Moltbot-Rob Integration
Deploy a local AI coprocessor (Mac Mini M4) as an employee Operator Node, governed by Rob. Enhance employee workflow locally while gating critical actions through centralized intelligence.
Laptop as primary workspace, Mac Mini as headless AI coprocessor. Recommended baseline: M4 Pro 48GB. Total investment ~$2,579 per employee.
Executive Summary
The Moltbot concept deploys a Mac Mini M4 as a local AI brain for each employee. The Mini handles inference, browser automation, and workflow learning locally, while Rob serves as the central governor — issuing scoped execution grants, brokering external APIs, and logging all actions. The employee works from their laptop as usual; the Mini operates as a silent coprocessor.
Why This Matters
A Moltbot pays for itself within 6–12 months compared to equivalent cloud GPU costs ($2,400–$6,000/yr). Zero data leaves the premises, zero per-token API costs for local inference, and always-on availability regardless of internet quality.
Architecture Overview
The Moltbot operates as a governed node in the Rob ecosystem. It receives tasks from and reports status to the central Rob platform. All high-risk actions require Rob approval via scoped, short-lived execution tokens.
Moltbot (Front-End / Operator)
- Grant-aware execution: checks high-risk actions against threshold, requires Rob approval
- Secure Rob command channel: persistent outbound connection, cert-based identity, message queue
- Local workflow learning: tracks apps, sites, sequences, patterns
- Stores structured outputs locally, sends summaries to Rob
Rob (Back-End / Governor)
- Operator registry: track devices, employees, permissions, tools, revocation
- Execution Grant Service: scoped, short-lived tokens (scope, duration, limits, domains)
- Broker APIs: handles external tools/keys (CRM, email, payments, travel, sheds)
- Audit & telemetry: log actions, metrics, health; standardize outputs
Mac Mini M4 Configuration Lineup
Apple's 2024 Mac Mini redesign is 5.0 × 5.0 × 2.0 inches and weighs just 1.5–1.6 lbs — smaller than a hardcover book. Built-in power supply (no external brick), silent operation, and Apple Silicon unified memory make it ideal for an always-on AI coprocessor that moves between office and home.
| Configuration | Chip | CPU / GPU | RAM | Storage | MSRP | Notes |
|---|---|---|---|---|---|---|
| Base M4 | M4 | 10C / 10C | 16GB | 256GB | $599 | Too tight for AI |
| Mid M4 | M4 | 10C / 10C | 16GB | 512GB | $799 | RAM-limited |
| M4 + RAM | M4 | 10C / 10C | 24GB | 512GB | $999 | Budget sweet spot |
| M4 Pro Base | M4 Pro | 12C / 16C | 24GB | 512GB | $1,399 | Pro chip, same RAM |
| M4 Pro + RAM + SSD | M4 Pro | 12C / 16C | 48GB | 1TB | $1,999 | RECOMMENDED |
| M4 Pro Max Bin | M4 Pro | 14C / 20C | 48GB | 1TB | $2,199 | Top-bin CPU upgrade |
| M4 Pro Ultimate | M4 Pro | 14C / 20C | 64GB | 1TB | $2,399 | Maximum RAM; 70B models |
Avoid: 8GB & Intel Macs
16GB is the absolute floor for any local LLM. At 16GB, the system will constantly swap to SSD, degrading performance and SSD lifespan. Intel Macs lack unified memory architecture — on Apple Silicon, GPU and CPU share the same RAM pool, meaning a 48GB M4 Pro gives models access to the full 48GB. Intel Macs also lack the Neural Engine and have vastly worse performance-per-watt.
