Rug-Proof Architecture Sidesteps Subscription Bans

Anthropic abruptly blocked third-party tools like OpenClaw from Max subscriptions, affecting 135,000 users and forcing API pay-per-token for heavy workloads. Nick Puru rebuilt a full multi-agent system directly on Claude Code—Anthropic's terminal-based coding agent—included in Pro ($20/mo) or Max ($100-200/mo) plans. No API keys, no extra fees, no rug-pull risk since it's their own product.

Core setup: Telegram as chat interface (free, phone/desktop). Messages hit a lead agent (orchestrator) that reads shared memory files (soul.md for personality: 'direct, no fluff, proactive'; user.md for personal info; memory.md for long-term context; daily markdown logs). Lead decides: handle simple tasks itself or delegate via JSON files in /shared/tasks/ to specialists like research (web search, analysis) or content (posts, scripts in your voice). Agents communicate via shared tasks folder, all powered by subprocess calls to Claude Code. Heartbeat script runs every 30 minutes: scans logs/memory, notifies via Telegram only on novel insights (24-hour dedupe). tmux keeps bot.py, heartbeat.py, and task manager running in parallel panes, auto-restarting.

Each agent is a folder with two config files (claude.md instructions, soul.md personality). Scales to 100+ agents identically—no rework. Total cost: just your subscription. Trade-off: Relies on terminal/tmux (beginner-friendly with copy-paste prompts); Claude Code self-heals errors via plain-English fixes.

"Now, nobody can pull the rug on me because I'm using their own product exactly how it was designed to be used." — Nick Puru, explaining why Claude Code dodges bans while OpenClaw failed.

Plain-English Prompts Build Everything in Minutes

No coding required: Claude Code generates/runs/fixes via natural language. Start in empty folder (e.g., 'second-brain'):

  1. Prompt for foundation: Describes memory system, bot.py (user-ID gated, reads memory before responding, appends logs), heartbeat.py, basic skills (web search, file creation), start.sh launcher. Claude creates files, installs venv Python deps (python-telegram-bot, etc.).
  2. Test: Get Telegram bot token/user ID from BotFather/@userinfobot. Run start.sh in tmux. Send message—bot responds, logs exchange.
  3. Upgrade to multi-agent: New prompt in fresh terminal pane: Create /agents/lead/, /agents/content/, /agents/research/. Lead instructions: 'For content tasks (posts/newsletters), JSON-delegate to content; research to research.' Symlinks shared memory/tasks. Update bot.py to route via lead. Add tmux task-manager pane. Smoke test end-to-end.

Full prompts/community commands in Nick's free Skool group. Uses Cursor IDE optional for extensions/terminal. If bot token conflicts (e.g., Claude plugins), kill via pkill -f telegram. Claude self-diagnoses: 'Heartbeat pane empty? Mid-call or error—fixes incoming.'

Demo results: 'LinkedIn post about why business owners should stop relying on third-party AI tools' → Lead dispatches research agent (scrapes Anthropic ban news) → Delegates summary to content → Polished post back in Telegram. Added user info ('Nick, Reprise AI, YouTube channel') auto-updates user.md.

"You can do all of this yourself. Build out this agent and your own open claw in a single afternoon without much work. You don't have to learn how to code or anything like that." — Nick Puru, promising no-dev accessibility after live build.

Specialization Without Complexity Scales Effortlessly

Single-agent v1 (weeks prior, linked video) handled basics. v2 adds delegation: Lead as 'boss' checks memory/files, JSON-tasks specialists (e.g., {"task": "research Anthropic ban", "agent": "research"}). Content agent mimics voice from memory. Research pulls web data. All share soul/user/memory—no siloed state.

Expands easily: Add /agents/leadgen/ or /agents/sales/ with same pattern. Channels beyond Telegram possible. Heartbeat prevents notification spam. Logs readable markdown for audits.

Trade-offs: Terminal-bound (not web UI); polling-based (tmux keeps alive, but laptop sleep kills—use VPS for 24/7). Max plan recommended for heavy use (faster, higher limits). No vision/tools beyond basics (add via skills prompts). Self-healing shines: Claude Code iterates 'describe, build, test, fix.' Built live on-camera: foundation 20 mins, multi-agent upgrade similar.

Production proof: Runs thousands of tasks/mo on subscription alone. Vs. API: Predictable cost, unlimited under fair use. Vs. OpenClaw: Customizable, ban-proof.

"Claude code, it is Anthropic's own product. So they can't ban their own product." — Nick Puru, highlighting the key insight missed by most after the ban.

Results: Production-Ready in Hours, Zero Code

Live demo: 3-pane tmux (bot/heartbeat/task-manager). Complex query uses all agents seamlessly. Post-ban resilience: Every call subprocesses Claude Code—no external deps. Nick's background: 2 years implementing AI for companies, scaling agencies. Free resources: Skool (16k members), newsletter for prompts/updates.

Replicate: Claude Pro/Max, Telegram, terminal. 1 afternoon → Delegating multi-agent 'second brain' for research/content/automation. Future-proof as Anthropic iterates Claude Code.

"The approach. So anything that I describe it's just going to be what I want in plain English and claude code. It will be building it all for me." — Nick Puru, on the 'describe-build-test-fix' loop that powers no-code builds.

Key Takeaways

  • Start with Claude Code on Pro/Max: Zero extra cost, ban-proof vs. third-party wrappers like OpenClaw.
  • Telegram + tmux = Persistent, multi-pane agent runtime; copy-paste setup in empty folder.
  • Lead agent JSON-delegates via /shared/tasks/: Simple scaling to 100+ specialists sharing memory.
  • Prompts over code: 'Describe memory/bot/heartbeat' → Claude builds/tests/fixes; join Skool for exact text.
  • Heartbeat every 30min scans logs/memory for proactive Telegram pings (24h dedupe).
  • Trade-off terminal for power: VPS for always-on; self-healing minimizes babysitting.
  • Test incrementally: Bot token from BotFather, user ID from @userinfobot, venv for isolation.
  • Customize voices/skills per agent folder: soul.md personality + claude.md rules.