Run Claude Code free by pairing it with a local model you install on your own machine. Four methods compared, from simple prompting to a model dropdown. Start today.
How to run Claude Code free without losing your cloud models
You run Claude Code for free by adding a local model next to it: install the model on your own machine, tell your agent it exists, and prompt the agent to delegate work to it. The four methods below, from a video by Julian Goldie published on 2026-10-06, all preserve the cloud models rather than replacing them.
First, a definition. A local model is an AI model that runs entirely on your own computer. It costs nothing per prompt, works offline, and keeps your code private because nothing leaves the machine. Claude Code, Anthropic agentic coding tool that runs in the terminal, and Codex, OpenAI's coding agent, can both operate such a model for you.
The transcript describes a member, Chad, who already linked a local coding model through Codex and wanted that model to appear inside his existing agent setup as an extra option. His goal shapes all four methods: keep the big cloud models available, add a free local choice beside them.
Method 1: install a local model and let your agent command it
Method 1 is the simplest way to run Claude Code free: install a local model, then tell your agent to use it. You can do this inside Codex or inside Claude Code, and you can even ask the agent to perform the installation for you.
In the video, the example model is an LFM2-family model from Liquid AI with about 2.6 billion parameters, described in the transcript as "LFM2.5 2.6B". Any local model works; this one was chosen because it was already installed. Tools like Ollama or LM Studio are common ways to get a local model running.
The workflow is prompt-driven. You type something like "create a web page using my local model". Claude Code receives the request, passes the job to the local model, and the local model builds the page. You stay in control the whole time through ordinary prompts. This description is the speaker's first-hand demonstration, not an independent benchmark.
Method 2: a local engine inside an agent operating system
Method 2 gives the local model a permanent home: you set it up as a local agent engine inside what the video calls an agent OS, a home base where multiple AI agents live and share one setup. Claude Code and Codex can then use that engine directly at any time.
The difference from Method 1 is structural. Instead of mentioning a model ad hoc in a prompt, the engine is a standing part of your environment. Agents plug into the shared home rather than each being configured from scratch. The video presents this as the author's own product, so treat the workflow description as vendor information.
One honest limitation from the speaker: there is no dropdown menu in Method 2. You do not click a button to switch models. Your agent operates the engine for you, which some people prefer and others find inconvenient.
Method 3: a model dropdown in an open-source harness
Method 3 is the only approach in the video that puts your models in a dropdown menu inside one app. A harness, in the speaker's terminology, is the app that wraps around a model: Claude Code is a harness, Codex is a harness, and the video describes a third, open-source one built around DeepSeek, the Chinese AI lab's open-weight model family.
The setup works through a customization or "create a mode" feature: you ask the app to expose your local models in the model selector, then add your cloud models alongside them. After that, the local models and the cloud models all appear as options in one dropdown, and you pick per task.
The speaker is direct about the trade-off: more customization means more setup complexity. If you are new to local models, start with Method 1 or 2 and return to the dropdown approach later. Note that the specific harness app is named only in the video and could not be verified against a primary source on 2026-10-10, so confirm the tool yourself before installing anything.
Method 4: a free API window, no local model needed
Method 4 requires no local model at all. The video mentions a free API, accessed through a service the transcript calls "Token Harbor", with a limited open window. You set the API up, let Codex control it, and use it while the free period lasts.
This method is the weakest documented of the four. The service name comes only from the video transcript and could not be matched to a verifiable primary source. Because free API windows are time-limited by nature, verify that any such offer still exists before building on it, and never route sensitive code through a service you cannot vet.
Which method should you pick?
Pick Method 1 if you want the fastest free setup, Method 2 if you want a permanent agent home, Method 3 if you insist on a visual model dropdown, and Method 4 if a verified free API window is open right now.
| Method | What you install | Switching | Setup effort | Evidence source |
|---|---|---|---|---|
| 1. Agent prompts local model | One local model | Via prompts | Lowest | Speaker demo |
| 2. Local engine in agent OS | Agent OS + local engine | Agent-operated, no dropdown | Medium | Vendor (video author) |
| 3. Open-source harness | Harness app + models | Dropdown in one app | Highest | Video, unverified app |
| 4. Free API | API access only | In the agent | Low | Video, unverified offer |
A simple first week, as suggested in the video: day one, pick Method 1 or 2. Day two, get a local model running and ask your agent to handle the install. Day three, repeat the web-page test and watch the hand-off. Day four, check the free API if it is still open. Day five, try the harness dropdown if you want one.
FAQ
- Is a local model really free to run? Yes in the sense that there is no per-prompt charge: the model runs on your own hardware using your own electricity. You still pay for the hardware and, for larger models, need enough memory to hold the weights.
- Do I have to give up Claude Code to use a local model? No. All four methods add the local option beside Claude Code rather than replacing it. The cloud models stay available whenever you want them.
- Does a local model work offline? Yes. Once the weights are on your machine, inference needs no internet connection, which also keeps your code private during the run.
- Which method has a dropdown to switch models? Only Method 3, the open-source harness described in the video. Methods 1 and 2 switch models through prompts or agent operation, with no visual selector.
- Can I trust the free API offer from the video? Verify it first. The service named in the transcript could not be checked against a primary source, and free windows expire, so confirm availability before you depend on it.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits