Updated October 2026
Run a local AI agent with Ollama
Most Ollama apps are chat windows: you type, the model answers. An agent goes further. It can read and write files, use tools and the browser, remember context, run on a schedule and reply to you in chat apps. CoWork OS is a free, open-source desktop app that turns a local Ollama model into that kind of agent.
Ollama chat app vs Ollama agent
| Typical Ollama chat app | CoWork OS with Ollama | |
|---|---|---|
| Chat with a local model | Yes | Yes |
| Work with your files and folders | Usually not | Yes, inside the workspace you choose |
| Use tools, MCP servers and skills | Rarely | Yes |
| Scheduled and background tasks | No | Yes |
| Reply from Telegram, Slack or WhatsApp | No | Yes |
| Approvals before risky actions | Not needed | Per-task access profiles |
1. Install Ollama and pull a model
On a Mac with Homebrew:
brew install ollama
ollama pull llama3.2
ollama serve
On Windows, install Ollama from ollama.com and run ollama pull llama3.2 in a terminal.
2. Connect Ollama in CoWork OS
- Install CoWork OS and open Settings → AI & Models → Model Access.
- Choose Ollama (Local).
- Leave Ollama Server URL as
http://localhost:11434unless Ollama runs elsewhere. - Click Refresh Models and pick your model. The API key is optional and can stay empty for a local server.
- Click Test Connection, then Save Settings.
3. Run your first local task
Pick a test folder as the workspace, keep the Ask for approval access profile, and try something you can check, such as "Summarise the documents in this folder into a one-page brief." Once models are installed, CoWork OS can run with Ollama offline.
Which Ollama model should you use?
| Model | Good for |
|---|---|
llama3.2 (3B) | Fast, simple tasks on modest hardware |
qwen2.5:14b | More involved writing and multi-step tasks |
deepseek-r1:14b | Reasoning-heavy work |
Agent tasks that call many tools are harder for small models. If a task stalls, try a larger model or route that task to a cloud provider; CoWork OS lets you mix both.
Context length
Ollama defaults to a short context window on smaller GPUs, which cuts agents off mid-task. CoWork OS asks Ollama for a 32,768-token context by default. You can change it with the COWORK_OLLAMA_NUM_CTX environment variable if your hardware needs a smaller or larger window.
More detail is in the providers documentation.
Frequently asked questions
Can Ollama run an AI agent?
Yes. Ollama serves the model; an agent app such as CoWork OS adds files, tools, memory, schedules and approvals around it so the model can act, not just chat.
What is the best Ollama GUI with agent features?
CoWork OS is a free, open-source desktop app that uses Ollama models for full agent tasks: files, tools, MCP servers, scheduled jobs and chat-app channels, with approvals before risky actions.
Does CoWork OS work offline with Ollama?
Yes. Once Ollama and your models are installed, CoWork OS can run tasks offline. Features that need the internet, such as web search, still need a connection.
Which Ollama model works best for agents?
Larger models handle tool use better. llama3.2 is fine for simple tasks; qwen2.5:14b or deepseek-r1:14b suit more involved work if your machine has the memory.
CoWork OS itself is free and open source. Model providers, gateways, connected services, and account plans may charge separately and apply their own eligibility and usage rules.
Last reviewed 2026-10-20 against official documentation. Features, plans and prices change; check the linked sources before deciding. CoWork OS is not affiliated with Ollama.