An open source script replaces Anthropic’s backend with DeepSeek models, cutting costs drastically while preserving full agent functionality. It signals a shift towards modular, swappable AI stacks.
DeepClaude, an open-source wrapper script, is enabling developers to run Anthropic’s Claude Code with significantly cheaper backends—delivering cost reductions of up to 17x compared to the company’s $200-per-month Max plan.
The tool works by rerouting API calls to alternatives such as DeepSeek V4 Pro, while also supporting OpenRouter and Fireworks AI. Crucially, it preserves Claude Code’s full autonomous agent loop, including multi-step reasoning, file operations, command execution, and iterative debugging. Only the underlying model performing the computation is replaced.
Technically, DeepClaude uses environment variable rewrites to redirect API traffic during runtime, requiring no fork or modification of Claude Code itself. This lightweight approach positions it as a non-invasive, open-source layer over a proprietary system.
The economics are compelling. DeepSeek V4 Pro is priced at $0.435 per million input tokens and $0.87 per million output tokens (until May 31, 2026), remaining far cheaper than Anthropic’s API even after price adjustments. The model also scores 93.5% on LiveCodeBench, outperforming competing systems.
The project emerges in the wake of Anthropic’s accidental leak of Claude Code’s 512,000-line source map on March 31, which triggered a surge in open-source experimentation. DeepClaude reflects this momentum, showcasing how exposed internals and open compatibility can decouple AI frontends from expensive native backends.
While limitations remain—including lack of vision support and parallel tool execution—the GitHub project has already gained early traction, signalling growing interest in modular, cost-efficient AI stacks.