What Models Fit by RAM Tier
| RAM | Models That Fit | Speed | Use Cases |
|---|---|---|---|
| 24GB | Llama 3.2 8B, Qwen2.5 14B, DeepSeek-R1-Distill 7B/14B, Mistral 7B, Phi-3 14B | 15–30 tok/s | Code completion, summarization, translation, basic reasoning |
| 48GB | All above + Qwen2.5 32B, Llama 3.1 32B, DeepSeek-R1-Distill 32B, Mixtral 8x7B, CodeLlama 34B | 8–15 tok/s | Complex reasoning, long-context coding, multi-model serving |
| 64GB | All above + Llama 3.3 70B (Q4), DeepSeek-R1-Distill 70B (Q4) | 3–8 tok/s | Near-frontier reasoning, complex analysis |
Key Insight
RAM is more important than chip variant for AI workloads. A base M4 with 24GB outperforms an M4 Pro with 16GB for model inference. The M4 Pro's extra GPU cores help with inference speed, but you cannot run a model that doesn't fit in RAM. Buy RAM first, chip speed second.
Physical Specs
Accessories & Peripherals
The employee uses this as a workstation at the office but takes it home where they work most of the time. The accessory setup should support both locations.
Monitors
| Product | Size | Resolution | Price | Best For |
|---|---|---|---|---|
| Samsung ViewFinity S7 (S70D) | 27" | 4K | ~$220 | Best budget 4K desk monitor |
| Dell S2725QC | 27" | 4K USB-C | ~$300 | Single-cable USB-C |
| ARZOPA 16" Portable | 16" | 2.5K | ~$100 | Budget portable for second location |
| ViewSonic VP16-OLED | 15.6" | 1080p OLED | ~$300 | Pro color accuracy portable |
Keyboard & Mouse
| Product | Price | Notes |
|---|---|---|
| Logitech MX Keys S + MX Master 3S (Mac) | ~$170 | RECOMMENDED — Multi-device pairing (switch between Mini & laptop instantly), superior ergonomics |
| Apple Magic Keyboard (Touch ID) + Magic Mouse | ~$278 | Native macOS, Touch ID for auth |
Carrying & Cables
| Item | Price | Purpose |
|---|---|---|
| Lacdo Hard Mac Mini M4 Case | ~$30 | EVA hard shell, fits Mini + all accessories |
| Thunderbolt 4/5 cable (0.8m) | ~$40 | Thunderbolt Bridge to laptop |
| Cat6a Ethernet cable (6ft) | ~$10 | Direct Ethernet |
| USB-C cable | ~$15 | Portable monitor connection |
| HDMI 2.1 cable (6ft) | ~$12 | Backup monitor connection |
Connection Options
The Moltbot connects to the employee's laptop and to the central Rob server. Different connection methods serve different needs.
Mini ↔ Laptop (Local Tethering)
| Method | Speed | Latency | Setup | Cost | Best For |
|---|---|---|---|---|---|
| WiFi (same LAN) | 100–600 Mbps | 2–10ms | None | $0 | Quick API calls, lightweight use |
| Ethernet (direct/switch) | 1 Gbps | <1ms | Low | $10–$20 | Reliable, low-latency traffic |
| Thunderbolt Bridge | ~10 Gbps | <0.5ms | Medium | $30–$40 | File transfer, high-bandwidth inference |
| 10Gb Ethernet (BTO) | 10 Gbps | <0.5ms | Medium | +$100 | Sustained throughput at scale |
Remote Desktop / Screen Sharing
For times when the employee needs visual access to the Mini's desktop (model management, debugging, etc.):
| Solution | Latency | Quality | Cost | Notes |
|---|---|---|---|---|
| Parsec | 4–8ms LAN | Excellent (H.265) | Free / $8/mo | Best latency. Hardware-accelerated. |
| Tailscale + Apple Screen Sharing | Medium | Fair | Free | Zero-config mesh VPN. Punches through NATs. |
| RustDesk (self-hosted) | Low–Medium | Good | Free | Open-source, self-hostable relay |
| Apple Screen Sharing | 50–150ms | Fair | Free | Built-in, zero setup. LAN only. |
Recommendation
Tailscale (mesh VPN, always on, free tier) for network connectivity between Mini and Rob. Parsec for low-latency screen sharing when visual access is needed. For the primary Moltbot use case (API-level communication), Tailscale alone is sufficient — the Mini's services are reachable over the Tailscale IP, no screen sharing needed.
