1. Install the official Mac app
Use the App Store listing or the official Mac download. On Mac, Local Code needs Apple Silicon and macOS 26.0 or later. See official downloads and platform availability for iPhone and iPad information.
Installation requirements do not tell you which local model will fit. Before downloading one, check unified memory in About This Mac and available storage. A large checkpoint, temporary download files, context caches and your project outputs all need room.
2. Choose where the model runs
For on-device work, download a model from the app's catalog. Qwen 3.8 27B in its smaller quantization is a useful starting point for a 16 GB Mac. Larger quantizations and DeepSeek models need more resources. The memory and storage guide shows which models fit each Mac.
You can instead choose a supported cloud provider or an OpenAI-compatible server. This can reduce local model storage needs, but inference then uses a connection and sends the request to the selected service. The agent's tools still run in the app's sandbox. Review the provider and data-flow boundaries before including private material.
3. Give the app a workspace
Create a folder for your first experiment and copy in a small, non-sensitive input file. Grant access to that workspace in Local Code. The agent starts in a project folder, but its permitted workspace is the access boundary: do not treat sibling projects in the same workspace as isolated from one another.
Keep your original data elsewhere while learning the workflow. Tell the agent which files to read and where to write outputs. Grant additional folders or personal-data permissions only when the task needs them.
4. Ask for a concrete output
Read sales.csv in this project. Describe its columns and any missing values. Create a monthly revenue chart and a short report in an outputs folder. Keep the source CSV unchanged and explain how you calculated each total.
This gives you several things to inspect: the input assumptions, computed totals, chart and report. The built-in Python and document tools can work with local files. If a task requests a new package, a website or an online service, that step may require a connection even when the model is local.
5. Review and iterate
Open the generated files. Check a few totals against the original data, inspect the chart labels and confirm that the document opens correctly. Ask the agent to fix a specific problem or adjust the result. For coding tasks, run relevant tests and review the diff before using the output.
If the task stalls, reduce its scope, shorten the context or try a smaller model. For offline use, prepare the model and dependencies while connected, then test a representative task with the network unavailable. See more document, spreadsheet, coding and research workflows.
Official references
See the App Store compatibility information and the official sandbox and tool documentation.