Install Hermes Agent and connect a local model
In this lesson I install Hermes agent, the open source personal agent from NURS research and I connect it to a model that runs on this Mac, no cloud account and no key are needed and no prompt leaves the machine. Before you start you need Ollama with a model that can call tools. If you do not have it, install it with Homebrew, start the service and pull Gemma 4 E4B. here is the check all LAMA answers and the model list tools among its capabilities now Hermes the official installer is a shell script I read it before running it it clones the Hermes repository into a hidden folder in my home and downloads its own copies of Python, Node, Ripgrep and FFmpeg all checked by hash it installs nothing system-wide and it needs no sudo. I add three flags skip setup because I configure the model myself, skip browser so it does not download chromium and skip computer use so the agent cannot drive the mouse and the keyboard. This takes a few minutes so I speed up the video. The installer picks the latest stable release, clones it, prepares a private Python environment, builds the terminal interface and copies a set of bundled skills. At the end, it adds the folder of the Hermes command to the path in my shell startup file. version 0.2 3.1 the stable release this course uses next model Hermes needs a context window of at least 64 000 tokens its instructions and tool descriptions alone take about 10,000 Ollama gives the model a shorter window by default so I make my own variant of Gemma 4 with a larger window in a two-line model file from names the model I already have numchintx sets the context window to 65536 tokens nothing is downloaded again the variant shares the weights Now I point Hermes at Olama. Olama speaks the same web interface as OpenAI's service, so the provider is custom. The base URL, the address Hermes calls, is Olama's address with slash v1, and the model is my variant. The interactive command Hermes model does the same with the menu. I use config set because you can copy it. One more step makes a big difference with a local model. Every message sends the description of every enabled tool. I turn off the tool sets this course does not use, the browser, code execution, image and voice tools, web search, delegation and computer use. The status shows the model and the custom provider. The settings live in one file config.yml in the hidden folder .hermes in my home. in my home. Time for a first conversation. I start Hermes with the Hermes command. On the first start Hermes asks once whether to share anonymous usage counters with News Research. No thanks is selected and I keep it. The banner lists the model, the folder I am in, the tools and the skills. Below is the prompt. The model runs on this Mac so the first answer of the session takes a little longer. Olama reads the whole instructions first The answer comes from Gemma 4 on this Mac. The status line under the prompt shows how much of the 64000 token window this session uses. Hermes is installed, it talks to a local model and it keeps its settings in config.dml. You are ready to give it real work.