Ardur Runtime governance for AI agents

Ardur

Get Started

Install Ardur and run your first governed AI session in 5 minutes.

Pick your path

Ardur works anywhere Python 3.10+ runs. Choose the setup that matches your setup.

Mac (Apple Silicon or Intel)

# 1. Clone the repo
git clone https://github.com/ArdurAI/ardur.git
cd ardur/python

# 2. Create a virtual environment
python3 -m venv .venv
source .venv/bin/activate

# 3. Install dependencies
pip install pyjwt cryptography

# 4. Verify it works
PYTHONPATH=. python -c "from vibap.passport import generate_keypair; generate_keypair()"

Done. You can now issue mission passports and run the governance proxy.

Linux (Ubuntu / Debian / Fedora)

# 1. Clone and set up Python
git clone https://github.com/ArdurAI/ardur.git
cd ardur/python
python3 -m venv .venv && source .venv/bin/activate
pip install pyjwt cryptography

# 2. Optional: build the Go AAT engine
cd ../go && go build ./...

VM / Sandbox / Remote Server

Same as Linux above. The proxy listens on 127.0.0.1 by default — if you need remote access, set up an SSH tunnel or reverse proxy. The proxy supports mutual TLS for production deployments.

Docker (coming soon)

A Docker Compose file and prebuilt images are on the roadmap. For now, clone the repo and run directly.

Connect your AI agent

With Ollama (local models)

Ardur works with any model running in Ollama. The proxy is provider-agnostic — it evaluates tool calls, not model outputs.

# Start Ollama with a local model
ollama pull <your-model>
ollama serve

# Run the governance proxy
PYTHONPATH=python python -m vibap.cli hub start

With Ollama (cloud models)

For larger models via Ollama’s cloud API:

export OLLAMA_API_KEY="your-api-key"

# Run the full governance test
PYTHONPATH=python python tests/run_cloud_model_test.py "$MODEL_NAME"

This runs a real-world test: a cloud model builds a complete web application while every tool call goes through Ardur’s governance check.

With Claude Code

Ardur ships a native Claude Code plugin:

# Initialize a mission profile
PYTHONPATH=python python -m vibap.cli profile init

# Protect your Claude Code session
PYTHONPATH=python python -m vibap.cli protect claude-code

See the Claude Code plugin README for the full setup.

With LangChain / LangGraph / AutoGen

Runnable quickstarts live in the examples directory:

Run your first governed session

Here’s the shortest end-to-end path:

# 1. Start the governance proxy
cd python
PYTHONPATH=. python -m vibap.cli hub start

# 2. In another terminal, issue a mission passport
PYTHONPATH=. python -m vibap.cli issue \
  --agent-id "my-agent" \
  --mission "read files in /tmp and write reports" \
  --allowed-tools read_file write_file \
  --resource-scope /tmp \
  --max-tool-calls 50

# 3. Use the token to start a session
# (The CLI prints the token — copy it)
curl -k -X POST https://127.0.0.1:<port>/session/start \
  -H "Content-Type: application/json" \
  -d '{"token": "<your-token>"}'

# 4. Evaluate tool calls through the proxy
curl -k -X POST https://127.0.0.1:<port>/evaluate \
  -H "Content-Type: application/json" \
  -d '{"session_id": "<session-id>", "tool_name": "read_file", "arguments": {"path": "/tmp/test.txt"}}'

Each /evaluate call returns PERMIT or DENY with a signed receipt.

Next steps