Running An AI Agent On Someone Else's Free Computer

There is a particular kind of satisfaction in getting a machine to do work for you while you pay nothing for the machine. This is roughly the appeal of running an autonomous AI agent on a free cloud server. The agent runs on hardware you do not own, on a bill you do not pay, using a language model someone else is also hosting for free.

Hermes AI Agent is an open-source framework that can execute tasks using large language models. It is available on GitHub, and it does not demand much. Oracle Cloud Free Tier offers Always Free compute instances, including ARM-based Ampere A1 machines with up to 4 OCPUs and 24 GB RAM.

OpenRouter, meanwhile, provides free access to select models such as Mistral 7B and Llama 2, with rate limits attached. Put those three pieces together and you have a small, self-directed worker living on a free machine. The catch, as usual, is in the setup.

Three Free Things, One Small Tower

The first free thing is the computer. Oracle's Always Free tier is unusually generous by cloud standards. The ARM-based Ampere A1 shape offers up to 4 OCPUs and 24 GB of RAM, which is more than most people's first laptop. There is also an AMD-based VM.Standard.E2.1.Micro with 1 OCPU and 1 GB of RAM.

The second free thing is the agent software. Hermes AI Agent is open source, so nobody is charging you a licence fee. It typically requires Python 3.8 or newer, plus packages such as requests, openai and langchain. That is the ordinary tax of running modern Python tooling.

The third free thing is the brain. OpenRouter offers free access to various LLM models, including Mistral 7B and Llama 2, subject to rate limits. Free models are the part most likely to disappoint you quietly, and we will get to that.

Keys Before Doors

Oracle Cloud instances require SSH key pairs for access, and you generate them during instance creation. If you have never done this, the process is less mysterious than it sounds. You create a key pair, keep the private half on your own machine, and Oracle holds the public half.

Think of it as a lock you install before you build the door. The private key never leaves your laptop. The public key sits on the server, waiting to check that whoever knocks is holding the matching half. Lose the private key and you lose the server.

One practical note: pick the ARM Ampere shape if you can. With 24 GB of RAM you have room to install dependencies without watching memory nervously. The 1 GB micro instance will run an agent, but it will feel like hosting a dinner party in a telephone box.

The API Key And The Base URL

Once the instance is running, the setup follows a familiar pattern. Create an Oracle Cloud account, launch a compute instance, connect over SSH, then install Python and the required dependencies. Clone the Hermes Agent repository from GitHub. At that point you have software with no one to talk to.

That is where OpenRouter comes in. You need an API key from openrouter.ai, and you set the base URL to https://openrouter.ai/api/v1. Hermes Agent can be configured to use OpenRouter by setting environment variables or by modifying its configuration file. Either route works.

The environment variable approach is usually less trouble, because it keeps secrets out of the repository. The configuration file approach is easier to read later. Choose based on whether you trust your future self to remember what you did.

Where The Free Model Starts To Creak

Here is the honest caveat. Free OpenRouter models carry usage limits, and an autonomous agent is a hungry thing. It does not ask one question and stop. It plans, calls tools, checks results and tries again, which means a single task can burn through a surprising number of requests.

Mistral 7B and Llama 2 are capable models, but they are not the sharpest tools available. An agent built on a smaller free model will sometimes misunderstand instructions, loop, or produce confident nonsense. This is not a reason to avoid the experiment. It is a reason to keep your expectations calibrated.

Monitor your usage. If you find yourself hitting limits constantly, upgrading is the straightforward answer. The technology works. The question is whether your patience outlasts the rate limiter.

Locking The Door Behind You

A free server is still a server, and it is reachable from the internet. The basics matter more than they sound. Use a non-root user for everyday work rather than operating as root. Configure firewall rules so only the ports you actually need are open. Keep the system updated.

None of this is exotic. It is the same advice that applies to any machine you expose to the world, and it is easy to skip when you are excited about getting the agent running. Skipping it is how a weekend project becomes somebody else's mining rig.

There is also a quieter point about autonomy. An agent that can execute tasks is a program with permissions. Give it only the permissions it needs, and check what it has been doing. The free tier makes experimentation cheap. It does not make carelessness cheap.

Final Thoughts

The interesting part of this setup is not the money saved. It is the shape of it. Three separate organisations, each giving away a different piece, combine into something that behaves like a small employee. Nobody designed that arrangement. It emerged from free tiers, open source and rate limits.

That is worth noticing, because it will not last forever. Free compute and free inference are marketing decisions, and marketing decisions change. What you learn building this will outlive the free tier it was built on.