Learning path · Local models

Local AI agents: OpenClaw, Hermes Desktop and VS Code

A local model is the brain. An agent app is the hands. These three free tools turn a model running on your own computer into an agent that can actually do work, with nothing sent to the cloud.

Checked October 15, 2026 against each project's official docs. Every step links to its source.

Which one is for you?

ToolWhat it isWho it's forHardwareCost
Hermes Desktop (Hermes Agent by Nous Research)A desktop app for an open-source agent with memory and skills. Its Local Models screen downloads and runs models for you.Beginners who want the fewest stepsA GPU or Mac with 8GB+ runs the small models; 16GB+ runs 27-35B modelsFree, open source (MIT)
OpenClawA personal assistant on your computer that you message from WhatsApp, Telegram, Slack and other chat apps.People who want a 24/7 assistant they can textmacOS 15+, Windows 10/11 or Linux, plus enough memory for your model (see the tiers below)Free, open source (MIT), no paid edition
VS Code + Ollama, or ContinueThe free code editor. Its Chat can use local Ollama models, and the Continue extension adds local autocomplete.Anyone working on code or text files who wants privacy8GB minimum, 16GB+ recommended (Continue docs)Free; no GitHub account or Copilot plan needed for local models

Step 0 (all three): install Ollama and pick a model size

Ollama is a free app that downloads and runs open models. All three tools below can use it.

  1. Go to ollama.com/download, choose macOS, Windows or Linux, and install it like any app. You should see the llama icon in your menu bar (Mac) or system tray (Windows). If not, open Ollama from Applications or the Start menu.
  2. Open Terminal (Mac: Cmd+Space, type Terminal) or PowerShell (Windows: Start, type PowerShell).
  3. Download a model that fits your computer. Agents need a model that supports tool calling; Gemma 4 does. Sizes are from the official Ollama page:
    • 8GB laptop: ollama pull gemma4:e2b (4.6GB to 7.5GB download)
    • 16GB laptop or Mac: ollama pull gemma4:e4b (6.6GB to 9.5GB)
    • 24GB to 32GB Mac, or a 24GB GPU: ollama pull gemma4:26b (16GB to 19GB). This is the size that handles agent work best.
  4. Agents need room to think. OpenClaw, Hermes and VS Code all recommend a context window of at least 64K tokens: open Ollama's Settings and raise the context length (see Ollama: context length). Small models on 8GB machines will be slow at agent tasks; that's normal.

Not sure what fits? Use the model picker.

Hermes Desktop: the easiest local agent

What it is: Hermes Agent is an open-source agent from Nous Research. Hermes Desktop is its app for macOS and Windows. It can install its own local model engine (llama.cpp), so you don't even need Ollama.

  1. Go to hermes-agent.nousresearch.com and download the desktop package. On a Mac, open the DMG and drag Hermes.app to Applications. On Windows, open the .appinstaller file.
  2. Open Hermes, then go to Settings → Providers → Local Models (or choose "Run models locally" during onboarding).
  3. Click Install runtime. It downloads a few hundred MB and checks it.
  4. In the catalog, each model is marked green (fits your GPU), amber (uses system RAM, slower) or red (too big). Pick a green one and click Download, then Use. New chats now run on your computer.
  5. Already use Ollama? In a terminal run ollama launch hermes-desktop instead; Ollama installs Hermes if needed and connects your model.

If it doesn't work: if every model shows red, your machine doesn't have enough memory; switch to a cloud model in the model picker. Sources: installation, local models, Ollama: Hermes Desktop.

OpenClaw: a local assistant you can text

What it is: a free, open-source personal assistant (previously called Clawdbot) that runs on your computer and answers you in chat apps. It can browse, fill forms, manage files and run commands, so treat it like giving someone your keyboard.

  1. Do Step 0 and pull a model, for example ollama pull gemma4.
  2. Download the OpenClaw desktop app from openclaw.ai (macOS 15+, Windows 10/11, Linux). It installs everything it needs, including Node.js. Or, in a terminal, run ollama launch openclaw and Ollama sets it up for you.
  3. In onboarding, choose Ollama, then Local only. Pick your model when asked.
  4. Choose sandboxed access first, then connect one chat app (for example Telegram) and send it a simple request.

If it doesn't work: if the model replies with raw code-looking text instead of acting, the Ollama address is wrong. It must be http://127.0.0.1:11434 without /v1. Sources: install, Ollama setup, Ollama: OpenClaw.

VS Code with local models

What it is: VS Code is Microsoft's free editor. Its Chat (the same panel GitHub Copilot uses) can run on your own models through "bring your own key", with no GitHub account or Copilot plan, even fully offline.

  1. Install VS Code (version 1.127 or newer) and make sure Ollama is running with at least one model, for example ollama pull qwen3.6 on a 24GB+ machine or gemma4:e4b on 16GB.
  2. In VS Code, click the Extensions icon, search Ollama, and install the official extension from the Ollama publisher.
  3. Open Chat, click the model picker at the bottom of the chat box, and choose a model under Ollama.
  4. Optional: install Continue from the Marketplace for local tab autocomplete. Run ollama pull qwen2.5-coder:1.5b and add it with the autocomplete role (VS Code Chat with your own model doesn't do inline completions).

If it doesn't work: no Ollama models in the picker? Check Ollama is running, run ollama list, then run "Ollama: Refresh Models" from the Command Palette. Work accounts on Copilot Business or Enterprise need the admin's BYOK policy on (it's on by default). Sources: VS Code: language models, Ollama: VS Code, Continue: Ollama guide.

Keep going