DeepClaude: Run Claude Code Agent Loop on DeepSeek V4 Pro

May 4, 2026 — DeepClaude is the most discussed AI dev tool launch on Hacker News today. The concept: keep Claude Code's excellent agent loop (file editing, code search, shell execution, linting) but route LLM inference through DeepSeek V4 Pro/V4 Flash. Result: Opus-level reasoning at Haiku-level pricing.

What It Is

DeepClaude (github.com/aattaran/deepclaude) is a thin shell script + local proxy that configures Claude Code to talk to DeepSeek's API instead of Anthropic's. The idea is simple: Claude Code has arguably the best agent loop on the market, but Opus is expensive. DeepSeek V4 Pro delivers comparable reasoning quality — at a fraction of the cost.

The killer feature: a local Python proxy (port 8080) that tracks costs across all sessions and switches models mid-session (V4 Pro for reasoning, V4 Flash for simple edits).

How It Works: 4 Lines of Shell

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic

export ANTHROPIC_AUTH_TOKEN=sk-...

export ANTHROPIC_MODEL=deepseek-v4-flash

exec claude "$@"

DeepSeek hosts an Anthropic-compatible endpoint (/anthropic). Claude Code speaks native Anthropic protocol, DeepSeek translates it to the V4 architecture. The proxy (src/deepclaude_proxy.py) adds two things:

1. Session-wide cost tracking (in-memory + log file)
2. Live model switching during a session

HN Discussion Highlights

Pro: "I honestly can't tell the difference from Opus, but I'm paying a fraction." — Multiple users report using DeepClaude for a week since V4 launched, with identical output quality.

Con: "Four lines of shell script as #1 on HN is slop." — Counter-argument: the proxy with cost tracking and model switching is the real value, but it's buried in a README that looks LLM-generated.

Privacy debate: DeepSeek doesn't offer training opt-out. Workaround: OpenRouter with data_collection: "deny" or HuggingFace Inference (Together AI, DeepInfra) — most commercial providers don't train on customer data. One HN comment: "Why do people ask about data security with DeepSeek, Kimi, GLM but never with Anthropic, OpenAI, Google? US data protection protects US citizens only — foreigners' data isn't protected, and companies aren't even required to disclose when they collect it."

DeepSeek V4 Pro is ~5x cheaper than Opus, V4 Flash ~10x cheaper than Haiku.

Why This Matters

DeepClaude demonstrates that LLM API layers are becoming increasingly interchangeable. The best agent loop (Claude Code) is no longer tied to the most expensive model provider. This will put pressure on pricing across the board — good for developers and for platforms like Best-AI.org that sort tools by quality, not brand.

For Best-AI.org: this belongs in AI Dev Tools / Cost Optimization.

• Price: Free (open source) + DeepSeek API costs
• GitHub: github.com/aattaran/deepclaude — 479★ and climbing today
• Category: Developer workflow enabler, not a GUI tool