Ollama local inference · LaunchAgent 24/7 · Memory Tree · TCO
Who is this for? In 2026, OpenClaw (MIT, channel-first autonomous agent) and OpenHuman (GPL-3.0, desktop super-assistant with Memory Tree) both support fully local inference via Ollama—but you still need a macOS host that stays awake. A laptop sleeps, a Linux VPS cannot run OpenHuman’s Tauri UI, and buying an M4 upfront costs thousands before you prove the workflow. Bottom line: renting a dedicated physical Mac Mini M4 (SSH + VNC, often live in under ten minutes) is the fastest path to a private, always-on agent stack. In this guide: why rent · product comparison · RAM tiers · node checklist · OpenClaw / OpenHuman / Ollama steps · cost table · security · FAQ. Pair with our rent vs buy TCO article and OpenClaw + Ollama hybrid setup.
The competitive question shifted from “which cloud API is strongest?” to “how do I run an agent continuously, privately, and without token bills that scale linearly with usage?” OpenClaw and OpenHuman both assume stable processes, writable disks, and background reaction to messages or desktop events.
24/7 uptime: Telegram/WhatsApp channels and OpenHuman’s periodic memory sync expect no lid-close sleep or home-router dropouts.
Native macOS: LaunchAgent daemons, TCC screen-recording prompts, and browser CDP automation break on headless Linux SSH-only hosts.
Apple Silicon inference: 16GB unified memory handles 7B–13B models comfortably; M4 Pro with 64GB reaches 70B-class quantized weights with Metal-backed Ollama.
Cash flow: Buying an M4 is roughly $599–$2,000+ upfront; VNCMac-style rental from about $19.8/day or ~$196/month turns CapEx into cancellable OpEx—ideal for a 60–90 day proof.
Numbers you can cite: a Mac Mini M4 idles around 4–6W for 24/7 duty; cloud GPU instances often cost $2–5/hour, while a local 7B model on M4 commonly lands near 30–45 tokens/s depending on quantization and thermals.
Renting also solves the “wrong continent” problem: pick a VNCMac region close to your team so SSH and VNC latency stay predictable, then snapshot configs before you offboard. You are not buying a shared VM slice—providers in this category ship non-virtualized Mac mini hardware, which matters when Metal-backed Ollama performance is the whole point of the experiment.
| Dimension | OpenClaw (MIT) | OpenHuman (GPL-3.0) |
|---|---|---|
| Shape | CLI + 20+ messaging channels | Tauri v2 desktop (Rust + React 19) |
| Memory | SOUL / MEMORY files + plugin ecosystem | Memory Tree, Obsidian vault, ~20 min tool sync |
| Local AI | Ollama via openclaw onboard | local_ai.* in config.toml (Ollama / LM Studio) |
| Voice / meetings | Mostly plugins | Native STT/TTS, Google Meet agent (stronger in v0.54+) |
| Runtime | Node.js ≥ 22 (v24 recommended) | Install script; 16GB+ RAM recommended |
On a rented Mac Mini M4, share one Ollama service at 127.0.0.1:11434. Set OLLAMA_MAX_LOADED_MODELS=1 so two 13B weights do not exhaust 16GB unified memory.
| Config | Models | Dual-agent note |
|---|---|---|
| M4 · 16GB | Qwen2.5 7B, Llama 3.2, Gemma3 small | OpenClaw channels + light OpenHuman local |
| M4 · 24GB | Phi-4 14B, Qwen 14B quant | One resident 13B + Gateway headroom |
| M4 Pro · 48–64GB | 32B–70B quant, model rotation | Parallel experiments without buying a Studio |
EU/US readers weighing GDPR-style privacy should default sensitive threads to Ollama in openclaw.json and OpenHuman settings instead of shipping prompts to third-party APIs.
Pick region, RAM tier, and billing on the pricing page; store SSH keys and VNC URL.
First VNC session: confirm automatic date/time (OAuth and certs depend on it).
Install CLT: xcode-select --install.
Pre-approve TCC: screen recording, accessibility, microphone for voice features.
