Skip to content
Giant Antique Postage Stamp style editorial illustration for the news article: DeepClaude: run Claude Code's agent loop on DeepSeek V4 Pro or any Anth
LaunchIndustryDeveloper

DeepClaude: run Claude Code's agent loop on DeepSeek V4 Pro or any Anthropic-compatible backend

By Harsh Desai
Share

TL;DR

DeepClaude proxies Claude Code's autonomous agent loop to DeepSeek V4 Pro, OpenRouter, or Fireworks, claiming roughly 17x cheaper runs versus standard Claude pricing.

What changed

DeepClaude is an MIT-licensed local proxy (port 3,200) that intercepts Claude Code's API calls and routes them to cheaper Anthropic-compatible backends: DeepSeek V4 Pro by default, plus OpenRouter, Fireworks, and Anthropic itself. The full Claude Code surface (file reads and writes, bash, subagent spawning, multi-step autonomous loops) keeps working unchanged. Backends can be switched live mid-session with slash commands or CLI flags, and the proxy tracks per-call costs.

Why it matters

For developers running long-horizon agent loops, the cost ceiling on Claude Code at scale gets uncomfortable fast. DeepClaude exposes the exact same UX while letting a developer dial down to a cheaper backend when the task does not need frontier reasoning. The README claims a light-usage profile drops substantially with a fallback backend.

What to watch for

The interesting variable is parity, not pricing. Claude Code's behaviour depends on subtle prompt and tool-call patterns the underlying model has been trained on; DeepSeek V4 Pro is a strong model but not a drop-in clone. Worth tracking which Claude Code workflows degrade quietly on the cheaper backends, and whether DeepClaude evolves heuristics to route per-task rather than per-session.

Who this matters for

  • For developers paying for Claude Code at scale: a local proxy on port 3200 that routes to DeepSeek V4 Pro, OpenRouter, or Fireworks with the full Claude Code surface (file ops, bash, subagent spawning, multi-step loops) working unchanged.
  • For agent builders running long-horizon tasks: the live mid-session backend switching means you can route trivial tasks to cheaper models and reserve frontier reasoning for the parts that need it.
  • For anyone tracking Claude Code parity benchmarks: this is the test bed. DeepSeek V4 Pro is a strong model but not a drop-in clone, and the workflows that break first will signal what Claude Code actually trained against.

Harshs take

DeepClaude is the kind of project that exists because Claude Code's pricing curve forces it to exist. The question is not whether developers will route Claude Code through a cheaper backend; they obviously will. The question is which workflows degrade quietly versus loudly when you swap the underlying model.

Anyone considering this should run their actual agent loops against both backends side-by-side for a week before committing, because the failure modes are subtle. Claude Code's behaviour depends on patterns the underlying model has been trained on. DeepSeek V4 Pro is a strong model but not a drop-in clone, and the subtle degradations only show up on the workflows you actually care about.

by Harsh Desai

Source:github.com

More AI news

Everything AI. One email.
Every Monday.

New tools. Model launches. Plugins. Repos. Tactics. The moves the sharpest builders are making right now, before everyone else.

No spam. Unsubscribe anytime.