gtui 👾

gtui is the high-performance Terminal User Interface (TUI) for gro. It provides a dedicated, low-latency environment for managing autonomous agents, monitoring tool execution, and visualizing the "Virtual Memory" systems of your LLM runtime.

Visualizing the context: Persistent agents in a dark-mode world.

Features

• Real-time Stream Monitoring: View live token generation and tool calls with the high-fidelity logging you expect from gro --verbose.
• Dynamic Context Visualization: Track active memory pages, importance weights, and "Virtual Memory" state in a clean, terminal-native layout.
• Provider Switching: Fast-toggle between Anthropic, OpenAI, xAI, and local models.
• Resource Intensity Control: Visual feedback for the @@thinking(0.0-1.0)@@ lever—see exactly how much compute your agent is burning.
• MCP Integration: A dedicated panel for connected MCP servers and available tools.

Installation

# Clone the repo
git clone https://github.com/tjamescouch/gtui.git
cd gtui

# Install dependencies
npm install

# Link for global usage
npm link

Usage

Launch gtui to wrap your existing gro sessions:

gtui --model grok-fast

Keybindings

• Ctrl + T: Toggle thinking intensity
• Ctrl + M: Cycle memory modes (Virtual/HNSW/Fragment)
• Ctrl + L: Clear active viewport
• Tab: Switch between chat and tool-logs

Architecture

gtui is built to be as efficient as the runtime it manages. It utilizes:

• Blessed / Ink: For the terminal rendering engine.
• Unix Domain Sockets / IPC: For zero-latency communication with the gro background process.
• Metal Performance Shaders (MPS): (Experimental) Local visualization hooks for Mac Studio.

Security

gtui spawns gro as a subprocess with --bash enabled. The LLM has shell access, file read/write, and any configured MCP tools. gro's built-in tool approval prompts do not apply in this mode (no TTY — gro runs as a piped subprocess).

Recommendations:

• Do not run gtui with MCP servers that provide network access unless you trust the model
• Use in containerized environments for untrusted workloads
• Consider running without --bash by modifying the spawn arguments

License

MIT