For Claude Code users, this repo is interesting because it isn’t trying to replace the whole experience — it’s trying to keep the workflow and swap out the model backend. That’s a very practical kind of hack: if the tool loop is the valuable part, maybe you don’t always need to pay for the most expensive brain attached to it.
deepclaude is a wrapper/proxy that lets you use Claude Code’s autonomous agent loop with non-Anthropic backends.96.4% on LiveCodeBench and is dramatically cheaper than Anthropic’s Opus-tier pricing.ANTHROPIC_BASE_URLANTHROPIC_AUTH_TOKEN--backend ds)--backend or)--backend fw)--backend anthropic)
What strikes me is how pragmatic this is. A lot of AI tooling talk is about “agents” in the abstract, but this repo is about preserving the thing developers actually feel day to day: the loop, the terminal, the edits, the commands, the momentum. I think that’s the right framing. If Claude Code is the interface you like, then the model is increasingly just one interchangeable component.
The cost angle is obviously the hook, and it’s a strong one. A cheaper agent loop is genuinely attractive if you’re doing lots of iterative coding, refactoring, or repo exploration. I’d be curious whether the real-world experience feels as smooth as the README suggests once you hit messy codebases, longer tool chains, or tasks that depend on stronger reasoning. The repo itself does not hide that Claude Opus can still be better on harder problems, which I appreciate.
I also like the live switching idea more than I expected. That’s the kind of feature that sounds small but could matter a lot in practice: use the cheap backend for routine work, then flip to Anthropic when the problem turns thorny. That feels more honest than pretending one model is optimal for everything.
What worries me a bit is the usual compatibility layer tradeoff. “Anthropic-compatible” is useful, but it’s not the same as native support, and the README already notes gaps around vision, MCP tools, and prompt caching behavior. So this is exciting, but not magical. If your workflow leans heavily on Claude-specific ecosystem features, this might be more of a partial substitute than a drop-in replacement.
If I were a Claude Code user, I’d probably try this for everyday coding sessions and keep Anthropic as the escape hatch. That seems like the real promise here: not replacing Claude Code, but making it cheaper to use more often.
