minxg-beta 2026.7.31

MINXG — modular AI worker platform with a chat CLI, an OpenAI-compatible v1 gateway, opt-in extensions (ADB/ROOT/files), and a self-developed temporal driver engine. 376 operators in 11 categories (306 of them mathematical across 6 pillars). Pure Python. Published on PyPI as `minxg-beta` — `pip install minxg-beta`.

MINXG-beta — AI Orchestration Platform

Version: 2026.07.31
Status: ⚠️ UNSTABLE — This build is undergoing major architectural refactoring. The TUI (CLI chat) has been rewired as the default entry point with the agent orchestrator permanently merged into normal conversation. The Web UI is still accessible via minxg web but the TUI currently has known rendering issues. Do not use this build for production work.

---

What is MINXG-beta?

MINXG-beta is a modular AI orchestration platform written in Python. It provides:

• A chat CLI with streaming, tool calling, and multi-provider model support
• An OpenAI-compatible /v1 API gateway (MCP server mode included)
• A worker system with 73+ worker classes exposing 600+ tools across file I/O, networking, crypto, math, text, system, AI, data, image/audio/video/PDF processing
• 7 mathematical pillars: Geometric Algebra, Category Theory, Information Geometry, Topology, Chaos Theory, Fiber Bundles, Symbolic Differential Algebra
• Polyglot bridges to C/C++, Rust, Go, Java, Julia, R, WebAssembly, and Datalog
• Multi-channel support for Telegram, Discord, Slack (opt-in, long-polling/websocket)
• A skill system for agent instruction packs (markdown-based, no code execution)
• A built-in multi-agent collaboration layer (Commander / session / reviewer / shared task board) for planning, parallel execution, and conflict-aware handoff
• Self-evolution and reverse engineering tooling (academic/interop scope)

Platforms tested: Linux, macOS, Docker. Windows and Android/Termux are supported on a best-effort basis — see the platform notes below.

Recent maintenance notes

• ⚠️ This build is transitional. The TUI (prompt_toolkit-based chat) is now the default entry point (minxg without arguments). The agent orchestrator (NexusOrchestrator) is permanently merged into conversation — there is no toggle to disable it because without it the system cannot converse.
• Known issues: TUI input focus, full-screen rendering, and API 401 errors are still being resolved. The Web UI (minxg web) remains functional.
• Agent functionality (50-agent orchestration) is now always-on and integrated directly into the normal chat flow. The old company_mode / Nexus toggle has been removed in favor of seamless merging.
• The minxg command no longer starts the Web UI by default — use minxg web explicitly for the browser interface.
• CLI history and compatibility tests now include a small fallback for legacy Android path checks.

---

Honest Comparison

We previously positioned MINXG-beta against other tools with inflated claims and unnecessarily aggressive language. That was a mistake. Here's a sober look:

Capability	MINXG-beta	Notes
Tool breadth	600+ tools across 73+ workers	Wide but shallow in some categories
MCP server	✅ Yes	Exposes all workers as MCP tools
OpenAI-compatible API	✅ Yes	/v1/chat/completions, streaming
Mathematical operators	300+ (7 pillars)	Niche strength; not general-purpose
Polyglot bridges	8 runtimes	C/C++, Rust, Go, Java, Julia, R, WASM, Datalog
Multi-channel (Telegram/Discord/Slack)	Opt-in	Long-polling / websocket, no public port
Agent framework	Built-in multi-agent orchestration	Commander/session/reviewer flow; functional but still evolving
RAG system	Vector + hybrid search	Basic implementation
Workflow engine	DAG with branching/loops	Early stage
Guardrails	Input/output validation, PII, toxicity, prompt injection, hallucination, code injection prevention	Comprehensive regex-based screening
Windows support	Best-effort	CI uses mocked paths; real-device testing needed
Android/Termux support	Best-effort	Degraded networking without termux-api

We are not "better than everyone." We have genuine depth in math/science operators and polyglot bridges, but we lag on messaging-platform reach, community skill breadth, and production hardening compared to established projects like OpenClaw or LangChain. If you need a wide messaging ecosystem or a mature plugin marketplace, those are better choices today.

---

Quick Start

# The easy way
pip install minxg-beta
Successfully installed minxg-beta-0.19.0

# The "from source" way
git clone https://github.com/pineapple-ooo/MINXG-beta-Beta.git
cd MINXG-beta-Beta
pip install -e .

# Verify installation
minxg doctor

Run minxg with no arguments to start the chat CLI. Use minxg gateway --mcp for the MCP server.

---

Core Features

Chat CLI (minxg)

• Streaming responses with tool-call visualization
• 32+ model providers (OpenAI, Anthropic, Google, DeepSeek, xAI, local via Ollama, etc.)
• Mid-conversation model switching (/model)
• Multi-language UI (12 languages, live-switchable with /lang)
• Theme system (8 built-in themes)

API Gateway (minxg gateway)

• OpenAI-compatible /v1/chat/completions endpoint
• MCP server mode: minxg gateway --mcp
• All 73+ workers exposed as callable tools
• Cost tracking, rate limiting, request logging

Built-in Multi-Agent Collaboration (Always-On)

The agent orchestrator is permanently merged into the normal conversation flow. When you type a message in the TUI, it's routed through the NexusOrchestrator which provides:

• Tool-calling loop with iteration budget, anti-loop guards, and dedup cache
• Streaming response with real-time token, thinking, and tool-call events
• Infinite context via layered memory (working + episodic + semantic)
• Session persistence across restarts via SQLite-backed session store
• Multi-agent routing (commander / session / reviewer / shared task board) for planning, parallel execution, and conflict-aware handoff
• No toggle to disable — the agent IS the conversation layer. Turning it off would make the system incapable of conversing.

