← Back to archive

Repo of the Day

MegaDyneSystems/prism32

Published: Sep 28, 2026

Open repository ↗

Build anything on anything, Self evolving AI super assistant designed to run at extremely fast speeds on any eras hardware with no comprimises

Summary

Prism32 is a single-file, stdlib-only Python AI assistant that turns a terminal into a multi-agent command harness. It talks to OpenAI-compatible APIs, runs shell commands the model emits in fenced execute blocks, spawns subagents on different models, and can generate and load new plugins mid-session. The project ships installers for pip, Homebrew, Scoop, npm, and several niche platforms such as OpenWrt, Synology NAS, and Termux.

What it is useful for

Prism32 is useful for engineers who want a local-first AI shell that runs on minimal hardware without Node, browsers, or a database server. The README describes verified end-to-end runs on a 27MB OpenWrt router, NetBSD 10.1 i386, Synology DSM 7.1.1 on ARMv7, and Fire TV / Termux on Android 5.1.1, and a low-RAM mode auto-activates below 64MB. Practical scenarios covered by the docs include:

  • Auditing a server, where the model emits commands, Prism32 runs them, and the output feeds back until the goal is structurally complete. The /mission planner breaks the goal into 3-6 steps that run as parallel subagent shards and can be controlled with status, pause, resume, interject, and wait.
  • Mixing providers per task: a strong reasoning model on the main agent while bulk scanning subagents run on a cheap local llama.cpp, Ollama, LM Studio, vLLM, or Jan server, picked from one unified /model browser.
  • Working on hardware that already has Python 3.7+ and a shell, including robots, IoT hubs, 3D-printer companions (OctoPrint, Klipper, Repetier), and SBCs, where the agent can read /dev nodes and generate stdlib plugins to talk to them on the fly via /extend.
  • Coordinating several installed AI CLIs (OpenCode, Codex CLI, Claude Code, Aider, Gemini CLI, Goose, etc.) from one REPL, since Prism32 scans for them and exposes their availability in context.

How engineers can use it

The fastest documented install is pip install prism32 followed by prism32 --setup-runtime and prism32. Source users can run git clone https://github.com/MegaDyneSystems/prism32.git && cd prism32 && bash install.sh && prism32, with a no-root variant that lands at ~/.local/bin/prism32. A local llama.cpp or Ollama server is pointed at with python3 prism32.py --api http://127.0.0.1:8080 --model local-model; OpenRouter works the same way with --api https://openrouter.ai/api/v1 and an API key.

Once inside the REPL, the cheat sheet lists /provider (an interactive picker that TCP-probes llama.cpp on 8080, Ollama on 11434, LM Studio on 1234, vLLM on 8000, and Jan on 1337), /model to assign main and subagent slots, /extend <goal> to ask the model to generate a stdlib-only plugin and load it without a restart, and /bash <cmd> to run a shell command directly. Documented limitations to keep in mind: Windows has reduced terminal interjection support, the agent can rewrite its own code when /evolve on is set, and Marlin/RepRapFirmware/Hubitat are MCU-only and require pointing Prism32 at a companion Linux host.