Devin works with every AI model, so I ditched my expensive cloud setup for a local LLM

Published Sep 12, 2026, 6:00 AM EDT Anurag is an experienced journalist and author who’s been covering tech for the past 5 years, with a focus on Windows, Android, and Apple. He’s written for sites like Android Police, Neowin, Dexerto, and MakeTechEasier. Anurag’s always pumped about tech and loves getting his hands on the latest gadgets. When he's not procrastinating, you’ll probably find him catching the newest movies in theaters or scrolling through Twitter from his bed. I have been looking for a good harness for my local models, something as good as Codex. I have tried tools like OpenCode and Aider, but for some reason, they just don't have the same finesse as Claude Code or Codex. They also lack many of the features you'd expect from a coding agent now. More recently, I started using Devin, which is an alternative to the popular harnesses available today. It is no longer just a coding agent. Devin is more like an operating system for AI engineers. Its local version offers many interesting features, including the ability to run local models. Devin offers everything Codex has And then some Devin now overlaps heavily with Codex on the core things you expect from a modern coding agent. It can work directly with a local codebase, use terminal tools, edit files, run tests, and even work across multiple agents. On that baseline, the two products are much closer than they were a year ago. But beyond the basics, Devin is designed as an orchestration layer for software engineering in general, not just a single coding assistant. You can start work locally and hand it off to a cloud Devin, spin up multiple isolated agents to work on separate parts of a problem, and use Devin Desktop as a control center for different coding agents. For context, Codex is built around OpenAI models, while Devin’s broader strategy is to sit above the model layer and route work across different agents and model providers. Cognition wants Devin to remain the best place to use whichever coding models are strongest. Devin can also monitor engineering systems and triage incidents. If you want, it can even investigate issues, interact with full computer environments, and launch specialized multi-agent workflows for tasks such as security review. Setting up Devin using a local LLM You need a bunch of things to start with Devin Desktop can be paired with a locally running LLM, although you are not replacing the model inside Devin Local itself. Instead, Devin Desktop acts as the interface and agent manager, while an ACP-compatible agent connects to your local model. I am doing this through OpenCode, but the same setup will work with other coding agents as well. Devin Desktop supports the Agent Client Protocol (ACP), which lets compatible agents such as OpenCode, Codex CLI, Claude Agent, and Gemini CLI run alongside Devin in the same interface. For a fully local setup, install Ollama or another LLM serving framework and download a coding model. I am using Qwen 3.5-9B, but if your hardware supports it, you can run a better model like Qwen3-coder-30B-A3B. OpenCode supports local Ollama models and can connect to Ollama's OpenAI-compatible endpoint. It also supports locally hosted models through llama.cpp. Once the model is running, configure OpenCode to use it and verify that the model works from OpenCode itself. OpenCode can then expose itself as an ACP agent. In Devin Desktop, open Devin User Settings, go to Agents, enable OpenCode, and restart the app. Devin Desktop can then launch OpenCode as a local subprocess, while OpenCode sends inference requests to Ollama running on the same machine. This gives you the Devin Desktop experience for managing agents and code while keeping the actual model inference on your own hardware. This also means you are not limited to whatever models Cognition chooses to expose through Devin itself. Any sufficiently capable local model supported by your agent harness can potentially become another agent inside Devin Desktop. The setup works well Better if you have a powerful local model The experience is surprisingly close to using any other coding agent inside Devin Desktop. You open a repository, start a session with the OpenCode agent, describe what you want changed, and the agent can inspect files, modify code, run commands, and iterate on the result. Devin Desktop provides the surrounding workspace, so you can keep multiple agent sessions open and switch between Devin, OpenCode, and other ACP-compatible agents without changing applications. The quality of the experience depends heavily on the local model and your hardware. I have not had the best experience with Qwen 3.5-9B, especially when asking the model to perform a big task. However, the setup works really well when I break the task into smaller bits. You'd also see more latency than you might expect if your hardware is only barely capable of running the model. You're not losing anything here The good news is that a lot of newer Macs and MacBook Pros, or a decent PC with enough RAM and a good GPU, can run much better local models than a 9B model. And once you move up to a stronger model, the output gets noticeably better too. The best part is that you are not paying for every prompt or burning through API credits, so if you already have the hardware, there is not much downside to trying it. Sure, the very best cloud models will still be better for harder tasks, but getting a solid coding agent running locally for basically no ongoing cost is still a pretty good deal.