Bring your local model
An Ollama AI agent workspace
Ollama runs compatible language models on your own hardware. LocalAI Cowork adds the workspace around the model: tasks, selected files, tools, approvals, reusable workflows, and results.
Why connect Ollama to an agent workspace?
An Ollama chat can answer questions. Agentic work also needs context and controlled actions. LocalAI Cowork can keep a requested outcome visible, pass the selected context to the model, expose allowed tools, pause for approval, and preserve task state. The model remains replaceable rather than defining the whole product.
A practical starting setup
- Install Ollama on the same Windows machine and pull a model suited to your hardware.
- Open LocalAI Cowork settings, choose Ollama, and confirm that the endpoint and model are available.
- Create a test project with copies of non-sensitive files.
- Ask for a small, verifiable result and compare the output with the source material.
- Increase tool permissions only after the read-only flow behaves as expected.
Choosing a model
Model size is a trade-off. Larger models often plan and follow instructions more reliably but require more memory and run more slowly. Smaller quantized models are easier to run locally but may need narrower tasks and more review. Use a model whose context window can hold the relevant instructions and file excerpts; adding every file usually makes results worse, not better.
When to use a remote model
Local models are not always the best tool. A complex research, reasoning, vision, or large-context task may benefit from a hosted endpoint. LocalAI Cowork also supports configurable compatible providers so you can create separate profiles. Sending a task to one of those profiles transfers the required prompt and context to that provider.
Security basics
Do not place API keys in project files or prompts. Store provider credentials through the app's credential interface, keep backups, and review commands before approval. Ollama reduces external model traffic; it does not make arbitrary MCP servers, web tools, or shell commands safe.