agentsculptor is a free, open, and local experimental AI-powered development agent designed to analyze, refactor, and extend Python projects automatically. It uses an OpenAI-like planner–executor loop on top of a vLLM backend, combining project context analysis, structured tool calls, and iterative refinement.

flowchart TD
    %% === NODES ===
    A([💬 User Request])
    B([🧠 Planner Agent])
    C([🔁 Agent Loop])
    D{{🛠️ Tools}}

    %% Tools
    D1([🧪 Run Tests])       
    %% Stadium shape
    D2([📄 File Creation])
    D3([🖋️ Refactor Code])
    D4([🔗 Update Imports])
    D5([🎨 Format Code])
    D6[/💾 Backup/]          
    %% Parallelogram

    E([🏆 Success])

    %% === EDGES ===
    A --> B -->|Generates JSON Plan| C --> D
    D -.->|run_tests| D1
    D <-.->|create_file| D2
    D <-.->|refactor_code| D3
    D <-.->|update_imports| D4
    D <-.->|format_code| D5
    D -.->|backup_file| D6

    D1 -->|❌ Tests Fail| B
    D1 -->|✅ Tests Pass| E

    %% === STYLES (softer colors) ===
    style A fill:#BBDEFB,stroke:#64B5F6,color:#0D47A1
    style B fill:#E1BEE7,stroke:#BA68C8,color:#4A148C
    style C fill:#FFE0B2,stroke:#FFB74D,color:#E65100
    style D fill:#CFD8DC,stroke:#90A4AE,color:#263238

    %% Normal tools - soft blue
    style D2 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D3 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D4 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D5 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B

    %% Special tools
    style D1 fill:#C8E6C9,stroke:#81C784,color:#1B5E20   
    style D6 fill:#FFCCBC,stroke:#FF8A65,color:#BF360C   

    style E fill:#C5E1A5,stroke:#AED581,color:#33691E

AgentSculptor is an experimental project created for the OpenAI Open Model Hackathon. It has only been tested with gpt-oss-120b via vLLM.

Other models may work, but compatibility is not guaranteed at this stage. Use at your own risk and please share feedback if you experiment with different setups.

• agentsculptor
  • ⚠️ Disclaimer
  • Table of Content
  • 🚀 Getting Started / Usage
    • 1. Install from PyPI
    • 2. Or install from source
    • 3. Set environment variables
    • 4. Run CLI commands
    • 5. Other examples
    • 6. Workflow Overview
  • 🚀 Features
  • 📦 Repository Structure
  • 👨‍💻 Developer Guide
    • 🟢 Option A — Development in VS Code DevContainer
      • 1. DevContainer
      • 2. Dockerfile
      • 3. Python Dependencies
      • 4. Start
    • 🔵 Option B — Without DevContainer (Dedicated virtual environment)
  • Workflow
    • 🛠️ Tools Available
    • 💡 Usage Notes
    • ⚙️ How It Works
      • 📋 Planner Agent (Reasoning Layer)
      • 🔁 Agent Loop (Execution Layer)
      • 🛠️ Tools (Action Layer)
      • ♻️ Re-Planning & Self-Healing (Future)
  • 🚀 Roadmap
  • 🤝 Contributions
    • 1. 🛠️ Setup Your Environment
    • 2. 🔧 Make Your Changes
    • 3. 🧪 Test Your Changes
    • 4. 📦 Commit & Submit
    • 5. ⚡ Code Review
    • 6. 🎉 Celebrate!
  • 📄 License

Python 3.12 or higher is required. Follow these steps to install and use agentsculptor:

pip install agentsculptor

Clone the repository and install in editable mode (for local changes):

git clone https://github.com/Perpetue237/agentsculptor.git
cd agentsculptor
pip install -e .

agentsculptor requires a running vLLM server. Set:

export VLLM_URL="http://localhost:8008"
export VLLM_MODEL="openai/gpt-oss-120b"

Adjust according to your server setup.

Generate or refactor code with agentsculptor-cli. For example, to create a basic Dockerized FastAPI app:

agentsculptor-cli ./test_project "create files for a basic dockerized FastAPI application. The initial app should just return a JSON with 'hello to the OpenAI community'"

This will:

• Generate a minimal FastAPI app structure.
• Create Dockerfiles for containerization.
• Initialize the project in ./test_project.

agentsculptor-cli ./test_project "create fast api app with clear instructions on how to run it."

agentsculptor-cli ./test_project "Merge multiple Python modules into a single file."

agentsculptor-cli ./test_project "Refactor functions and automatically generate unit tests."

agentsculptor-cli ./test_project "Refactor my codebase and modernize. Update imports across the project after refactoring."

