Your machine. Your model.
Code with Ollama.
Skip cloud request quotas.
Connect a local model to OliveCode, work in your own project folder, and verify that the agent really edited your files.
Before you start
You need OliveCode, Ollama, enough disk space for a model download and enough memory to run it. Ollama supports macOS, Windows and Linux. A model that can chat is not necessarily good at editing files or using tools. Use the hardware recommendations in OliveCode’s Local menu as a starting point, then try a small coding task.
There is no cloud-model request quota for genuinely local inference. The app does not give your computer unlimited resources: longer context, larger models and other running apps increase memory requirements. OliveCode’s approximate model budget is guidance rather than a reserved amount of RAM.
1. Install and start the local runtime
Install Ollama from its official download page and start it. Use ollama list in Terminal to see models already registered with that runtime. The presence of an entry alone does not prove it is a downloaded local model: cloud entries can appear there too.
Install OliveCode using the command for your operating system. Open a new terminal, change into a project folder, then launch:
olivecode2. Choose a model in Local
Click Local beside the prompt, or type /local. OliveCode inspects the hardware and shows an approximate memory budget. Select a downloaded model that is ready on this computer, or choose a recommended download. Review its size before starting; larger models need more storage and memory.
Wait for the download to finish, then select the model. Check the model name beside your prompt. A model whose name includes a cloud variant is not a substitute for a downloaded local model. There is no single model we can promise is best on every computer.
3. Check actual coding, not just a chat response
Start in a disposable project. For a repeatable check, use the ten-test slugify example and the prompt on our free coding-agent guide.
Watch for actual file reads, edits and a test run in the expandable work details. A message saying “I created the file” is not enough. Inspect the file on disk, verify only the requested code changed, and run bun test ./slug.test.ts yourself. If the model only prints a tool call as text, try a more capable tool-using local model rather than treating its claim as completion.
If something does not work
| What you see | What to check |
|---|---|
| Ollama is unavailable | Start the Ollama app or service, then reopen Local. Confirm ollama list works. Do not expose the service publicly to make a local connection work. |
| Downloaded model is missing | Compare ollama list with OliveCode’s local-ready list. Cloud-only registrations are not local downloads. Confirm you are using the same Ollama runtime and model directory. |
| Very slow output or out-of-memory errors | Close heavy apps, use a smaller model or shorten the task/context. A large advertised context window does not mean it fits in your available RAM. |
| It chats but does not edit files | Inspect the work details for real tools and errors. Try a capable tool-using model and a small explicit task. File editing and reasoning quality vary. |
| You want disk space back | Open Local → Delete models to remove a downloaded model. Recheck with ollama list; do not manually remove shared blob files. |
What stays local?
Local inference runs on the computer hosting Ollama. Choosing that model does not make every OliveCode feature offline: cloud-based Assist models, browser tools, Telegram and other integrations may send data to services. Review the selected models, disable unwanted cloud features and use the privacy information to choose a workflow for your project. This guide does not claim to be a network-isolation audit.
For runtime-specific details, see Ollama’s quickstart, tool-calling documentation, and model storage locations.