MaruOpsInfrastructure and AI notes

About (Korean)Contact (Korean)

[Local LLM 20] Connect Hermes Agent and Review Tools and Memory

Hand taking notes next to a laptop

Reference photo · Hand taking notes next to a laptop · Photo: JESHOOTS.COM jeshoots / Source / CC0 · Resized for listings; illustrative photo, not the tested environment

Building a Local LLM · 20/30 · Connect with an Agent

Full series and learning sequence · Practice code ZIP

An agent uses the model’s response to request a tool, executes the permitted operation and returns its result to the model. Connect Hermes Agent to the local model with a narrow initial set of data and tool permissions. Check those boundaries before using it for repeated work.

Prepare practice folders for Ollama and a regular user account. Associate providers and models in a location free of sensitive files.

Start with a normal user experience

Create a lab folder in a separate regular user environment. It doesn’t use the admin terminal by default. Once you’ve confirmed that Ollama is running, start with the current official integration path. For versions without launch functionality, review Ollama updates or Hermes official installation documentation.

New-Item -ItemType Directory -Force "$env:USERPROFILE\hermes-lab"
Set-Location "$env:USERPROFILE\hermes-lab"
ollama --version
ollama launch hermes

On the installation/provider selection screen, select the model downloaded locally. :cloud model and messenger connection are excluded from this exercise. Windows native and WSL2 paths are supported, but here we proceed in an environment such as Windows Ollama.

Check the provider address directly

itemValue to associate with local Ollama
providerCustom endpoint or Ollama integration guide
API base URLhttp://127.0.0.1:11434/v1
API keyNo physical key required for local default Ollama
modelExact tags downloaded, excluding cloud suffix

The provider and model can be selected again with hermes setup or hermes model. The /v1 used here is different from the following OpenClaw native provider address. Even when setting up a connection to Ollama, check the API path required for each program.

hermes model
hermes tools
hermes prompt-size

First Request and Reading Tools

First, have them summarize a hypothetical inspection note. Check the responses and then add only the reading tools you need. Turn off unused browser, external search, and writing tools from the tools management screen. File access is limited by account permissions and provided quarantine settings.

The directive “Read only this folder” does not limit actual file permissions. A process can access only the scope permitted by its execution account. Use a practice account or container, and check the current Hermes security documentation for the behavior of command authorization and isolation.

Memory·skill·external connection management

Memorizing and reusing skills between conversations is convenient, but it can also leave behind incorrect procedures or sensitive values. Check where it is stored and how to delete it. The created skill is used again after reading the command, access address, and necessary permissions.

Even if the inference model is local, search tools, auxiliary models, and fallback providers can connect externally. Here we start by verifying a short request without an external provider key. If the tool call format is broken, check model capabilities, server compatibility, and prompt size.

reference material

Documentation and model tags checked: October 18, 2026. Primary lab environment: Windows 11 x64.