A local Qwen–Ollama–OpenCode setup built a usable screenshot-renaming Mac app, but its PDF-organizing attempt stalled during debugging. David Gewirtz reported those results from a test on his 128GB M4 Max Mac Studio; they describe his tasks and setup, not a general performance benchmark.
What the local setup built
Gewirtz paired OpenCode v2.2.5, the coding agent, with Ollama, which runs models, and qwen3-coder-next:q4_K_M, the coding model. He reported a 64,000-token context and described the model as having 80 billion parameters in total, with 3 billion active at a time. His Mac Studio was the machine he used, not a stated minimum requirement.
The clearest success was a utility for renaming screenshots. It opened a folder, displayed an image, let him enter a filename, and saved it before moving to the next image. The project took almost six days. Gewirtz spent four days trying to get the file list working, discarded that code, then asked a fresh OpenCode session to build the list. The finished app lacked features such as zoom, resizable panes, filename templates and quick next/previous navigation.
Why the PDF organizer stalled
The same setup made limited progress on a PDF organizer: it could identify the subject of an individual PDF, but Gewirtz said the task did not work reliably across larger batches. His five-day attempt became bogged down in debugging a new macOS 27 on-device document-understanding system.
That contrast matters when judging the result. The screenshot tool was a narrow task with a clear sequence; the PDF project involved debugging a newer system feature and repeating the work at scale. Gewirtz’s account shows what happened in those particular projects, not how every local coding agent will perform.
Treat a local coding agent as supervised work: keep tasks narrow and approve file changes.
Gewirtz said he kept tasks small, restarted sessions for individual fixes, and required approval for file and shell actions. Those are practices from his test, not a guarantee that an agent will handle every action safely.
Another local CLI, with a different setup
Local-CLI is a separate coding-agent command-line tool. Its current repository documents an MIT license, a first-run endpoint wizard, planning and execution features, and a supervised mode that asks for approval before file modifications. The documented installation requires Node.js 20 or later:
npm install -g local-cli-agent
local-cli
| Workflow | Coding agent | Model connection | What changes |
| Gewirtz’s reported test | OpenCode v2.2.5 | Qwen through Ollama, running locally | Replaces Claude Code’s agent and model provider with a local stack |
| Local-CLI | Local-CLI | A local, on-premises or cloud OpenAI-compatible endpoint | Uses a separate CLI and lets the user configure its endpoint |
Local-CLI’s repository lists local options such as Ollama, vLLM and LM Studio, alongside cloud services including Azure OpenAI and Google Gemini. Choosing a cloud endpoint is a different setup from using a local model. The repository describes the tool’s features; Gewirtz’s account does not test Local-CLI.
