Codegen - Self-Extending AI System

An LLM agent that generates its own Python tools on demand using the Model Context Protocol (MCP). The agent can create, register, invoke, inspect, and remove tools at runtime — all within a single process.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    main.py (entry point)                │
│                                                         │
│  ┌──────────────┐    anyio     ┌──────────────────────┐ │
│  │  SelfExtend- │◄────────────►│  DynamicMCPServer    │ │
│  │  ingAgent    │  mem streams │  (mcp_server.py)     │ │
│  │  (agent.py)  │              │                      │ │
│  └──────────────┘              │  ┌────────────────┐  │ │
│         │                      │  │ DynamicTool-   │  │ │
│         │ OpenAI API           │  │ Registry       │  │ │
│         ▼                      │  │ (tool_registry)│  │ │
│    LLM (Ollama /               │  └────────────────┘  │ │
│    OpenAI / etc.)              │  ┌────────────────┐  │ │
│                                │  │ CodeGeneration-│  │ │
│                                │  │ Pipeline       │  │ │
│                                │  │ (code_gen...)  │  │ │
│                                │  └────────────────┘  │ │
│                                └──────────────────────┘ │
└─────────────────────────────────────────────────────────┘

Requirements

• Python ≥ 3.11
• A running LLM endpoint (Ollama, OpenAI, LM Studio, etc.)

Installation

# 1. Clone the repository
git clone https://github.com/ms1963/Codegen.git
cd codegen

# 2. Create and activate a virtual environment
python -m venv .venv
source .venv/bin/activate      # Linux/macOS
# .venv\Scripts\activate       # Windows

# 3. Install dependencies
pip install -r requirements.txt

# 4. Configure environment
cp .env.example .env
# Edit .env with your LLM settings

Running

Interactive mode (default)

cd codegen
python main.py

Single query mode

cd codegen
python main.py --query "Calculate the compound interest on $10,000 at 5% for 10 years"

With custom LLM settings

cd codegen
python main.py \
  --base-url https://api.openai.com/v1 \
  --api-key sk-... \
  --model gpt-4o \
  --log-level DEBUG

With Docker

# Build
docker build -t self-extending-ai .

# Run interactive (with local Ollama)
docker run -it \
  -e LLM_BASE_URL=http://host.docker.internal:11434/v1 \
  -e LLM_MODEL=qwen2.5:14b \
  -v$(pwd)/data:/data \
  codegen

# Run single query
docker run --rm \
  -e LLM_BASE_URL=http://host.docker.internal:11434/v1 \
  -v $(pwd)/data:/data \
  codegen --query "What is the square root of 144?"

Environment Variables

Variable	Default	Description
LLM_BASE_URL	http://localhost:11434/v1	LLM API base URL
LLM_API_KEY	ollama	API key
LLM_MODEL	qwen2.5:14b	Model name
TOOLS_FILE	tools.json	Tool persistence file
LOG_LEVEL	INFO	Logging level

Built-in Meta-Tools

Tool	Description
generate_and_register_tool	Generate a new Python tool from a description
list_registered_tools	List all currently registered dynamic tools
get_tool_source	Inspect the source code of any registered tool
remove_tool	Remove a tool from the registry
get_registry_stats	Get usage statistics for all tools

Interactive Commands

Command	Action
exit / quit	Stop the system
clear	Reset conversation history
stats	Show registry statistics

Issues Fixed (30 total)

Severity	Count	Examples
🔴 Critical	10	Deadlocks, missing functions, deprecated APIs
🟡 Medium	13	Edge cases, type handling, error loops
🟢 Low	7	Consistency, persistence, minor bugs