Use a local model in chat
Pick a local model in the composer and use it with the agent, Claude Code, Codex CLI, or OpenCode.
Once a model is installed, using it is one click in the composer. The same picker works whether the chat runs on the agent, Claude Code, Codex CLI, or OpenCode.
Pick a local model
- Click the model name at the bottom of the composer.
- The picker opens with a Local section at the top, listing every model in your library. Below it are Cloud and any providers you have added.
- Choose a model. It starts the moment you send your first message, so there is nothing to launch by hand.
Mach Studio remembers your choice: new chats open with the model you used last.
Catalog models you have not installed yet appear in the Local section too, with a download icon next to them. Click it to install from the picker and the chat switches to that model when the download finishes.
Use it with Claude Code, Codex CLI, and OpenCode
Local models are not limited to the built-in agent. The Harness menu, just below the model row, lets you move a chat to Claude Code, Codex CLI, or OpenCode. The model picker still shows your Local section, so you can run any of them against a model on this computer, with every prompt and response staying on your machine.
Pick the harness, pick a local model, and send. See Switching harnesses for how handoff between harnesses works.
Tips
- Agentic coding works best with models that advertise Tool use in Browse. Filter for that capability when choosing a model to run with a harness.
- Models with Vision can read screenshots and images you drop into the composer.
- If a model feels slow, pick a smaller one or close other memory-hungry apps. The Server page under Local models shows live throughput and memory use.
- To change the model's default sampling or context length, open Runtime options from its card in the library.