DeepClaude – Claude Code agent loop with DeepSeek V4 Pro

· ai coding devtools · Source ↗

TLDR

  • Shell wrapper that redirects Claude Code’s API calls to DeepSeek V4 Pro or OpenRouter, keeping the full tool loop at up to 17x lower cost.

Key Takeaways

  • Sets ANTHROPIC_BASE_URL, ANTHROPIC_AUTH_TOKEN, and model env vars per-session so Claude Code’s file, bash, git, and subagent tools run unchanged against a cheaper backend.
  • DeepSeek V4 Pro is priced at $0.87/M output tokens vs Anthropic’s $15/M; OpenRouter route hits $0.44/$0.87 with US servers.
  • DeepSeek’s automatic context caching drops repeat-turn costs to $0.004/M, making long agent loops especially cheap.
  • Known gaps: no vision/image input, no parallel tool use, no MCP server tools through the compatibility layer.
  • Remote control mode splits traffic: WebSocket bridge stays on Anthropic infrastructure, model calls route through a local proxy to DeepSeek.

Hacker News Comment Review

  • The core env-var trick is a one-liner; commenters posted minimal shell scripts achieving the same redirect without the wrapper, questioning the project’s added value beyond backend switching and cost tracking.
  • DeepSeek V4 Pro’s $0.87/M price is a 75% promotional discount expiring 2026-05-31; post-discount rate rises to $3.48/M output, materially changing the cost story.
  • Skepticism on harness quality: per terminal-bench@2.0, Claude Code ranks among the weaker agentic harnesses; alternatives like opencode, pi.dev, and lmcli were suggested as better starting points.

Notable Comments

  • @syntex: warns DeepSeek may not be trained on Claude Code’s specific tool-call contracts, planning structure, and error-recovery patterns, degrading real agentic performance.
  • @varenc: notes the README buries the genuinely non-trivial features (mid-session model switching, unified cost tracking) and that an LLM wrote the README without surfacing them.

Original | Discuss on HN