Firewall: expose OpenClaw Gateway port 18789 only behind TLS reverse proxy or IP allowlists.
curl -fsSL https://openclaw.ai/install.sh | bash
openclaw onboard --install-daemonThe onboard wizard wires LLM provider (Ollama), Telegram/WhatsApp tokens, and ~/.openclaw/openclaw.json. --install-daemon registers LaunchAgent for login persistence—critical on rental nodes you keep online 24/7. After upgrades, run openclaw doctor --fix when configs drift.
Minimal Telegram test: BotFather token → onboard → phone /start → verify routing in the VNC console. Silent failures usually trace to heartbeat/thinking routes—see our no-reply troubleshooting posts on the blog index.
Hardening: openclaw security audit --fix; bind Gateway to 127.0.0.1 unless TLS termination is in place.
curl -fsSL https://raw.githubusercontent.com/tinyhumansai/openhuman/main/scripts/install.sh | bashStable tag v0.53.22 (May 9, 2026); mainline v0.54.x adds local voice and IDE bridges. Complete OAuth for Gmail/Notion/Slack inside the VNC browser. Local AI ships disabled until you opt in:
[local_ai]
runtime_enabled = true
opt_in_confirmed = true
provider = "ollama"Memory Tree tips: let the agent ingest tool activity automatically; pair with an Obsidian vault if you want human-readable archives. v0.54+ local voice needs microphone TCC plus a small summary model such as gemma3:1b pulled in Ollama.
brew install ollama
brew services start ollama
ollama pull qwen2.5:7b
ollama pull llama3.2:latestFor always-on agents, tune launchd env vars: OLLAMA_KEEP_ALIVE=30m, OLLAMA_MAX_LOADED_MODELS=1, OLLAMA_NUM_PARALLEL=2. OpenClaw declares model IDs under models.providers.ollama; OpenHuman points local_ai.provider to the same endpoint. Avoid running OpenHuman voice mode and OpenClaw browser MCP stress tests simultaneously on 16GB.
| Option | Year-one cash (order of magnitude) | Local 13B | 24/7 fit |
|---|---|---|---|
| Buy M4 16GB | ~$599–$800 + power | Yes | You own downtime risk |
| VNCMac monthly | ~$196 × active months | Yes (physical) | Upgrade RAM tier without new hardware |
| Cloud GPU VM | $2+/hr adds up fast | Yes, data leaves box | No macOS LaunchAgent/GUI path |
| API-only | Token-linear | No | Weak privacy story |
Three takeaways: (1) A 90-day experiment with both agents plus Ollama usually costs less than an idle purchased Mini. (2) Three-year guaranteed uptime at a fixed RAM tier can favor buying. (3) Heavy API users who move chat to local 7B/13B often beat twelve months of tokens versus rental—same economics as our Hermes 24/7 host article.
Use openclaw secrets / SecretRef; never commit Telegram tokens. OpenHuman cloud subscription can coexist with local models—route sensitive tickets to Ollama by default.
Authenticate any public Gateway; enterprises may set OPENCLAW_PROXY_URL.
Before offboarding, archive ~/.openclaw, OpenHuman config, Obsidian vault, and Ollama model list (weights are re-pullable; configs are not).
Yes with one resident sub-13B Ollama model. Rent 48GB+ for 32B parallel work.
OpenClaw is mostly SSH-friendly; OpenHuman OAuth and TCC still need VNC for first-time setup.
Hermes targets Nous Skill evolution; OpenClaw targets channels; OpenHuman targets desktop memory. All want the same 24/7 macOS host—you can isolate directories on one rented Mini.
OpenClaw wires agents into the messaging apps you already use; OpenHuman stitches memory across desktop tools; Ollama removes per-token billing for the conversations you can keep local. The hard part is not the install script—it is keeping a macOS machine awake without melting your laptop or fighting headless Linux permissions.
Buying a Mac Mini M4 makes sense after you prove year-round load. Renting a physical M4 node first—finishing install and TCC over VNC—lets you convert CapEx into an experiment you can stop. VNCMac offers dedicated Mac Mini M4 hardware, full SSH/VNC, and RAM tiers you can raise when 13B is no longer enough.
Agent value compounds with uptime—give it a Mac that does not sleep. Open the Mac Mini M4 plans when you are ready.