Using Cody and Ollama for Local Code Completion

Coding assistants can be useful, but hosted models require network access. In this January 2024 tutorial I show Cody in Visual Studio Code with Ollama and Code Llama 7B for local autocomplete. Cody sponsored the video. The local claim applies to the configured completion provider; it does not establish that every Cody feature, account operation or telemetry event is offline.

I use a MacBook Pro with M2 Max and find the 7B model responsive on that machine. Performance and memory needs depend on hardware and configuration. The simple examples shown here are demonstrations, not a comprehensive comparison with other coding assistants.

Install Visual Studio Code for your operating system, open Extensions and search for Cody AI. I install the prerelease extension, open Cody and sign in using Google; GitHub sign-in is also offered in the demonstrated flow. Installation, account setup and model download require network access in this walkthrough.

At the time I describe a free tier, rate limits and a temporarily free Pro offering. Those prices and model defaults are historical and may have changed. My focus is to switch autocomplete to a local model rather than assume every feature uses that model.

Install and start Ollama, then open a terminal and download the demonstrated model with `ollama pull codellama:7b-code`. I already have it cached, and the displayed download is approximately 3.8 GB. See [Ollama’s Code Llama instructions](https://ollama.com/blog/run-code-llama-locally); model tags and sizes can change.

In the then-current Cody extension settings, find Autocomplete > Advanced: Provider and select `unstable-ollama`. I leave the separate model override at null, noting that the displayed control refers to the Fireworks provider. This was an experimental 2024 configuration: newer versions may use different settings, and [Sourcegraph’s documentation](https://sourcegraph.com/docs/cody/capabilities/ollama) should be checked for current support.

To inspect the selected completion provider, open the VS Code Output panel and choose Cody by Sourcegraph. The log shows the Ollama provider and Code Llama model. This is evidence of the provider used for that completion, not an audit proving every network operation or feature is local.

I type a comment requesting a Fibonacci method, and Cody suggests code that I accept with Tab. The response looks fast on my computer. Generated code still needs review and testing; this example alone does not establish correctness on all inputs.

The lightbulb menu offers an “Ask Cody to generate” action. I request code to print numbers from 1 to 100 and show the options to accept, retry, undo or inspect the diff. The tutorial does not independently establish that this command uses the same local inference path as autocomplete.

Using an existing face-detection script, I add a line and show Cody suggesting a comment based on surrounding code. I then select code and request an edit to change the output image name. I accept the result and show model/provider information in the output panel.

Cody also offers chat and other features that, in this demonstration, are not powered by the local model. I select a hosted Claude model and ask what the referenced face-detection file does. Available models, plans and data handling are tied to the service and configuration, so local autocomplete does not make this chat local.

I demonstrate adding documentation and removing comments, then mention code explanation, detecting code smells and generating unit tests. I would find tests particularly useful, but generated changes and tests still need human review. I prefer Cody’s feature set to GitHub Copilot in this demonstration; that is my opinion rather than a controlled comparison.

Cody sponsored the video. The walkthrough shows one historical configuration for local completion and separate hosted capabilities; the available features, pricing and licensing can change over time. Code Llama’s weights have their own license and conditions, separate from Cody’s software and service terms.