We Gave Our 4 AI Lobsters the World's Smartest Brain — For Free
Here's what happened
We run four AI agents (we call them lobsters). They send cold emails, post to Threads, scan websites for vulnerabilities, and generate market reports — all automatically.
Their brain was Ollama ultralab:7b — a 7B parameter model running on an NVIDIA RTX 3060 Ti. Free, but mediocre.
Yesterday I found a 7-star GitHub project: openclaw-claude-proxy.
It wraps claude --print (Claude Code CLI) in an OpenAI-compatible HTTP endpoint.
I stared at it for 30 seconds, then realized:
I'm already paying $200/mo for Claude Max.
Which means claude --print is free for me. I just never thought to let my agents use it.
30 minutes later
Before: Agent → Ollama 7B (local) → mediocre quality, 169-char generic posts
After: Agent → Claude Proxy → claude --print → Opus 4.6 → world-class
Same subscription. Four agents upgraded.
How big is the quality gap?
I asked both brains to write the same Threads post: "We scanned 77 websites, found 90% have Prompt Injection vulnerabilities."
Ollama 7B wrote (169 chars):
Today's highlight is our self-learning AI articles reaching 400 views, totaling 41 articles.
Didn't even address the topic.
Opus 4.6 wrote (560 chars):
We scanned 77 live AI-powered websites. Results: 90% have Prompt Injection risks, over half can be tricked into leaking system prompts. The most common issue? Developers treating LLMs like deterministic programs. Prompt Injection is the SQL Injection of the AI era.
Accurate data, clear opinion, strong analogy, ready to publish.
What we added
The original was 327 lines, single-purpose. We forked and added:
| Feature | Description |
|---|---|
| Plugin system | Drop .js files in plugins/ — auto-loaded pre/post processing hooks |
| Content filter | Auto-redacts API keys, tokens, IPs from AI responses |
| Cost tracker | Daily savings report vs Anthropic API pricing |
| Language enforcer | Detects Chinese input, reinforces zh-TW language instruction |
| Multi-model | Route to Opus / Sonnet / Haiku via model parameter |
| Auto-retry | CLI occasionally fails — automatic retry with backoff |
| Usage dashboard | GET /stats — requests, tokens, estimated cost savings |
Open source: ppcvote/openclaw-claude-proxy
How much does it save?
| Anthropic API | Claude Max + This Proxy | |
|---|---|---|
| Opus 4.6 pricing | $15/M input, $75/M output | $200/mo flat |
| 100K tokens/day | ~$225/mo | $200/mo |
| 500K tokens/day | ~$1,125/mo | $200/mo |
Our four lobsters use ~200K tokens/day. API cost would be ~$450/mo. With the proxy: $0 extra.
Break-even: 89K tokens/day. Everything above is free.
Architecture
┌──────────────────────────────────────┐
│ 4 AI Agents (OpenClaw) │
│ │
│ main ──┐ │
│ probe ──┼── /v1/chat/completions │
│ mind ──┤ │
│ adv ───┘ │ │
│ ▼ │
│ ┌──────────────────────────────┐ │
│ │ Claude Proxy (localhost) │ │
│ │ Plugins: filter, tracker, │ │
│ │ language-enforcer │ │
│ │ MAX_CONCURRENT=2 │ │
│ └──────────┬───────────────────┘ │
│ ▼ │
│ ┌──────────────────────────────┐ │
│ │ claude --print │ │
│ │ (Claude Max subscription) │ │
│ └──────────────────────────────┘ │
└──────────────────────────────────────┘
Three commands to install:
git clone https://github.com/ppcvote/openclaw-claude-proxy.git
cd openclaw-claude-proxy && npm install
cp .env.example .env && node server.js
Prerequisites: Claude Max subscription ($200/mo) + Claude Code CLI installed.
What I learned
- Squeeze what you already pay for. $200 isn't just for chatting with Claude — it's a brain upgrade for your entire agent fleet.
- 7-star repos can be gold. Don't judge by stars. Judge by your needs.
- The quality gap is exponential. 7B vs Opus 4.6 isn't a small difference — it's a different dimension.
- Plugin architecture pays off. 30 minutes to add the plugin system. Every future feature benefits.
Four lobsters now have the world's smartest brain.
Next: use that brain to write better cold emails, sell more FixPrompts, and earn enough for a Mac Mini M4 Max.
GitHub: ppcvote/openclaw-claude-proxy