DeepClaude – Claude Code agent loop with DeepSeek V4 Pro

Original Article ↗ Hacker News Discussion ↗

Purpose of DeepClaude / Using DeepSeek with Claude Code

  • Main idea: route Claude Code’s CLI through DeepSeek V4 (Pro/Flash) or other models via Anthropic-compatible/OpenRouter endpoints and env vars.
  • Several comments show one‑liner shell scripts achieving this; some note this has been possible “since the beginning” via Anthropic_BASE_URL + token + model env vars.
  • DeepClaude adds a local proxy that can switch models mid‑session and track combined costs across Anthropic and non‑Anthropic models, though this is under‑emphasized in its README.

Is DeepClaude Necessary?

  • Some see the repo as trivial “env-var glue” unworthy of a full project or HN post.
  • Others value a packaged solution that streamlines setup and highlights the pattern.
  • There’s confusion about tool-call compatibility; clarifications note DeepSeek exposes an Anthropic-compatible endpoint but it’s missing some features.

Alternative Harnesses and Tools

  • Many suggest ditching Claude Code entirely in favor of open-source harnesses: OpenCode, Pi, Hermes, Forge Code, Langcli, custom CLIs, etc.
  • Opinions differ on quality: some find OpenCode less effective than Claude Code; others claim Forge Code or Codex-style harnesses outperform Claude Code in benchmarks.
  • Claude Code is praised for rich plugins/skills and MCP ecosystem, criticized for regressions, cost-optimization at quality’s expense, and being closed-source.

Model Quality and Use Cases

  • DeepSeek V4 Pro/Flash are widely reported as strong coding models, often “close to” or better than Claude Sonnet and approaching Opus for some tasks, at much lower cost.
  • Other favored models: GLM 5.1 (surprisingly strong, sometimes near Opus), GPT 5.5 + Codex for some languages, Gemini Flash for speed, Kimi/Qwen/MiniMax for cost.
  • Some argue you need the best model for planning, architecture, and debugging; others say “good enough” cheaper models suffice, especially in mixed workflows (e.g., premium for design, cheaper for implementation).

Cost, Subsidies, and Data Privacy

  • Cost engineering is a major theme; some report burning tens of dollars quickly even with cheap models.
  • DeepSeek’s very low pricing is noted as heavily discounted and time-limited; debate over whether inference is subsidized.
  • Repeated concerns about data training/retention, especially with Chinese providers; mitigations include OpenRouter’s zero‑data‑retention flags and alternative inference providers.
  • Some distrust US firms less, others point out US data protections are also limited; geopolitical tensions color perceptions.

Harness Design, Contracts, and Local Inference

  • Building robust agentic harnesses is described as harder than expected: models must reliably honor tool formats, planning structures, permissions, and error recovery; weaker models expose harness fragility.
  • Local and hybrid setups (llama.cpp, Ollama, custom harnesses, DGX‑like boxes, upcoming DeepSeek V4 local support) are seen as key for privacy and long‑term cost control.

Meta: LLM Slop, Benchmarks, and OSS Process

  • Strong backlash to “vibe‑coded” LLM-generated repos that are minimal but heavily promoted; some call for age/history gates on HN projects.
  • Benchmarks (Terminal Bench, LiveCodeBench) are cited but also criticized as unrepresentative or vulnerable to “cheating.”
  • Separate thread on OSS projects auto‑closing issues to cope with low‑quality, LLM‑generated bug reports and PRs.