OpenClaude: Free Open Source Claude Code Alternative Runs Any AI Model For Free

本站添加: 2026-05-06

[00:00:00]Anthropic's Claude code is probably the strongest coding agent out there right now, but you have to pay for it and you are stuck with whatever model Anthropic gives you. A new open-source project called Open-Claude just changed that. It takes the same Claude code workflow, the same tool calling, the same terminal interface, and lets you run it with any model you want. Open AI, Gemini, Deep Seek, local Ollama, even your existing GitHub Copilot subscription. And in some tests, the output actually looks better than Claude code itself. Open-Claude is a community project on GitHub and it already has around 25,000 stars. So, this is not some random side project.

[00:00:43]The repo itself is clear that it started from the Claude code code base and got rewritten to support multiple providers.

[00:00:50]It is not affiliated with Anthropic at all. The name throws people off, but think of it like this. Claude code is the closed door version. Open-Claude is the open door version where you bring your own model and your own keys. The big idea is one CLI for everything. You install it once and inside the same terminal, you can flip between Open AI compatible APIs, Gemini, GitHub models, Codex, Ollama for local stuff, and Atomic Chat for Apple Silicon.

[00:01:21]It uses a simple {slash} command called {slash} provider to set this up. And you can save profiles so you do not have to re-enter keys every time.

[00:01:30]Same workflow as Claude code, but you control the brain behind it.

[00:01:35]Now, let me put this in context with other open-source coding agents because Open-Claude is not the only one in this space. You probably heard of Open-Code.

[00:01:44]That one has around 152,000 stars, supports about 75 providers through models.dev, and has things like LSP integration and multi-session support. Aider is another popular one.

[00:01:57]Then there is Codex CLI from Open AI, Gemini CLI from Google, Klein, and a few others. Each has its own angle. So, why pick Open-Claude over those? Honestly, it comes down to one thing. If you already love how Claude code works, the prompts it understands, the way it plans tasks, the way it handles tool loops, then Open-Claude gives you that exact feel without the lock-in. Claude code is closed source and tied to Anthropic billing. Open-Claude sits right in the middle. Familiar interface, no provider lock, no subscription required. All right, let me show you the actual install. You only need two things first, node.js and bun. Once those are ready, the easiest install is just one line.

[00:02:42]Type npm install {dash} g at gitlab/open-claude.

[00:02:47]That pulls the package globally so you can run it from anywhere. If you want the source build, you can clone the repo, run bun install, then bun run build. The repo also has a Windows quick start, a Mac and Linux quick start, and even an Android install guide if you want to run this on a phone, which is honestly wild. Oh, one heads-up, if the install warns about Ripgrep not found, just install Ripgrep on your system and confirm rg {dash} {dash} version works.

[00:03:17]Open-Claude uses it for fast file search. Now, the part I actually want to focus on, the provider setup. This is where Open-Claude really opens up. Once you type Open-Claude in your terminal, you get an interactive shell. From there, you have a few options. You can run {slash} provider for guided setup with saved profiles. You can run {slash} onboard {dash} GitHub if you want to use your existing GitHub Copilot subscription, which is what I am doing.

[00:03:43]Or you can just export environment variables before launching.

[00:03:48]For Open AI or any Open AI compatible API, you set three envars. Claude code use open ai equals one, your Open AI API key, and the model name.

[00:03:59]That same setup works for Open Router, Deep Seek, Grok, Mistral, LM Studio, basically anything that exposes a {slash} v1 endpoint. For local Ollama, you point the Open AI base URL at localhost on port 11434, pick a local model, and you are running fully offline with zero API cost. For me though, the GitHub Copilot route makes the most sense because I already pay for Copilot. So, I run {slash} onboard {dash} GitHub inside Open-Claude, and then I run {slash} model and pick Claude Sonnet 4.6 from the list.

[00:04:36]Now, I am running the same Sonnet model that Claude code uses, but through my Copilot subscription instead of paying Anthropic separately. That is a real cost saving if you are already on Copilot.

[00:04:48]Gemini works the same way through {slash} provider, and the repo says it supports API keys, access tokens, and even local ADC workflow. GitHub models gets its own onboarding flow with saved credentials. Codex uses your existing Codex credentials if you have them.

[00:05:05]There is also agent routing, which is cool. You can route different agents to different models. So, your explore agent runs on cheap Deep Seek, your plan agent runs on GPT-4o, your front-end dev agent runs on something else. You configure that in your settings.json file under agent models and agent routing.

[00:05:25]One more thing worth mentioning. Web search and web fetch work out of the box on non-Anthropic models using DuckDuckGo as the free fallback. That means Deep Seek, Gemini, even local Ollama models can browse the web inside Open-Claude.

[00:05:41]If you want a stronger setup, you can plug in a Firecrawl API key. The free tier gives you 500 credits and then web fetch handles JavaScript rendered pages properly. Okay, time for the actual test. I gave both Open-Claude and Claude code the exact same prompt. Build a single HTML file with three.js that renders a rotating 3D Rubik's Cube. Auto rotate on X and Y. Soft gradient background, smooth lighting, and a floating label below the cube. Same Sonnet model on both sides, same prompt, just different agents.

[00:06:18]Let me show you the Open-Claude result first. The file got generated really fast, like noticeably fast. I hit enter, the agent planned, wrote the code, saved the HTML, and it was done. I opened it in the browser and look at this. You see how clean this is? The cube has a glossy white finish on top, soft orange tinting on one side, blue on the other. The lighting is doing real work here. It looks almost like a render out of Blender. The faces have that subtle reflection, the edges are smooth, and there is even an outline glow around the whole shape. The label at the bottom says built with Open-Claude in a clean pill shape. And the best part, I can grab it with my mouse and rotate it manually. That was not even in the prompt. Open-Claude added orbit controls on its own. Now, the Claude code version. Same prompt, same Sonnet model.

[00:07:10]It took longer to generate. That part I can confirm. The output runs, but look at the difference. Flat red and blue faces, solid gray top. No specular highlights, no real depth, no orbit controls, so I cannot drag it with my mouse. The label says built with Claude code, which is fine, but the whole thing looks like a basic three.js tutorial example. It works, but it does not impress. Same model, same prompt, different result. So, why did this happen?

[00:07:40]The thing is, Claude code goes through Anthropic's API with their own system prompts and their own scaffolding wrapped around the model.

[00:07:48]Open-Claude talks to the model more directly through whatever provider you picked with its own scaffolding that seems to push the model toward more complete, more polished output. There is research backing this up, by the way.

[00:08:01]Tembo did a comparison earlier this year and found that the same model can score very differently across agents because the scaffolding around the model matters as much as the model itself.

[00:08:12]So, the takeaway, if you want a coding agent that feels exactly like Claude code, runs any model you want, costs nothing extra if you already have Copilot or Open Router or just a free Ollama setup, and sometimes produces nicer output than the real Claude code, Open-Claude is genuinely worth installing. It is not perfect. Smaller local models still struggle with long tool chains, and if a provider has lower output caps, the agent has to adapt. But for daily coding work, this is one of the most flexible setups I have tried.

[00:08:44]All right, so that's it from the video and I hope you enjoyed it. If you did, please like this video and subscribe to the channel and I'll see you in the next video.