By Adam Jenner, 10+ years building automation for ASIC, Westpac, and Australia's major banks. Founder of Phoenix AI. · github.com/adamjen · adamjenner.com.au
Build Your Local
AI Agent Team
What's inside: Chapter 1 and the whole book free this week — Pi: Building Local AI Agents — just give your email (then $9.95, free for Prime members). Plus a ready-to-run config bundle to run a multi-agent AI coding harness entirely on your own hardware.
Four years ago I was learning coding with ChatGPT, asking questions, then copying and pasting code locally, all manually. Two years ago, things changed. I started running AI locally: Ollama, LM Studio, Cline. I no longer needed to copy and paste code; I could do it all in one place. My goal was always the same: get to the point where local models actually work, and can also do the work, with no subscriptions, no API keys, no monthly bills. Four months of deep work building the Pi harness means you can skip the painful trial and error. You get something that works from day one. The agent files aren't generic prompts. They're refined from hundreds of real sessions.
Give the harness a complex task and walk away. Even if it's slow, it's running on your machine, your data never leaves, and you can come back later to see what it built.
Prefer to skip the email? Get the book on Amazon →
Get Chapter 1 + the Book — Free
Chapter 1 of Pi: Building Local AI Agents is free — and the whole book is free on Amazon this week. Enter your email below: the chapter lands in your inbox and you'll get the Amazon link to grab the book. No card, no subscription.
From Zero to a Working
Multi-Agent Team
Install Pi from Scratch
Set up the Pi harness on Linux, Windows (WSL2), or Mac (Apple Silicon), understand the ~/.pi/ folder structure, and run your first session with a cloud provider.
Compile & Run llama.cpp with CUDA
Build the local inference engine from source, load a GGUF model, and get your first local AI response on NVIDIA CUDA, Apple Metal, or CPU.
Manage 14 Models with llama-swap
Configure the reverse proxy that loads, unloads, and routes requests across multiple models, so you're not limited to one model at a time.
Wire the Backend End-to-End
Connect llama-swap to Pi via settings.json and models.json, understand the three config files that make the system work, and know why each model was chosen.
Deploy Agents, Skills & Extensions
Set up 13 specialized agent definitions, load ~93 skill references on demand, and hot-reload 6 extensions without forking.
Run a Live Multi-Agent Project
Watch the orchestrator delegate, researchers gather context, debuggers diagnose and fix, and coders implement, all locally, all in real time, from prompt to finished product.
Everything You Need
to Go Local
The Book: Pi: Building Local AI Agents
14 chapters of real configs and real failure data — the 18 production bugs, root-caused, with the fix for each. Free on Amazon this week, then $9.95.
Training Guide Ebook
The Local LLM Agent Harness, Second Edition: a complete PDF reference with annotated configs and decision trees.
Ready-to-Run Configs
settings.json, models.json, AGENTS.md, and 5 agent definitions — drop into your ~/.pi/ and go.
Skills & Extensions
3 skills (pipeline orchestration, goal pipeline, librarian) and 3 extensions (subagents, VCC, web access) in the Starter Bundle.
llama.cpp MTP Setup Guide
Step-by-step build instructions for multi-token prediction support, tested on RTX 3090 (24GB), 16GB cards, and Apple Silicon (Metal). Works on Linux, Windows (WSL2), and Mac. This customised setup gives you a speed boost over standard inference.
Pick Your Tier
Available Now
The book is FREE on Amazon this week. The bundle adds the ready-to-run configs that make it run.
Runs on any modern PC. NVIDIA GPU (6GB+ VRAM) recommended. Apple Silicon and CPU-only also supported. See FAQ for details.
Pi: Building Local AI Agents
eBook · Amazon · FREE this week, then $9.95 (free for Prime)
The technical deep-dive: 14 chapters on building a production local agent harness. The real bugs, the real configs.
