DEV Community

dtzp555-max
dtzp555-max

Posted on

I built a local proxy so Claude Code/OpenClaw can use your Pro/Max subscription — no API key needed

#claude #devtools #openai #productivity

Background

I've been using OpenClaw with Claude Opus 4.6 and Sonnet 4.6 for coding tasks. These are hands-down the best models for agentic coding — strong reasoning, clean code output, no hallucination loops.

The problem: Claude doesn't support OAuth. To use Claude models via any third-party tool, you need an API key. And running Opus 4.6 on heavy agentic tasks burns through API credits fast.

Meanwhile, my Claude Pro/Max subscription sits there with unlimited usage... doing nothing for my local agents.

The Fix: claude-proxy

A lightweight local proxy that intercepts OpenClaw's API calls and routes them to the Claude CLI — which runs under your existing subscription.

How it works:

OpenClaw agent → localhost:3456 (claude-proxy) → claude CLI → Claude Pro/Max
Enter fullscreen mode Exit fullscreen mode

No API key. No billing changes. Full Opus 4.6 access.

Setup (3 steps) 

# 1. Make sure Claude CLI is installed and logged in
claude --version

# 2. Clone + auto-configure
git clone https://github.com/dtzp555-max/openclaw-claude-proxy.git
cd openclaw-claude-proxy
node setup.mjs

# 3. Set default model
openclaw config set agents.defaults.model.primary "claude-local/claude-opus-4-6"
openclaw gateway restart
Enter fullscreen mode Exit fullscreen mode

That's it. All your agents now run on Opus 4.6 at zero additional cost.

Model selection UI (included)

You get a TUI picker to switch between claude-haiku-4, claude-sonnet-4-6, and claude-opus-4-6 at runtime — no config file editing needed.

Why this matters

A lot of people downgrade to cheaper models to save API costs, then blame their prompts when results get worse. The model is the bottleneck.

I tested every major model for agentic coding tasks. My ranking:

  • Claude Opus 4.6 — best for complex multi-step tasks
  • Claude Sonnet 4.6 — great balance of speed and capability
  • GPT-5.4 / Codex — solid but pricier per token
  • Cheaper models — frustrating: vague answers, loops, contradictions

If you're already paying for Claude Pro/Max, there's no reason to run your agents on a degraded brain.

GitHub

👉 https://github.com/dtzp555-max/openclaw-claude-proxy

Happy to answer questions. Works on macOS, Linux support coming.

Top comments (0)