Local AI Software Stack
Model Runners
| Tool | Type | License | Best For |
|---|---|---|---|
| Ollama | CLI + REST API | Open source | Primary engine — headless, API-driven, perfect for coprocessor pattern |
| LM Studio | GUI + API | Free (closed) | Interactive model browsing, MLX optimization |
| MLX (Apple) | Framework | Open source | Maximum Apple Silicon optimization, lowest RAM overhead |
Browser Automation
| Tool | Description | Integration |
|---|---|---|
| Playwright MCP | Microsoft's browser automation via Model Context Protocol. LLMs drive browsers through structured commands. | Ollama/LM Studio via MCP |
| Stagehand | AI-native browser automation SDK. Natural language + code hybrid. | Direct API |
| Browser Use | Python framework bridging AI agents and browser control. | Ollama API compatible |
Workflow Automation
| Tool | Description | Runs On |
|---|---|---|
| n8n (self-hosted) | Open-source workflow automation, 400+ integrations, native AI nodes connecting to Ollama. Docker-based. | Mac Mini via Docker |
| Open Interpreter | Runs code locally via natural language. Controls OS, terminal, browser. | Python, local |
| Claude Code / Aider | AI coding assistants operating as autonomous agents. | CLI, local or API |
Recommended Stack
n8n (workflow orchestration) + Ollama (local inference) + Playwright MCP (browser automation). This gives a complete local AI automation pipeline with zero cloud dependencies for routine tasks.
Security & Governance Integration
Authentication Layers
| Approach | Description | Complexity |
|---|---|---|
| Tailscale ACLs | Device identity via WireGuard keys. ACLs restrict which Moltbots can reach Rob's API. Zero-config. | Low |
| mTLS (Mutual TLS) | Both Rob and Moltbot present X.509 certificates. Rob's CA signs Moltbot's cert at provisioning. Defense in depth. | Medium |
| SSH Certificates | Rob CA signs short-lived SSH certs for command execution channels. | Medium |
Message Queue Patterns
| Pattern | Protocol | Latency | Best For |
|---|---|---|---|
| MQTT | MQTTS (8883) | Low | Heartbeats, command dispatch, device telemetry. IoT-native. |
| WebSocket | WSS | Low | Real-time task streaming, interactive AI sessions. |
| HTTP/REST | HTTPS | Medium | Simple task dispatch and result retrieval. |
| Redis Pub/Sub | TCP | Very low | Fast message passing (Rob already uses Celery/Redis). |
Permission Gating (CML Mirror)
| Layer | Mechanism | Description |
|---|---|---|
| Device Registration | Provisioning Token | One-time token from Rob admin. Moltbot uses it to register and receive its device certificate. |
| Session Tokens | JWT with Scopes | Short-lived JWTs (1h expiry): inference.local, browser.automate, file.access, tool.execute |
| Permission Tiers | CML Levels | CML-0 (read-only queries), CML-1 (local actions), CML-2 (external actions requiring approval) |
| Rate Limiting | Token Bucket | Max 100 browser actions/hr, max 10 external API calls/hr per device |
Setup Options
Three physical configuration options, depending on employee workflow and IT maturity. The employee in question works primarily from home, with time at the office — the Mini travels between both locations.
Option A
Laptop as primary workspace, Mac Mini as headless AI coprocessor. Employee works normally on their laptop; Mini runs headless, serving API requests and executing workflows.
Pros
- Employee keeps familiar laptop workflow
- Mini can be purpose-built and locked down for AI
- Runs 24/7 even when laptop sleeps
- Offloads heavy inference from laptop
- Easy to upgrade without disrupting employee
- Centrally manageable by IT
Cons
- Two devices to maintain
- Requires network between devices
- Initial setup more technical
Option B
Mini as full workstation with monitor, keyboard, mouse. Employee remotes in from laptop when away from desk via Parsec or Screen Sharing.
