qwerty-agent 2.0.0

Zero-dependency symbolic AI agent. Pure Python stdlib. Runs anywhere.

Qwerty v2 — Zero-Dependency Symbolic AI Agent

Pure Python stdlib. No numpy. No torch. No sklearn. 10MB total. Runs on a Raspberry Pi.

Qwerty is an autonomous software agent that reads files, edits code, runs commands, searches the web, learns from every interaction, and composes multi-source answers — all with zero external dependencies. It's a symbolic AI: deterministic, constitution-bound, auditable, and completely offline-capable.

Why Qwerty?

Qwerty	Other AI Agents
pip install → runs	Requires Python ML stack (numpy, torch, transformers, 800MB+)
Offline by default	Cloud API or local LLM (8GB+ RAM)
10MB on disk	2-20GB for local models
Constitution hardcoded at engine level	Safety is a prompt (jailbreakable)
Deterministic — same input, same output	Non-deterministic (sampling)
Runs on ARM (Raspberry Pi Zero)	Needs x86 GPU or cloud

Features

• 14 tools: read, write, edit, search, find, run, web search, wikipedia, learn, analyze errors, and more
• Fuzzy NLP: handles "runn ls -la", "plz tel me bout the kernul", "idk wat uefi is" — 50+ shorthand maps, Levenshtein, n-gram Jaccard, contraction expansion, character dedup
• Knowledge base: 8 curated domains + 127 Wikipedia articles covering programming, OS dev, networking, math, physics, chemistry, biology, technology, and general knowledge
• Genius composition engine: retrieves, scores, and assembles multi-sentence responses from 2500+ sentences across all knowledge files
• On-demand learning: lookup <topic> → Wikipedia API (free, caches for next time); web fallback → DuckDuckGo
• Constitution safety: 9 hardcoded rules block rm -rf /, credential leaks, file overwrites without confirmation, and other dangerous operations
• Persistent memory: every answer, every lookup, every fix is stored in memory/learned.json — learning compounds across sessions
• Recursive solve loop: react → think → explore → internet → report — 3 layers of fallback before giving up
• Protocol: JSON-RPC, MCP server, and CLI modes — plugs into any editor or toolchain
• 76 passing tests, zero external dependencies

Quick Start

git clone https://github.com/KasishStar/qwerty-agent
cd qwerty-agent
python agent.py

Or ask a question directly:

python agent.py "what is a kernel"

Interactive Mode

python agent.py

qwerty> what is UEFI boot
qwerty> lookup quantum computing
qwerty> list files
qwerty> read file agent.py
qwerty> learn that the answer is 42
qwerty> exit

Interface (with history and slash commands)

python interface.py

JSON-RPC Protocol

python protocol.py --process "what is a kernel"
python protocol.py --json "what is a kernel"
python protocol.py --schema
python protocol.py --mcp

Architecture

User Input → Normalize → Classify Intent → Constitution Check → Execute → Result
                              ↓
                    Genius Composition Engine
                    (2500+ sentences, 8+ domains)
                              ↓
                    Recursive Fallback Loop
                    react → routine → explore → internet → report
                              ↓
                    Persistent Memory (learned.json)

All layers are checked against the constitution before execution. No layer can run rm -rf /, overwrite user files without confirmation, or leak credentials.

How It Works

Qwerty is not an LLM. It's a symbolic AI — every decision is deterministic if/then logic. Its "intelligence" comes from:

1. Broad knowledge (Wikipedia articles + curated domains)
2. Fuzzy pattern matching (Levenshtein, n-gram Jaccard, keyword overlap)
3. Multi-source composition (Genius scores and assembles sentences)
4. On-demand learning (fetches and stores new knowledge at runtime)
5. Compounding memory (every success is saved for next session)

What Qwerty Can Do

Task	Example
Answer knowledge questions	"what is a microkernel"
Read/write files	"read file agent.py"
Edit code	"edit line 10 of test.py to say hello"
Run commands	"run cargo build"
Search code	"search for TODO in src/"
Learn new facts	"learn that X that Y"
Wikipedia lookup	"lookup TCP protocol"
Web search	"search web for rust uefi tutorial"
Analyze errors	Automatic on command failure
Remember across sessions	All in memory/learned.json

Requirements

• Python 3.10+
• No pip packages needed
• Linux, macOS, or Windows
• ARM (Raspberry Pi) supported
• Internet optional (for web search and Wikipedia)

Project Structure

qwerty-agent/
├── agent.py              # Main engine: constitution, tools, NLP, solve loop
├── genius.py             # Response composition engine
├── interface.py          # Interactive REPL with history
├── protocol.py           # JSON-RPC / MCP / CLI modes
├── constitution.json     # 9 safety rules
├── knowledge/            # 9 JSON knowledge files (8 curated + Wikipedia)
├── memory/               # Persistent learning (auto-generated)
├── workflows/            # Pre-defined multi-step workflows
├── knowledge_importer.py # Wikipedia article downloader
├── test_all.py           # 76 comprehensive tests
└── pyproject.toml        # Build configuration

Tests

python test_all.py

License

MIT — see LICENSE.

---

Created by KasishStar. Pure symbolic AI. No ML dependencies. Runs anywhere.