, spend limits, and logs Real files changed or lost. No label covers this, so the page explains each case.
An always-on machine for your agent Keeping an old laptop, a mini PC, or a rented server running for schedules and phone access, and keeping it private Privacy risk and Surprise bill
Use a chat plan you already pay for Running the agent on a ChatGPT, Claude, or GitHub Copilot plan through officially documented sign-ins Ban risk and Surprise bill
Relays and aggregators Using one company's key to reach many models, what a relay can see, and how to judge one Privacy risk and Ban risk
Uncensored models Abliterated and uncensored models, the hosts and shops that sell them, and the risks of using one in an agent Privacy risk and Ban risk

Your agent's own advanced page

These shared pages cover setups that work with any agent. Each agent's guide also has its own advanced page with features only that agent has, such as Agent Zero's model presets and Flare tunnel. It is page 7 in each alternative agent's guide and Step 8 in the Agent Zero guide.

What this section does not cover

Some routes promise free or cheap models, and this guide names them only to say why it gives no steps for them. Tools such as gpt4free scrape other people's services without permission, so someone else pays for what you use. Juggling several free accounts to get around a provider's limits breaks the rules you agreed to, and the provider can close every account at once. Shared, resold, or leaked keys are unauthorized at best and stolen at worst, and the owner can see every prompt you send through them.

Google ended Gemini CLI consumer sign-ins on 18 June 2026, and its Antigravity terms call using an Antigravity sign-in from a third-party tool a breach. Google banned accounts for exactly that earlier in the year. Use a chat plan you already pay for explains what happened, and it gives no steps for those routes.

Before you change anything

The pages set out each risky choice the same way: what you gain, what you risk, and when it makes sense. Those trade-offs follow from three questions about what the agent can reach, whether you can undo what it does, and whether anyone is watching. Safety versus capability explains them, and its ladder shows the order in which to add capability, one rung at a time.

Sources

Each page in this section lists its own sources at the end, with a link for every factual claim it makes. This overview adds no new facts beyond the page summaries above, and the Gemini dates come from the sources on Use a chat plan you already pay for.

Checked 14 November 2026 · prices, limits, and versions change