- 14 chapters, ~95,000 words
- 18 real production bugs, root-caused
- VRAM tier guides, 6 GB to 24 GB+
- Exercises + "Boss Fight" scenarios
- Complete routing table & dispatch protocol
Start here. The bundle adds the ready-to-run configs.
Prefer to skip the email? Get the book on Amazon →
Starter Bundle
$19.95 USD
PDF + configs + agent definitions + skills. Everything that runs.
- Training guide ebook (PDF)
- Ready-to-run config bundle
- 5 agent definitions & 3 skills
- Optimized llama.cpp fork
- 3 core extensions
- Tokyo Night theme
The book is free on Amazon this week — grab it with the harness.
Common Questions
NVIDIA GPU (CUDA): RTX 3090 (24GB VRAM) is the sweet spot. All 27B models fit entirely in VRAM at 35+ tokens/sec. A 16GB card like the RTX 4060 Ti works with partial CPU offloading at ~18-20 tok/s. Even 12GB or CPU-only runs at ~10 tok/s. Apple Silicon (Metal): unified memory handles 27B models on 32GB+ Macs. CPU-only: works at ~10 tok/s. The bundle includes guidance for each tier.
You don't need VRAM for all 14 models simultaneously; llama-swap loads one model at a time. The largest models (35B MoE) need ~30GB, which partially offloads to RAM on 24GB cards. The 27B models fit entirely in 24GB VRAM. Minimum: 6GB VRAM for the smallest models. Apple Silicon: unified memory counts as VRAM. A Mac with 32GB unified memory handles the 27B models comfortably. 16GB Macs work with the smaller models (7B and under) at good speed.
Starter Bundle ($19.95): PDF training guide, config files (settings.json, models.json, AGENTS.md), 5 agent definitions (orchestrator, researcher, debugger, coder, worker), 3 skills (pipeline orchestration, goal pipeline, librarian), 3 core extensions (pi-subagents, pi-vcc, pi-web-access), and a modified llama.cpp fork with MTP support.
Basic command-line comfort is expected, whether that's Linux, macOS Terminal, or Windows WSL2. No prior LLM experience required; the series starts from a clean install and walks through every step. If you can run git clone and cmake, you're ready. The bundle covers build instructions for all three platforms.
Linux: primary platform, fully tested on Ubuntu/Debian. Windows: works via WSL2. Mac (Apple Silicon): fully supported via llama.cpp's Metal backend. The MTP (multi-token prediction) features work across all three platforms. NVIDIA GPUs (CUDA) give the best throughput, but Apple Silicon M-series chips deliver excellent performance through Metal; models that fit in unified memory run smoothly.
Full refund within 7 days of purchase. If the bundle doesn't meet your expectations, reach out via the contact form below and we'll process your refund through Stripe. After 7 days, all sales are final for digital products.
Questions, troubleshooting, and updates: reach out via the contact form below. The config files and extensions are updated as Pi evolves.
Contact
Questions, requests, or just want to say hi?
More Products
Phoenix AI
The AI chat widget for Australian businesses. Runs locally, answers visitors 24/7, data stays in Australia. From $29/month.
Visit Phoenix AI →pi-docker
Free demo project built in 10 minutes with the PI Agent Harness. Sandboxed, isolated, fully local.
View on GitHub →Pi: Building Local AI Agents
Two years of hard-won knowledge running AI on your own hardware — 14 chapters (~95,000 words) of real configs, real failure data, and real numbers. No API keys. No per-token bills.
FREE on Amazon this week — then US$9.95 / A$13.83, or $0 with Kindle Unlimited.
What's inside
- → Hardware & VRAM math — what actually runs on 6 GB, 12 GB, and 24 GB GPUs
- → Model selection — Qwen, Gemma, MoE models, quantization (Q4–Q8), and the traps to avoid
- → The inference stack — llama.cpp, GGUF, llama-swap hot-swapping, 24/7 systemd
- → The agent harness — orchestrator, skills, extensions, and permission gates
- → Failure modes — the 18 real production bugs, each with its fix and a “WHY DOES THIS EXIST?” callout