LocalPibox.

A personal, local-first AI devstack built on Pi.dev.

Pi.dev Lemonade · local LLM Qwen GitHub Pages

A containerized Pi coding agent paired with a local large model served by Lemonade, MCP servers, and forked extensions — focused on local LLM development and tuned for a Ryzen AI Max+ 395 with 128 GB of unified memory. The model never leaves your machine.

Why local, why Lemonade

Because the target hardware can run large models comfortably, there is no reason to send code to a cloud API. Lemonade serves the model locally (http://127.0.0.1:13305/v1) — it is the default provider. The default model is Qwen3.6-35B-A3B-MTP-GGUF, driven at medium/low thinking via the reasoning_effort patch in the Pi fork. 128 GB of unified APU memory means big context windows and on-device reasoning: private, offline-capable, fast.

The stack

Repo	What it is	Upstream
devstack	container image + `lpb` launcher, VSCodium web IDE, CI, bootstrap	own
pi	Pi monorepo fork: Qwen reasoning, overflow fixes	earendil-works/pi
config	preset: settings, MCP servers, skills, agents	own
lemonade-pi-plugin	Lemonade provider: Qwen thinking + vision	lemonade-sdk
pi-subagents	sub-agents: parallel execution, live widget, custom agent types, mid-run steering	tintinweb
lpb-memory	persistent memory extension (subprocess reviews)	independent
localpibox	canonical project overview & stack reference	own
localpibox.github.io	this project site (GitHub Pages)	own

The `lpb` utility

Host-side launcher for the whole stack. Install once, then drive it from your shell:

curl -fsSL https://raw.githubusercontent.com/localpibox/devstack/main/scripts/install.sh | bash

Commands

Command	What it does
`lpb [project]`	Start a Pi CLI session (foreground); default `~`
`lpb /proj -- <pi-args>`	Pass args to `pi` (`-p`, `--session`, `--continue`, `--thinking`, `--name`)
`lpb --shell [project]`	Open an interactive bash shell in the container
`lpb --web [project]`	Start VSCodium (web IDE) at a project (background)
`lpb --ssh [pubkey\|path] [proj]`	Start an sshd server (background) for remote login
`lpb --stop`	Stop the container
`lpb --remove`	Stop + remove the container and its state dirs
`lpb --logs`	Stream container logs
`lpb --update`	Pull the latest image(s)
`lpb --config`	Show the launcher config file location
`lpb --help`	Show usage

`--web` (VSCodium) options

--host · --port · --token · --without-token · --data-dir · --user-data-dir · --ext-dir · --base-path

Examples

lpb                                          # Pi at ~
lpb /home/user/myproject                     # Pi at a project
lpb /myproject -- -p "summarize this repo"   # survey via CLI
lpb /myproject -- --session abc123           # resume a session
lpb --shell /home/user/myproject             # shell into the container
lpb --web --port 8080                        # VSCodium on :8080

Quick start

Prerequisites: a container engine (podman or docker) and, to run models on-device, hardware with sufficient RAM/APU (this stack targets machines like a Ryzen AI Max+ 395 with 128 GB, plus a local Lemonade/llama.cpp server on :13305).

curl -fsSL https://raw.githubusercontent.com/localpibox/devstack/main/scripts/install.sh | bash
lpb /path/to/your/project

Open the browser to http://localhost:3000 (token: devsession) for the VSCodium web IDE. Full setup, manual podman run, and multi-project usage are in devstack/README.

Architecture

Your shell → lpb → container (podman/docker)
                    ├─ Pi CLI (custom fork: Qwen reasoning_effort)
                    ├─ lemonade plugin → local Qwen LLM (127.0.0.1:13305/v1)
                    ├─ MCP servers (agent-browser, exa, …)
                    └─ extensions: lpb-memory, pi-subagents, powerline footer

Config preset from config seeds ~/.pi/agent/ on first run; extensions install on boot from settings.json#packages via pi update --extensions. Full architecture and the Qwen-thinking overflow notes live in devstack/README and devstack/doc.

Requirements

Container engine — podman or docker.
Local LLM server — Lemonade/llama.cpp serving Qwen3.6-35B-A3B-MTP-GGUF at 127.0.0.1:13305/v1.
Hardware — large models need serious memory: the reference machine is a Ryzen AI Max+ 395 with 128 GB unified APU memory. Smaller models run on less; your mileage depends on context size and model.
Network — outbound only to pull images and (optionally) package updates; model inference stays on-device.

Docs & guides

localpibox/README — canonical project overview and stack reference
Forking & Repointing — fork the whole stack and repoint lpb.stack.env at your own repos
Forked Repos & Upstream Policy — per-fork patch contents and upstream mapping
CONTRIBUTING — contributing directly, forking for your own stack, or feeding back experiences
Config-manager design — planned pi-config utility for MCP + extensions

Contributing

The stack is designed to be forked and personalized. Three paths:

Contribute directly — improve patches, add features, fix bugs. Each repo has its own CONTRIBUTING.md.
Fork for your stack — fork all repos, repoint lpb.stack.env, and run your own variation.
Feed back experiences — share what works reliably on your hardware with your models. Helps everyone converge on robust local configurations.

See devstack/CONTRIBUTING.md for the full guide.