How to use ego (lite) with a local model

Set up local Qwen with Ollama on Windows, connect it to your Agent Harness, and use ego-browser to complete a browser task.

This tutorial connects a local Qwen model to ego (lite) on Windows. The recordings use DeepSeek Harness with a model labeled Qwen3.5 9B Local Demo to search Airbnb for accommodation in Tokyo and return three listings.

There are three parts to the setup: Ollama runs the model, your Agent Harness manages the conversation and executes tools, and ego (lite) opens the web pages through the ego-browser skill. Downloading a model alone does not connect it to the browser.

You can use another local model and Harness if they support tool calling, loading ego-browser, and executing its commands on the Windows computer where ego (lite) is installed.

Before you start: prepare the software

Install these applications on the same Windows computer:

Ollama for Windows: downloads the model and serves it locally. Open the application after installation and leave it running.
Your Agent Harness: the recordings use DeepSeek Harness. It must support local Ollama or an OpenAI-compatible model connection and local command execution.
ego (lite): install it and finish setup so your Harness can load the ego-browser skill. Fully close and reopen the Harness after setup, then start a new session.

Keep an internet connection for the initial model download and the websites you want to browse. Model inference can run locally while the browser still accesses online pages.

Set up the local model with Ollama

The setup clip opens Ollama's website, searches for qwen3.5, and views the model family and its available sizes. Follow the steps below to install and connect the 9B variant used in this tutorial. The clip does not show the model download or the Harness provider configuration.

1. Check Ollama and download Qwen

Open PowerShell in Windows Terminal and run:

ollama --version
ollama list
ollama pull qwen3.5:9b
ollama list

Wait for the download to finish. The final list should include qwen3.5:9b. Use the explicit :9b tag so you download the intended size rather than relying on the family's default tag. The official model page lists the available download and model details. Leave additional memory available for the model's context and your other applications; the download size is not its total runtime memory requirement.

If PowerShell cannot find ollama, complete the Ollama installation and open a new terminal. If it cannot connect to the service, open Ollama from the Windows Start menu. For installation and service details, see the Windows setup instructions.

2. Set the context length

Browser tasks accumulate instructions, page text, and tool results. In Ollama's application settings, set the context length to 64k or higher, provided your computer has enough available memory. Ollama's context-length guide recommends at least 64,000 tokens for agent tasks and explains the additional memory needed.

For a CLI-only installation, you can instead start the service from PowerShell with:

$env:OLLAMA_CONTEXT_LENGTH = "65536"
ollama serve

Use this as an alternative to the running Ollama application. Quit the tray application normally first so you do not start a second service on the same port. Leave this PowerShell window open and use another terminal for the next commands. The environment setting above applies to this terminal session; set it again when launching a new CLI service.

3. Check that the model responds

In PowerShell, run:

ollama run qwen3.5:9b

Send a short message such as Reply with one sentence to confirm you are ready. Wait for a reply, then enter /bye to leave the model chat. Keep Ollama running.

This checks that the model can generate text. The browser connection will be checked when you run the task below.

4. Connect your Harness to local Ollama

In your Harness's supported model or provider configuration, add the local connection. For an OpenAI-compatible connection, use these values from Ollama's compatibility documentation:

Field	Value
Provider name	A name you recognize, such as `Ollama`
API type	OpenAI-compatible
Base URL	Your local Ollama connection address
Model ID	`qwen3.5:9b` for the official model downloaded above
Display name	A name you recognize, such as `Qwen3.5 9B`
API key, if the client requires one	`ollama` as a placeholder; local Ollama ignores it

Add the provider through your Harness's supported settings or configuration file. Use the file format supported by your installed Harness, and preserve any unrelated settings when adding the provider.

Use the exact model ID from ollama list. A display name alone does not create or rename an Ollama model.

Use the local connection address required by your Harness's Ollama integration. Check whether it expects a native Ollama connection or an OpenAI-compatible connection before saving the provider.

Save the configuration and choose the local model in the Harness's model picker. The recording already has Qwen3.5 9B Local Demo selected below the message input. That is a display label; use the model ID returned by ollama list when configuring your own connection. Different Harnesses use different provider settings or configuration files, so use your Harness's supported setup method.

Send a short test message from the Harness with your local model selected. While it is generating, run this in another terminal:

ollama ps

Check the running model and its CONTEXT value. An empty list means no model is currently loaded; repeat the check during generation. This lets you distinguish a saved model entry from a model actually running in Ollama.

Step 1: Set session permissions

With your local Qwen model selected, open the permission menu below the message input. In the recorded DeepSeek Harness interface, switch from Workspace Write to Full access, select I understand the risks and want to continue, and choose Enable Full access.

These are Harness permissions for executing tools on your computer. They are separate from Ollama's model configuration. If you use another Harness, choose its corresponding local command permissions and review any approval requests.

Step 2: Load the ego-browser skill

Start a new session. Type /ego-browser, select ego-browser from the slash-command picker, add a space, and enter your task in the same message. Confirm that your local Qwen model is still selected, then send the message.

The recording asks the agent to search Airbnb for accommodation in Tokyo and find three inexpensive options. You can use:

Use ego-browser to search Airbnb for accommodation in Tokyo, Japan. Compare three inexpensive listings and return their links and the prices shown on the page.

Step 3: Let the local model work

Qwen generates the next actions through the Harness. The Harness runs ego-browser commands, opens Airbnb in an ego (lite) Space, applies search filters, and reads listings. Switch to the agent Space to watch the browser actions while your own tabs remain in your Space.

Keep Ollama and the Harness running until the task finishes. The model needs both the local inference service and the Harness's tool execution to continue.

Step 4: Check the result

The recorded task ends with a table of three listings, links, and price information in the Harness conversation. The agent's browser pages remain available in ego (lite), so you can open the returned links and compare them with the answer.

Troubleshooting

Ollama cannot connect or the model is missing

Open Ollama, run ollama list, and check for qwen3.5:9b. Finish the download before retrying. If the service is running but the Harness fails to connect, check its connection type, base URL, and exact model ID against the table above.

The model does not appear in the Harness

Save the local provider configuration and restart the Harness if it does not refresh its model list. Check that the provider can connect to your running Ollama service. A model listed by Ollama still needs to be added or discovered through the Harness's provider integration.

The model replies, but the browser does not move

Check that the Harness loaded ego-browser and is issuing tool calls. If the skill is missing, finish ego (lite) setup and restart the Harness. Review local execution permissions and any pending approvals. The Harness needs to execute commands on the same Windows computer as ego (lite).

The task is slow, runs out of memory, or stops

Read the error in the Harness or Ollama before retrying. Close unnecessary memory-heavy applications and start with a smaller browser task. Check ollama ps during generation for the loaded model and context. If you choose a smaller model, download it and update the Harness's model ID as well; it still needs to support tool calling.

Open the agent Space, complete the sign-in or verification yourself, and then ask the agent to continue.

Go deeper

DeepSeek Harness explains the interface used in the recordings.
Your own Agent Harness explains the requirements for another Harness.
Skills explains how an agent gets access to the browser tools.
Space explains where the agent's pages live.
Ollama for Windows covers installation and service setup.
Ollama context length covers context settings and memory use.