A complete, tested guide to building an always-on AI agent with its own identity, crypto wallet, and three-tier inference fallback chain -- powered by a Mac mini and the Morpheus decentralized AI network.
Every step in this guide was tested on real hardware with real money. 8 bugs were hit during the original build and fixes for all of them are documented.
- A Mac mini running headless 24/7 with a dedicated user account for the agent
- An AI agent with its own email, GitHub, crypto wallet, and phone number
- Three-tier inference fallback: cloud AI (Claude) + decentralized (Morpheus) + local (Ollama)
- A defined persona that makes the agent consistent and predictable
- On-chain financial guardrails via a Safe multi-sig wallet
- A kill switch that halts everything immediately
| Resource | What It Is | Best For |
|---|---|---|
| Overview | 4-page summary | Understanding the architecture quickly |
| Full Guide | 17-page walkthrough | Reading through the complete build |
| Claude Code Prompt | Interactive agent prompt | Paste into Claude Code and get walked through the build step by step |
The Claude Code Prompt is the fastest path. Paste it into a Claude Code session and the AI will guide you through each phase interactively -- asking what you have, adapting to your setup, and flagging gotchas before you hit them.
1. Claude Opus (cloud) -- best quality, paid subscription
2. Morpheus (decentralized) -- no per-request cost, tokens returned after use
3. Ollama (local) -- free, private, always available
If Claude is down or rate-limited, the agent falls back to Morpheus. If Morpheus providers are sparse, it falls back to local Ollama. The agent always has a way to think.
Morpheus is a peer-to-peer AI inference network. Instead of paying per-request to a cloud API, you stake MOR tokens to open a 7-day session. When the session expires, your tokens come back. You're renting access, not spending money.
No API keys. No rate limits. No single company controlling access. Your agent gets inference that doesn't depend on anyone's billing system staying online.
Software guardrails are breakable -- a prompt injection or bug could bypass them. The agent's funds live in a Safe multi-sig wallet with on-chain spending limits enforced by the blockchain itself. The agent proposes transactions; spending above the threshold requires your co-signature. The contract is the security boundary, not code the agent could override.
| Phase | What | Time |
|---|---|---|
| 0. Discovery | Define what you're building, choose a name | 30 min |
| 1. Mac mini Setup | Headless, remote access, always-on | 2-3 hours |
| 2. Agent Identity | Email, secrets, wallet, GitHub | 2-3 hours |
| 3. Communication | Signal via VoIP (JMP.chat) | 1-2 hours |
| 4. Model Routing | Cloud + Morpheus + Local fallback chain | 3-4 hours |
| 5. Persona Files | SOUL.md, AGENTS.md, MEMORY.md, etc. | 1-2 hours |
| 6. Security | Hardening checklist | 1 hour |
| 7. Operations | Trust-building, ongoing monitoring | Ongoing |
Total: about a weekend for the basic agent. Add another day for the full Morpheus integration.
| Item | Cost | Type |
|---|---|---|
| Mac mini M4 (16GB) | ~$600 | One-time |
| HDMI dummy plug | ~$10 | One-time |
| Proton Mail Plus | ~$4/month | Recurring |
| VoIP number (JMP.chat) | ~$4/month | Recurring |
| Cloud AI subscription | ~$20-200/month | Recurring |
| ETH for gas (Base) | ~$10-50 | Refillable |
| MOR for staking | ~$10 | Reusable (tokens return) |
| Ollama (local) | Free | Needs a Mac with RAM |
Total recurring: ~$28-208/month depending on your cloud AI tier. The Morpheus layer reduces dependence on paid cloud inference over time.
These are real bugs from the original build, with fixes:
- signal-cli platform mismatch -- frameworks download the wrong binary on ARM Macs. Install via Homebrew.
- Signal captcha on headless machines -- solve on your laptop at signalcaptchas.org, copy the token.
- macOS Keychain locks in SSH -- run
security unlock-keychainbefore accessing secrets. - launchctl bootstrap fails over SSH -- SSH runs in the background domain. Manual start after SSH, auto on reboot.
- Morpheus proxy-router overflow (v5.11.0) -- reset allowance to 0, let the router manage.
- Morpheus model IDs wrong out of the box -- verify against the actual network.
- Apple Silicon is fast for local inference -- M3 Ultra 96GB runs Llama 3.3 70B. Don't underestimate it.
- Heredocs corrupt over SSH -- use single-line commands or scp files directly.
- Identity separation. The agent is not you. Own email, wallet, accounts. Your credentials never touch its machine.
- On-chain guardrails. Financial limits enforced by a Safe multi-sig contract, not by software.
- Agent proposes, human approves. Every financial move explained in plain language first.
- Operational transparency. Silence looks like death. The agent communicates what it's doing.
- Kill switch. "STOP ALL IMMEDIATELY" halts everything. No arguments.
Built your own agent? Found a bug not in the guide? Open a PR.
- Add gotchas you discovered
- Share alternative framework configs
- Improve the Morpheus setup instructions as the network evolves
- Add instructions for Linux (the guide is macOS-focused)
MIT