Formerly a separate "Nexus" / company_mode feature, the agent orchestration is now the foundation of every chat session. Run minxg to start. Type /help for available slash commands.

Worker System

73+ worker classes organized in 8 categories:

• I/O: file, network, database, web, archive, cloud, media
• Dispatch: system, shell, process, platform, ADB, root, notify, security
• Transform: state machines, persistence, events, rules, hot-reload, AI tools
• Scalar: text, datetime, math, string, version, color, markdown
• Aggregate: encoding, crypto, data, templates, i18n, ML, benchmarks
• DevTools: Android Forge, QuadForge, DevShell, ReverseStudio, AuditWorker, SelfEvolutionWorker
• Polyglot: Julia, R, Datalog, WASM workers
• Math Pillar: GeometryWorker (driver-geometry glue)

Mathematical Pillars (7 + Geometry)

• Geometric Algebra — multivector calculus
• Category Theory — functors, monads, morphisms
• Information Geometry — Fisher metric, natural gradient
• Topology — homology, cohomology, spectral sequences
• Chaos Theory — Lyapunov exponents, bifurcation, fractals
• Fiber Bundles — connections, curvature, parallel transport
• Symbolic Differential Algebra — jets, Lie brackets, differential ideals
• GeometryWorker — driver-geometry integration layer

Polyglot Bridges (8 runtimes)

Runtime	Bridge	Notes
C/C++	c_core/, cpp_core/	FFI via cffi / pybind11
Rust	rust_core/	minxg_rust_core crate, ctypes
Go	go_core/	JSON-RPC over stdio
Java	java_core/	JVM daemon + vector engine
Julia	julia_worker.py	minxg.contracts.runtime.julia
R	r_worker.py + r_scripts/	Statistical computing
WASM	wasm_worker.py	Wasmtime sandbox
Datalog	datalog_worker.py	Clingo / pyDatalog

Skill System (minxg skill ...)

• Markdown SKILL.md files with YAML frontmatter
• Local or remote catalogs (JSON file, raw GitHub URL works)
• minxg skill search / install / new / publish
• Also exposed to the chat agent as skill_search, skill_install, skill_new tools
• No code execution on install — skills are instruction packs, not plugins

Multi-Channel Gateway

• Telegram (long-polling), Discord (websocket), Slack (websocket)
• Disabled by default; enable in config/gateway.yaml
• No public port required — all outbound
• See gateway/channel_*.py for per-channel status and known gaps

---

CLI Reference

minxg                    # Start chat CLI (default)
minxg setup              # Run setup wizard
minxg config             # Show current configuration
minxg status             # Runtime status
minxg tools              # List available tools
minxg model [name]       # Set or view model
minxg api <url>          # Quick-set API base URL
minxg key <key>          # Quick-set API key
minxg lang [code]        # Switch display language
minxg gateway [--detach] # API gateway (foreground default)
minxg gateway --mcp      # MCP server mode
minxg doctor             # Self-check (config + tools + extensions)
minxg ext <sub>          # Extension management
minxg skill <sub>        # Skill management
minxg features           # Feature showcase
minxg themes             # Theme management
minxg export             # Export memories (json/markdown)
minxg import <file>      # Import memories
minxg help               # Show this cheatsheet

---

MCP Server Setup

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "minxg": {
      "command": "minxg",
      "args": ["gateway", "--mcp"]
    }
  }
}

Cursor / Windsurf / Any MCP Client

Same configuration; the --mcp flag exposes all workers as MCP tools.

---

Platform Notes

Windows

• Install Python 3.10+, then pip install minxg-beta
• Run in PowerShell or cmd.exe
• minxg doctor reports missing capabilities
• Command execution routes through cmd.exe correctly
• Not tested on real hardware in this release — CI mocks the Windows code paths. Please report issues.

Android (Termux)

• Install Termux from F-Droid
• pkg install python git && pip install minxg-beta
• Notifications use termux-api when available; fallback to log lines
• Networking tools degrade gracefully without inetutils
• Not tested on real device in this release — CI mocks Termux paths. Please report issues.

Linux / macOS / Docker

• Fully tested in CI
• Dockerfile provided for containerized deployment

---

Contributing

1. Fork the repository
2. Read ARCHITECTURE.md for the module layout rationale
3. Make your changes
4. Run pytest tests/ — all tests must pass
5. Submit a PR

See CONTRIBUTING.md for details.

---

License

MIT. Use freely. We take no responsibility for what your AI does.

---

A Note on Tone

Earlier versions of this README (up to 0.18.3) used hyperbolic, aggressive, and profane language. That was unprofessional and misleading. We apologize. This rewrite aims for honesty about what MINXG-beta does well, where it falls short, and what you can actually expect. If you find remaining instances of the old tone, please open an issue or PR.

---

Version 2026.07.31 — transitional build, agent-merged, not fully usable.