I built Claude Code's workflow with a local model, and it's surprisingly close

Published Sep 9, 2026, 5:30 PM 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. Claude Code can inspect a repository, read and edit files, run terminal commands, and check whether its changes actually work. The model plays a huge role here, of course, but much of the experience comes from the workflow built around it. I wanted to see whether I could recreate that workflow while keeping the model on my own computer. I combined Qwen 3.5 with LM Studio and Hermes Agent. I’m running the 14B model, so it’s nowhere near as good as Claude, but this setup brought me pretty close to what I had in mind, with a few limitations, of course. LM Studio runs the model Hermes handles the actual work LM Studio loads Qwen 3.5 on my computer and makes it available through an OpenAI-compatible API. I previously used Ollama, but I prefer LM Studio because it doesn’t force me to use the command line all the time, and its Hugging Face integration comes in handy. To run the agent, I use Hermes, which connects to that API instead of using LM Studio’s chat interface. While I prefer Hermes, you can use another tool, such as OpenCode or even Claude Code. OpenAI’s Codex also supports local models. Hermes Agent provides the tools Qwen needs to work inside a project. These include functions for searching, reading, patching, and writing files, along with terminal access for running commands. When Qwen requests a tool, Hermes executes it and sends the output back to the model. For example, a coding request can move from search_files to read_file, then to patching the relevant file and running a test command, without me copying code between a chat window and an editor. Of course, this setup isn’t as convenient as Claude Code, which packages the entire workflow into one product. My version splits it into three components: Qwen decides what to do, LM Studio runs the model, and Hermes carries out the actions. But given that none of this costs me anything, I’m happy to keep using it. Getting the local workflow running You can spin this up in minutes Qwen 3.5’s 14B model runs on my M5 MacBook Pro with 24GB of unified memory. I used a 4-bit quantized build and limited the context window to 32,768 tokens, which left enough memory for macOS, Hermes, and any terminal commands running alongside the model. LM Studio can expose a loaded model through an OpenAI-compatible server. You can start it from the Developer tab, after which the model becomes available at http://localhost:1234/v1. Hermes includes LM Studio as a provider, so I only had to select it through hermes model and choose the Qwen model already loaded in memory. Hermes’ coding toolset provides file search, editing, terminal access, memory, and task planning. An AGENTS.md file can also tell the agent how your project is organized, which build commands to use, and what checks it should run after changing code. Mine requires Hermes to inspect existing files before editing them and run the relevant tests afterward. Command approvals are worth keeping enabled when the agent runs directly on your computer. I used Hermes’ smart approval mode, limited file access to the project directory, and worked from a separate Git branch, so a bad edit wouldn’t affect the main copy. It works best when the job stays focused Though the experience isn’t the same as Claude Code Focused coding jobs gave me the closest experience to Claude Code. I tested the agent with a configuration bug that caused the project to ignore a user-defined setting. Qwen searched for references to the setting, inspected the configuration loader, changed the relevant file, and ran the existing tests through Hermes. The first patch caused a fallback test to fail. Qwen read the terminal output, adjusted the condition it had added, and ran the test suite again. I didn’t have to identify the file or paste the error back into the conversation. Each step took longer than it would with Claude Code, though, especially when Qwen had to process several files or a large terminal response. You'll start seeing the limits of the 32K context window once you ask for bigger changes. You can ask Hermes to compress the conversation, but some earlier details will disappear in the process. A better approach here is to split a feature into smaller requests. Qwen also produces an invalid tool call at times, which can stop the workflow until you repeat the request. Local agents are getting better Local agents are much better than they were a few years ago. Now, if you have a decent machine, such as a MacBook Pro or even a MacBook Air with more than 16GB of RAM, you can run a very good model that can do a lot for you for free. If you have a machine with a dedicated GPU and enough VRAM, you can even run something like a 32-billion-parameter model. Paired with a tool such as Hermes, it can produce output worth considering.