--- title: "How Do You Control an Olares One With a Coding Agent?" url: "https://goremy.ai/blog/articles/olares-cli-control-local-ai-server-with-codex" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-11-10T00:40:11.389Z" updated: "2026-11-10T17:04:27.475Z" topic: Local AI description: "How a coding agent like Codex or Claude Code can drive an Olares One: what the setup involves, example prompts for STL files and ComfyUI, and what to verify." --- # How Do You Control an Olares One With a Coding Agent? > How a coding agent like Codex or Claude Code can drive an Olares One: what the setup involves, example prompts for STL files and ComfyUI, and what to verify. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-11-10 ## How do you control an Olares One with a coding agent? You give the agent a command-line interface to the machine, then describe what you want in plain language. The agent turns that into commands: installing apps, moving files, starting workloads. The Olares One is a local AI server you normally reach through a browser desktop, so a CLI is the obvious bridge for an agent like Codex or Claude Code. The exact Olares CLI commands and skill names are not covered in the material this article draws on, so confirm them in the official Olares documentation. ## TL;DR - The Olares One is a **local AI server with a browser desktop**, so anything you click through today is a candidate for an agent to script. - A coding agent needs a **command-line path into the box**, which is the role the Olares CLI is meant to fill. - Apps already on the system, including **ComfyUI, Blender and OpenCode**, are the natural targets for natural-language prompts. - A documented hardware build on the Olares One used **OpenSCAD, STL export and Blender preview**, which is the same file flow an agent would automate. - Version numbers, command names and skill packaging should be **checked against official docs**, not assumed from a blog post. - Local models can run on the same box, so you can choose a **cloud agent or a local one** for the driving seat. ## What is the Olares One? The Olares One is a compact local AI server that began life as a Kickstarter campaign, which has since finished. You can now buy it directly from the maker. In a recent hands-on demonstration, the creator reached it through a web browser from a Windows laptop. The interface is a minimal desktop with a dock on the left and a set of installed apps. Apps visible in that demo included Hermes Agent, Blender, ComfyUI and OpenCode. The operating system has also gained the option to be used locally as a desktop, so you can plug in a display and use it like a mini PC, including for games. Note that the demonstration was sponsored by Olares, which is worth knowing when weighing any praise. ## What does a CLI add that the web desktop does not? The browser desktop is built for a person. A CLI is built for scripts, and by extension for agents. Coding agents are good at reading command output, deciding what to do next and retrying on errors. They are clumsy at clicking through a GUI. A CLI gives an agent three things: 1. A way to **list and install apps** without you opening a store page. 2. A way to **move files** between your machine and the server. 3. A way to **check state**, such as whether an app is running, so it can verify its own work. That is the general logic. Which of these the Olares CLI exposes, and under what command names, is something to read from its own help output and docs. ## How would you prompt an agent to fetch STL files? STL files are the format 3D printers consume, and the Olares One already fits into a workflow that produces them. In the hardware demo, a local model wrote OpenSCAD files for a speedometer enclosure (an ESP32, a GPS module and a small OLED display). Those files were not print-ready until exported. The creator then asked the model to export them as STLs and render previews, and later opened the results in Blender on the Olares desktop. An agent with CLI access could handle the file movement around that. Example prompts, written as illustrations and not as verified commands: - "List the STL files in the project folder on my Olares One and copy them to my local Downloads folder." - "Export every OpenSCAD file in this project to STL and tell me the output paths." - "Open the enclosure lid STL in Blender on the server and confirm the file loads." Keep the prompts specific about paths and filenames. Agents handle "copy this file from here to there" far better than "get my model." ## How would an agent drive ComfyUI? ComfyUI is a node-based image and video generation tool, and it was among the apps installed in the demo. It is a good fit for natural-language control because its workflows are saved as files and its jobs have clear inputs and outputs. For image-to-image generation, a prompt could look like this: - "Take the render at this path, run it through my image-to-image workflow in ComfyUI with this style prompt, and save the result next to the original." - "Run the same workflow at three different denoise strengths and name the outputs so I can compare them." The first time through, expect to supply the workflow yourself, or point the agent at a saved one. An agent can fill in parameters and queue jobs, but you should check the first outputs by eye before letting it batch anything. ## Can the agent be local instead of Codex or Claude Code? Yes, and the Olares demo is a good example. The creator ran OpenCode against a local 27B-parameter Qwen model at 4-bit quantization, hosted on the Olares One itself. The model fit on a 24 GB GPU. The system dashboard showed about 20.5 GB of VRAM in use out of a little over 24 GB, while CPU use stayed low. The system also reported 96 GB of memory available overall, shared with other services. The same creator had earlier tested Claude Opus 4.5 and GPT 5.2-Codex on a physical build, and the point of the newer test was to see whether a local model could match what frontier models managed some 11 months earlier. The local model showed solid built-in knowledge of the parts involved. It could tell GPS modules apart, and it flagged that a modeled OLED bezel would not fit the actual screen, then asked for measurements. The creator's caveat was that very large codebases are where local models are weaker. | Option | Where it runs | Trade-off | |---|---|---| | Codex or Claude Code | Cloud model, commands run against your server | Stronger on hard tasks, needs network and an account | | OpenCode with a local Qwen 27B (4-bit) | On the Olares One | Private and offline-capable, tight on VRAM at 24 GB | ## Is it safe to let an agent operate your server? Treat it like giving a new contractor shell access. A few habits help: - Start with read-only requests, such as listing apps and checking status. - Ask the agent to print a command before running anything that installs, deletes or overwrites. - Keep generated outputs in a dedicated folder so mistakes are easy to spot. - Watch the resource dashboard. A model and a ComfyUI job competing for the same 24 GB of VRAM will slow or fail. ## What should you verify before relying on this? Three things were not established by the sources used here, so check them yourself: 1. The exact Olares OS release that ships the CLI (the brief mentions 1.12.7). 2. The command names, install steps and any agent "skill" packaging for Codex or Claude Code. 3. Which authentication method the CLI uses to reach your device. The official docs and the CLI's own help output are the authority on all three. ## Frequently Asked Questions ### Do I need Codex or Claude Code specifically? No. Any coding agent that can run shell commands can in principle use a CLI. The Olares One can also host a local model that works with a tool like OpenCode, as one demonstration showed. ### Can an agent really move STL files and run ComfyUI for me? The pieces exist on the system: ComfyUI and Blender are installable apps, and STL export from OpenSCAD worked in a documented build. Whether one prompt chains them together depends on the CLI's capabilities and your workflow setup. ### Will a local 27B model be good enough to drive the box? For focused tasks it can be. A 4-bit Qwen 27B handled a multi-part hardware project and showed accurate part knowledge. It is a weaker choice for large, complex codebases. ### Is the Olares One only usable through a browser? No. It is typically accessed through a browser, but the OS now also supports use as a local desktop with a connected display.