Pros
- Single primary device to manage
- All compute in one place
- Simpler software stack
Cons
- Requires good internet for remote use
- 50–80ms WAN latency on remote sessions
- Can't work offline from laptop
- Peripherals needed at each location
Option C
Mini controls laptop apps — acts as an AI agent that observes and automates the employee's applications. The most advanced option, best deferred to v2.
Pros
- True AI assistant model
- Learns employee workflows
- Most "futuristic" option
Cons
- Highly complex to build
- Cross-device security implications
- Needs robust error handling
- Deferred to future roadmap
Rollout Playbook
Procure Hardware
- Order Mac Mini M4 Pro (48GB RAM, 1TB SSD) — $1,999
- Order Logitech MX Keys S + MX Master 3S combo — ~$170
- Order Samsung ViewFinity S7 27" 4K monitor (office) — ~$220
- Order ARZOPA 16" portable monitor (home) — ~$100
- Order Lacdo carrying case + cables — ~$90
Base Configuration
- macOS setup: create managed admin account, enable FileVault encryption
- Install Tailscale, join to Rob's tailnet with ACLs configured
- Install Docker Desktop for Mac
- Install Ollama, pull recommended models (Qwen2.5 32B, DeepSeek-R1-Distill 14B)
- Install Parsec for remote desktop access
- Verify Mini is reachable from employee's laptop over Tailscale
Rob Integration
- Register Moltbot in Rob's operator registry (device ID, employee, permitted tools)
- Deploy MQTT client for heartbeat/status reporting
- Configure mTLS certificates (Rob CA signs Moltbot's cert)
- Set up Execution Grant Service endpoints on Rob side
- Deploy n8n via Docker with Ollama integration
- Configure Playwright MCP for browser automation
Test & Validate
- End-to-end test: employee laptop sends inference request to Mini via Tailscale
- Test Rob governance: Moltbot requests execution grant, Rob approves/denies
- Test portability: Mini moves from office to home, reconnects via Tailscale
- Test remote desktop: Parsec from laptop to Mini over WAN
- Benchmark local inference speed and validate model quality
Employee Onboarding
- Set up employee's desk at office: monitor, Mini, peripherals
- Set up home station: portable monitor, carry case protocol
- Train employee on daily workflow: how to interact with Moltbot from their laptop
- Document standard operating procedures
- Monitor usage for first 2 weeks, tune models and workflows
Cost Summary
Per-Employee Investment
Recommended Setup Breakdown
| Item | Cost |
|---|---|
| Mac Mini M4 Pro, 48GB RAM, 1TB SSD | $1,999 |
| Samsung ViewFinity S7 27" 4K (office desk) | $220 |
| ARZOPA 16" portable monitor (home) | $100 |
| Logitech MX Keys S + MX Master 3S for Mac | $170 |
| Thunderbolt 4/5 cable (0.8m) | $35 |
| Ethernet cable (Cat6a, 6ft) | $10 |
| USB-C cable (portable monitor) | $15 |
| Lacdo carrying case | $30 |
| TOTAL | ~$2,579 |
Annual Cost Comparison
| Comparison | Annual Cost |
|---|---|
| ChatGPT Team (per seat) | $300/yr |
| Claude Pro (per seat) | $240/yr |
| GitHub Copilot Business | $228/yr |
| Moltbot (amortized 3 years, recommended) | ~$860/yr |
| Cloud GPU instance (comparable) | $2,400–$6,000/yr |
ROI Analysis
The Moltbot pays for itself within 6–12 months compared to cloud GPU costs. Additional benefits: zero data leaves the premises, zero per-token API costs for local inference, always-on availability regardless of internet quality, and the hardware has a 5+ year useful life with Apple Silicon's longevity.