DeepClaude

Overview

DeepClaude is the cleanest current demonstration of the agent-loop-vs-model decoupling. The Claude Code harness already respects ANTHROPIC_BASE_URL; DeepSeek already ships an Anthropic-compatible endpoint at api.deepseek.com/anthropic. DeepClaude is the four-env-var bridge plus a small Node proxy that handles edge cases like sub-agent routing and remote-task auth.

The pitch on the repo is “same UX, 17× cheaper” — which is empirically right on cache-heavy agent-loop workloads. On benchmark, DeepSeek V4 Pro is within 0.2 points of Claude Opus 4.7 on SWE-bench Verified, beats Sonnet 4.6 on Terminal-Bench and LiveCodeBench, and trails Opus 4.7 noticeably on multi-file architectural reasoning and SWE-bench Pro.

Why It Matters

The reason DeepClaude is the #1 HN story today isn’t the cost savings (those have existed via OpenRouter and Ollama for a while). It’s the cleanness of the swap. Three exports, one binary that doesn’t need to exist, and the entire Claude Code experience runs on a competitor’s brain at ~6% of the price. The implication for the broader market — that the harness is the product and the model is a swappable component — is the structural shift the cluster of Cursor SDK, Browserbase Skills, and OpenAI Apps SDK launches has been building toward all month.

For agent buyers, the read is: default to DeepClaude, route the architecture-heavy 20% of work to Anthropic. The harness is the constant, the brain is the variable.

Setup

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=sk-deepseek-...
export ANTHROPIC_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
claude

The fourth-and-fifth-and-sixth env vars handle sub-agent routing — without them, sub-agents will try to call Anthropic-named models on the DeepSeek endpoint and 404.

Use Cases

• Heavy feature-implementation workloads where SWE-bench Verified parity matters but architectural reasoning isn’t on the critical path.
• Indie/startup teams that can’t afford $500-2000/mo Claude API bills.
• Steady high-volume production agents that benefit from DeepSeek’s lack of rate-limit cliff.
• Any team comfortable routing data to PRC infrastructure (the major caveat).

Limitations

• DeepSeek’s Anthropic-compatible endpoint is policy-controlled by a single Chinese company and can be repriced or restricted.
• Computer Use and browser-based sub-agents work but are noticeably weaker than on Anthropic or Codex Pro.
• Architecture-heavy multi-file refactors trail Opus 4.7 quality.
• Compliance review for enterprise: data leaves the local machine and goes to PRC servers.

Pairs With

• Claude Code — the harness DeepClaude plugs into.
• DeepSeek-TUI — Rust-native alternative if you don’t want Claude Code in the loop.
• context-mode — multiplicative cost savings.

Verdict

The cleanest agent-loop cost-cutting move available in May 2026, and a structural signal that the substrate-vs-model decoupling has gone mainstream. Test it for one week on a normal workload before committing — the cost of being wrong is one Saturday afternoon and $5 of DeepSeek credits.