DeepClaude Lets You Run Claude Code With DeepSeek’s Brain for 17x Cheaper

A new open-source tool called DeepClaude lets developers run Claude Code—Anthropic’s autonomous coding agent—with DeepSeek’s model under the hood.

Pushed to GitHub by a developer going by aattaran, the project bills itself as the same Claude Code experience at “17x cheaper.” It’s a simple bash and PowerShell script — nothing fancy, no fork, no rewrite.

Using the tool, you can also run Claude Code with other backends like OpenRouter, or Fireworks AI—not just DeepSeek, though that’s apparently the main draw.

The trick is environment variables. Claude Code reads a handful of them to decide where to send API calls. DeepClaude rewrites those for the duration of a session, points them at a cheaper backend, then restores the originals when you exit.

The command line tool itself doesn’t change. File reading, editing, bash execution, multi-step tool loops, subagent spawning, all of it still works. Only the model doing the thinking is different.

But the “tool loop” is the whole point of Claude Code. When you ask Claude to fix a bug, it doesn’t answer in one shot. It reads files, runs commands, sees the output, decides what to do next, edits something, runs your tests, reads the failures, tries again. So it’s a process that involves thinking, acting, observing, thinking again, dozens of times, until the job is done.

That cycle is what people mean by “autonomous agent loop.” You give one instruction, the model takes thirty steps. It’s also the part that burns through tokens—and through your Anthropic budget.

DeepClaude keeps the loop intact and just swaps which model is making the decisions inside it.

The math is straightforward. Anthropic’s Max 20x plan, which gives heavy users full Claude Code access, runs $200 a month. DeepSeek V4 Pro through OpenRouter currently costs $0.435 per million input tokens and $0.87 per million output—a promotional rate that expires May 31, 2026, after which it doubles. Even at full price, it’s drastically cheaper than Anthropic’s API.

By default, DeepClaude routes traffic to DeepSeek V4 Pro, the 1.6 trillion-parameter open-weight AI behemoth model that Hangzhou-based DeepSeek released on April 24. (Parameters, for the uninitiated, are what determine a model’s capacity—the more parameters, the wider the model’s breadth of knowledge.)

DeepSeek V4 Pro scores 93.5% on LiveCodeBench—a benchmark that runs models on fresh competitive programming problems, meaning DeepSeek solved 93.5% of the tasks—placing it ahead of both Gemini 3.1 Pro and Claude Opus 4.6.

Some things are still a bit rough around the edges. Image input doesn’t work. DeepSeek’s Anthropic-compatible endpoint doesn’t support vision. Parallel tool calls are disabled. MCP server integrations don’t pass through. And on hard reasoning tasks, the “readme” file itself concedes Claude Opus is still stronger.

That said, for reasoning, coding, and creative tasks, this implementation is a very powerful and efficient solution to use.

A bit of context for why this exists: On March 31, Anthropic accidentally shipped Claude Code’s full source map to npm, exposing 512,000 lines of TypeScript. The leak triggered a wave of clones, Python rewrites, and tools that hook into Claude Code’s now well-documented internals.

DeepClaude is the natural next step. It doesn’t fork the code, but it exploits the fact that Claude Code’s backend is, by design, swappable.

There’s also a “remote control” mode that opens a Claude Code session in any browser via a claude.ai/code/session_… URL, with DeepSeek still doing the model work. That one requires a Claude.ai subscription, since Anthropic’s WebSocket bridge is hardcoded.

Whether Anthropic will care is another question. Claude Code is a paid product, but the API endpoints are open by design, the company built them to be Anthropic-compatible for third-party adoption. DeepClaude turns that compatibility outward.

The repo currently sits at almost a thousand stars and two forks. DeepSeek’s promotional pricing ends May 31.