1. prepare_context scans your project.
2. PlannerAgent outputs a structured JSON plan.
3. AgentLoop executes each tool step by step.
4. Tests are run automatically (existing or generated).
5. Changes are applied safely, with backups if needed.

• 📂 Project Context Builder – Parses files, functions, classes, and imports.
• 🧠 Planner Agent – Generates structured JSON plans of tool calls (no free-form text).
• 🔁 Agent Loop – Executes tool calls, tracks logs, and retries if needed.
• 🛠️ Tool Registry – Includes file creation, backup, import updates, code refactoring, and test execution.
• 🧪 Automated Testing – Runs pytest or generated test files before/after modifications.
• 🖋️ Code Refactoring – Uses LLM guidance with file structure hints for safer transformations.

.
|-- LICENSE                      # Open-source license file (MIT, Apache, etc.)
|-- README.md                    # Project documentation / usage instructions
|-- agentsculptor/                # Main Python package
|   |-- __init__.py              # Marks this as a Python package
|   |-- __pycache__/             # Auto-generated bytecode cache
|   |-- agent/                   # Agent logic (planner, executor, loop, etc.)
|   |-- llm/                     # LLM client + prompt handling
|   |-- main.py                  # CLI or entrypoint for running agentsculptor
|   |-- tools/                   # Helper modules (refactor_code, run_tests, etc.)
|   `-- utils/                   # Utility functions (file ops, logging, search, etc.)
|-- agentsculptor.egg-info/       # Metadata for packaging (setuptools)
|   |-- PKG-INFO                 # Package metadata (name, version, summary)
|   |-- SOURCES.txt              # List of source files included in the distribution
|   |-- dependency_links.txt     # Deprecated mechanism for dependency links
|   |-- entry_points.txt         # Console_scripts / plugin entry points
|   |-- requires.txt             # Runtime dependencies for installation
|   `-- top_level.txt            # Top-level package(s) exposed (here: "agentsculptor")
|-- setup.py                     # Build/install configuration for setuptools
|-- test_project/                # Example project used for testing/refactoring
|   |-- Dockerfile               # Docker container config for test project
|   |-- main.py                  # Example application entrypoint
|   |-- requirements.txt         # Test project dependencies
|   `-- test_main.py             # Unit tests for test project
`-- tests/                       # Tests for agentsculptor itself (not the example project)

This repository supports development using VS Code DevContainers.

Note: The vLLM server is run separately. Make sure it is running at VLLM_URL and serving VLLM_MODEL. See /vllm-server-config/README.md for instructions on running the vLLM server.

This repository provides a .devcontainer folder for an optimized development environment with docker.

Create .devcontainer/devcontainer.json:

{
    "name": "agentsculptor",
    "build": {
      "dockerfile": "Dockerfile"
    },
    "features": {},
    "containerEnv": {
      "PYTHONUNBUFFERED": "1"
    },
    "customizations": {
      "vscode": {
        "extensions": [
            "ms-python.python",
            "ms-toolsai.jupyter",
            "innerlee.nvidia-smi",
            "Leonardo16.nvidia-gpu",
            "RSIP-Vision.nvidia-smi-plus",
            "yzhang.markdown-all-in-one",
            "MermaidChart.vscode-mermaid-chart",
            "Gruntfuggly.mermaid-export"
        ]
      }
    }
  }

.devcontainer/Dockerfile:

FROM python:3.12-slim

ENV DEBIAN_FRONTEND=noninteractive
ENV TZ=Europe/Berlin

COPY . .

RUN apt-get update && apt-get install -y curl git && rm -rf /var/lib/apt/lists/*
RUN pip3 install --no-cache-dir -r requirements.txt

requirements.txt:

black
pytest
commitizen
requests

1. Open the project in VS Code.

2. Click Reopen in Container.

3. Make sure the vLLM server is running and the environment variables are set:

   export VLLM_URL=http://localhost:8008
   

4. Intall agentsculptor

   pip install -e .

now you can use the agentsculptor-cl.

1. prepare_context scans your project files.
2. PlannerAgent generates a JSON plan of tool calls.
3. AgentLoop executes each tool in sequence.
4. Tests run (if available or auto-generated).
5. Changes are applied safely (backup first).

Each tool in agentsculptor has a name, description, and parameters. You can use these tools via the CLI or programmatically.

