AEGIS is a prototype implementation for "Aegis: Towards Agentic Support for Software Artifact Reviews". It generates structured evidence about software artifacts referenced in scientific papers through a three-stage pipeline: extract artifacts, generate evidence, and assess results.

Warning: This is a research prototype intended for local use only. Do not deploy publicly. The application mounts the Docker socket into the container, which grants the container effective root access to the host. See Security below.

• Docker & Docker Compose
• Anthropic API key

git clone https://github.com/kkirchheim/aegis.git
cd aegis

cp .env.example .env
# Edit .env and set your ANTHROPIC_API_KEY

docker-compose up -d
# Open http://localhost:5000

Default admin credentials are admin / changeme (configurable via ADMIN_USERNAME / ADMIN_PASSWORD in .env). If ADMIN_PASSWORD is unset, a random password is generated and printed to the logs on startup.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

cp .env.example .env
# Edit .env and set your ANTHROPIC_API_KEY

PYTHONPATH=src python src/app.py

Open http://localhost:5000 to upload papers and view results.

python scripts/repro-cli.py \
  -u http://localhost:5000 \
  -k <api_key> \
  -p paper.pdf \
  -o results.json

1. Extract — Parse the PDF, identify software artifacts, and extract metadata
2. Generate Evidence — Run three types of evidence generators against extracted artifacts:
   • Deterministic — execution checks (scripts run inside Docker containers)
   • Semantic — LLM-based evaluation plugins
   • Interactive — agentic exploration of repositories in sandboxed containers
3. Assess — Aggregate evidence into a structured review context

src/           # Application source (Flask app, services, models, blueprints)
agent/         # Agent code + agent Dockerfiles
web/           # Frontend assets (static/, templates/)
tests/         # Test suite
docker/        # Main app Dockerfile
scripts/       # CLI tools
docs/          # Documentation

Code execution happens inside isolated Docker containers, built automatically on first use. To build manually:

scripts/build_agent_images.sh

To rebuild without Docker layer cache:

scripts/build_agent_images.sh --no-cache

The script builds these images:

docker build -t paper-reproducibility-agent:latest -f agent/Dockerfile .
docker build -t paper-reproducibility-agent-ml:latest -f agent/Dockerfile.ml .
docker build -t paper-reproducibility-agent-py27:latest -f agent/Dockerfile.py27 .
docker build -t paper-reproducibility-agent-py34:latest -f agent/Dockerfile.py34 .
docker build -t paper-reproducibility-agent-py36:latest -f agent/Dockerfile.py36 .
docker build -t paper-reproducibility-agent-matlab:latest -f agent/Dockerfile.matlab .
docker build -t paper-reproducibility-agent-matlab-official:latest -f agent/Dockerfile.matlab-official .

The Python 2.7, Python 3.4, and Python 3.6 legacy images run the control agent with modern Python 3, but repository commands default to the selected legacy Conda environment with pip. They also configure the old Anaconda free channel so legacy packages can be installed when a paper requires them. Select a legacy Python image in the upload form for papers that require old Python or legacy scientific Python versions.

The MATLAB Official image expects a local licensed MathWorks base image named matlab-licensed-test:r2025b. Create it by signing in to the official image, then committing that running container from another terminal:

docker run --rm -it mathworks/matlab:r2025b matlab -licmode onlinelicensing
docker commit <container_id> matlab-licensed-test:r2025b
docker build -t paper-reproducibility-agent-matlab-official:latest -f agent/Dockerfile.matlab-official .

To use a different licensed base image tag:

docker build \
  --build-arg MATLAB_BASE_IMAGE=my-matlab-licensed-base:r2025b \
  -t paper-reproducibility-agent-matlab-official:latest \
  -f agent/Dockerfile.matlab-official .

To install toolboxes into the official MATLAB agent image during build, pass MATLAB_PRODUCTS. For example, to add Computer Vision Toolbox, Statistics and Machine Learning Toolbox, and Symbolic Math Toolbox:

docker build \
  --build-arg MATLAB_PRODUCTS="Computer_Vision_Toolbox Statistics_and_Machine_Learning_Toolbox Symbolic_Math_Toolbox" \
  -t paper-reproducibility-agent-matlab-official:latest \
  -f agent/Dockerfile.matlab-official .

The image includes MathWorks Package Manager as mpm. MATLAB commands are wrapped so matlab -batch "..." automatically uses online licensing mode.

After changing an agent Dockerfile, rebuild the matching image before starting a new job so Docker does not reuse an older local image.

This application requires the Docker socket (/var/run/docker.sock) to be mounted so it can spawn agent containers. This effectively gives the application root-level access to the host system. Additionally, the agent containers clone and execute arbitrary code from repositories referenced in uploaded papers. There is a risk of prompt injection attacks. Use at your own risk.

Do not expose this application to the public internet. It is designed for local or trusted-network use only.

• Architecture — System design, database, caching
• API Reference — REST endpoints
• Development — Setup, configuration, workflow
• Testing — Running tests