GPAgent

A native C++/Qt AI agent with tool use, multi-provider LLM support, and self-improving tool selection.

Features

• Multi-Provider LLM Support: Claude (Anthropic), Gemini (Google), with automatic fallback
• Built-in Tools: File operations, bash execution, grep, glob, web search, web fetch, image reading
• Tool Selector: Self-improving tool selection using Tiny Recursive Models (TRM) (Jolicoeur-Martineau, 2025), combining recursive reasoning with BERT-style masked prediction for unsupervised learning from usage patterns
• Episodic Memory: Persistent storage of task history for learning and context
• Context Management: Automatic summarization to handle long conversations
• Modern Qt UI: Clean chat interface with streaming responses

Requirements

• C++20 compatible compiler (GCC 11+, Clang 13+, MSVC 2022+)
• CMake 3.21+
• Qt 6.5+ (Core, Quick, QuickControls2)
• OpenSSL development libraries

Building

# Clone the repository
git clone https://github.com/waddadaa/gpagent
cd gpagent

# Create build directory
mkdir build && cd build

# Configure (adjust Qt path as needed)
cmake .. -DCMAKE_PREFIX_PATH=/path/to/Qt/6.x.x/gcc_64

# Build
cmake --build . --parallel

Dependencies

The following dependencies are fetched automatically via CMake FetchContent:

• nlohmann/json - MIT License
• spdlog - MIT License
• cpp-httplib - MIT License
• yaml-cpp - MIT License

Configuration

API Keys

Set your API keys as environment variables:

export ANTHROPIC_API_KEY="your-anthropic-key"
export GOOGLE_API_KEY="your-google-key"
export PERPLEXITY_API_KEY="your-perplexity-key"  # For web search

Or configure them in the Settings page within the application. Keys are stored in:

• Linux: ~/.config/gpagent/config.yaml
• macOS: ~/Library/Preferences/gpagent/config.yaml
• Windows: %APPDATA%/gpagent/config.yaml

Runtime Data

Episodic memory, TRM models, and logs are stored in ~/.gpagent/

Usage

./GPAgent

The application provides a chat interface where you can:

• Ask questions and have conversations
• Request file operations (read, write, edit)
• Execute shell commands
• Search the web
• Analyze images

Architecture

gpagent/
├── include/gpagent/     # Header files
│   ├── agent/           # Orchestrator, planner
│   ├── core/            # Config, types, errors
│   ├── llm/             # LLM providers
│   ├── memory/          # Episodic memory
│   ├── context/         # Context management
│   ├── tools/           # Tool system
│   ├── trm/             # Tool Ranking Model
│   └── ui/              # Qt UI components
├── src/                 # Implementation files
├── Qml/                 # QML UI files
└── CMakeLists.txt

Tool Selector (TRM-based)

A self-improving tool selection system using Tiny Recursive Models that learns from usage patterns without requiring labeled training data.

Architecture

• Tiny recursive network: 2 layers, 512 hidden size (~5-7M parameters)
• Deep supervision: Learns from each step, not just final outcomes
• Recursive refinement: Iteratively improves predictions (T=3, n=6 recursions)

Training Objectives (Unsupervised)

Loss	Weight	Description
Contrastive	1.0	Similar tasks → similar tool choices
Next Action	0.5	Predict next tool from action sequence
Outcome	0.3	Predict task success/failure
Masked	0.2	BERT-style: mask a tool, predict from context

How It Works

1. Cold Start: Uses keyword matching + history boosting
2. Episode Collection: Stores task descriptions, tool sequences, and outcomes
3. Training: After 5+ episodes, trains on collected data
4. Inference: Predicts best tool with confidence scores

License

This project is licensed under the MIT License - see the LICENSE file for details.

Third-Party Libraries

• Qt 6 - LGPL v3 (dynamically linked) or commercial license
• nlohmann/json - MIT License
• spdlog - MIT License
• cpp-httplib - MIT License
• yaml-cpp - MIT License

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.