Claude Max subprocess as zero-cost LLM runtime

Every LLM call in your codebase goes through subprocess.run(["claude", "-p", "--output-format", "text", "--model", model, "--system-prompt-file", path,...]) rather than the Anthropic SDK. No ANTHROPIC_API_KEY anywhere. Claude Max subscribers ($200/mo flat) run unlimited workloads — L1 classifiers, L2 enrichment, L3 deep research with WebSearch+Bash — at zero marginal cost. A ~80-line llm.py wrapper module is copy-paste-portable across projects.

For internal tools and small-team automation, per-token billing distorts what's worth building. A founder-discovery CLI and a deal-intake bot share the same claude -p wrapper and run 4000 tweets/day classification plus ~40 Opus deep-research memos a day at $0 marginal cost. Same pattern works for any team running L1/L2/L3 pipelines, batch translation, content rewriting, or agentic workflows. Beyond a few thousand calls/month, the Max subscription pays for itself.

1. Install Claude Code globally: npm i -g @anthropic-ai/claude-code; claude auth login with a Max-tier account.
2. Write llm.py:claude_call(prompt, model='sonnet', system_file=None, allowed_tools=None, max_turns=1) that wraps subprocess.run([...]) and returns stdout.
3. Add JSON-extraction helper for prompts that should emit JSON (look for the first {–} block; one retry with "produce JSON only" prompt on parse fail).
4. Migrate your existing API calls to this wrapper.
5. For tool-using runs (WebSearch, Bash, Read), pass --allowedTools WebSearch --allowedTools Bash --max-turns 25 --timeout 420.
6. Watch the Max plan's rate limits — they exist but are generous; throttle batch jobs to 2–4 parallel.
7. Keep an SDK fallback for one specific edge case (e.g., when you need streaming responses) — the rest of the codebase stays on claude -p.

• Don't ship Max-authenticated calls into customer-facing production where each end-user fires LLM calls — the Max plan is for you and your team, not your users. Customer-facing usage needs a metered API key.