• Order matters: Typically, backup_file runs before refactor_code or update_imports.
• Chaining tools: You can combine create_file, refactor_code, and update_imports in a single workflow.
• Tests & safety: Run run_tests after code modifications to catch issues automatically.
• Formatting: format_code can be applied at the end of the workflow for clean, readable code.

The agentsculptor agent runs on an OpenAI-like planner–executor loop, designed to make structured, safe, and iterative changes to your codebase.

flowchart TD
    %% === NODES ===
    A([💬 User Request])
    B([🧠 Planner Agent])
    C([🔁 Agent Loop])
    D{{🛠️ Tools}}

    %% Tools
    D1([🧪 Run Tests])       
    %% Stadium shape
    D2([📄 File Creation])
    D3([🖋️ Refactor Code])
    D4([🔗 Update Imports])
    D5([🎨 Format Code])
    D6[/💾 Backup/]          
    %% Parallelogram

    E([🏆 Success])

    %% === EDGES ===
    A --> B -->|Generates JSON Plan| C --> D
    D -.->|run_tests| D1
    D <-.->|create_file| D2
    D <-.->|refactor_code| D3
    D <-.->|update_imports| D4
    D <-.->|format_code| D5
    D -.->|backup_file| D6

    D1 -->|❌ Tests Fail| B
    D1 -->|✅ Tests Pass| E

    %% === STYLES (softer colors) ===
    style A fill:#BBDEFB,stroke:#64B5F6,color:#0D47A1
    style B fill:#E1BEE7,stroke:#BA68C8,color:#4A148C
    style C fill:#FFE0B2,stroke:#FFB74D,color:#E65100
    style D fill:#CFD8DC,stroke:#90A4AE,color:#263238

    %% Normal tools - soft blue
    style D2 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D3 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D4 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B
    style D5 fill:#B3E5FC,stroke:#4FC3F7,color:#01579B

    %% Special tools
    style D1 fill:#C8E6C9,stroke:#81C784,color:#1B5E20   
    style D6 fill:#FFCCBC,stroke:#FF8A65,color:#BF360C   

    style E fill:#C5E1A5,stroke:#AED581,color:#33691E

Takes your natural language request (e.g., “merge utils.py and helpers.py into one module”).

Uses the project context (functions, classes, imports) to understand the current codebase.

Produces a structured JSON plan of tool calls – never free text. Example:

{
  "action": "refactor_code",
  "args": {
    "path": "utils/helpers.py",
    "instruction": "merge into utils.py"
  }
}

Iterates through the planned tool calls one by one.

Ensures dependencies are respected (e.g., backup → refactor → update imports → run tests).

Captures logs and errors for each step, with the ability to retry.

The agent never edits files directly. Instead, it calls specialized tools:

backup_file → snapshot before modifying.

create_file → safely generate new files.

refactor_code → apply structured code transformations.

update_imports → keep imports consistent.

run_tests → verify changes with pytest.

format_code → ensure style consistency with black.

This makes the workflow transparent, reproducible, and debuggable.

If a step fails (e.g., tests break), the agent will be able to re-plan automatically.

The loop can feed test results or error logs back into the Planner Agent, generating a new plan until success.

This section emphasizes the structured loop, safety-first approach, and extensibility of your agent.

Here’s what’s next for AgentSculptor:

• Vector store DB integration → persistent context & memory across sessions
• More refactoring tools → auto-docstring generation, linting, style normalization
• Framework templates → easy scaffolding for Flask, Django, Streamlit, etc.
• Interactive mode → iterative refinement with conversational feedback
• Multi-agent workflows → planner + executor agents collaborating on complex tasks
• VS Code extension → invoke AgentSculptor directly from your editor

We welcome contributions of all types—bug fixes, feature enhancements, documentation improvements, or new tools. Here's how you can help:

Clone the repository and install in editable mode:

git clone https://github.com/Perpetue237/agentsculptor.git
cd agentsculptor
pip install -e .

Set up required environment variables:

export VLLM_URL="http://localhost:8008"
export VLLM_MODEL="openai/gpt-oss-120b"

• Add features: Create new tools or improve the agent workflow.

• Refactor code: Improve readability, structure, or performance.

• Fix bugs: Check the issues list and pick tasks you can solve.

• Update documentation: Keep README, comments, and examples up to date.

Run the automated test suite or add your own:

pytest

• Use clear, descriptive commit messages.

• Push your branch and open a Pull Request (PR).

• Include a description of your change and reference related issues.

All contributions go through review for quality and safety. Be ready to iterate on feedback.

Once approved and merged, your contribution becomes part of agentsculptor!

Apache License. See LICENSE for details.