A containerized Pi coding agent paired with a local large model served by Lemonade, MCP servers, and curated 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.
| Destination | Where |
|---|---|
| 🌐 Project site | localpibox.github.io |
| 🚀 Get started | devstack — Quick Start |
| 📖 Full documentation | devstack/README |
| 🔧 Contributing | devstack/CONTRIBUTING |
8 repositories, one cohesive system:
| # | Repository | Role | Status |
|---|---|---|---|
| 1 | devstack | Container image, lpb launcher, CI |
🟢 original |
| 2 | pi | Pi monorepo fork: Qwen reasoning, overflow fixes | 🔧 fork of earendil-works/pi |
| 3 | config | Settings, MCP servers, skills, agents preset | 🟢 original |
| 4 | lemonade-pi-plugin | Lemonade provider: Qwen thinking + vision | 🔧 fork of lemonade-sdk |
| 5 | pi-subagents | Centralized subagent model registry | 🔧 fork of tintinweb |
| 6 | lpb-memory | Persistent memory (subprocess reviews) | 🟢 independent |
| 7 | localpibox | This repo: project overview & docs | 🟢 original |
| 8 | localpibox.github.io | GitHub Pages project site | 🟢 original |
Your shell → lpb → container (podman/docker)
├─ Pi CLI (forked: 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
One-line install:
curl -fsSL https://raw.githubusercontent.com/localpibox/devstack/main/scripts/install.sh | bashEvery forked repo carries LocalPibox work as a single squashed commit on top of upstream — one clean patch, always rebased.
| Fork | Upstream | What LocalPibox adds |
|---|---|---|
| pi | earendil-works/pi | reasoning_effort, context overflow Case 4, compaction tuning |
| lemonade-pi-plugin | lemonade-sdk | Qwen thinking-format, vision detection, reasoning-format handling |
| pi-subagents | tintinweb | Centralized model registry (parent-model inheritance) |
Patches are candidate upstream contributions — they go upstream when generally useful and not too opinionated. See each fork's README for details.
- Forking & Repointing — fork the whole stack and run your own variation
- Upstream Policy — per-fork patch contents and upstream mapping
- Contributing — contribute directly, fork for your stack, or share experiences
- Config-manager design — planned
pi-configutility - Qwen thinking overflow — context window mitigations
Three paths:
- Contribute directly — improve patches, add features, fix bugs
- Fork for your stack — fork all repos, repoint
lpb.stack.env, run your variation - Feed back experiences — share what works on your hardware with your models