Run DRAKULIS locally

DRAKULIS is a development-stage independent agent runtime. It does not require a proprietary cloud account for local operation. Start with a local model and the read-only tools, then inspect the evidence behind each task.

Requirements

Python 3.12+, Node 22.12+, and Ollama with a provisioned tool-capable model. Dependency and model installation initially need connectivity. Model quality and tool behavior depend on the model and inference server version; a small model is useful for testing but is not a production quality guarantee.

Install from your checkout

No public repository URL or package release is being advertised yet. From your existing source checkout:

python3 -m venv .venv
.venv/bin/python -m pip install -e '.[dev]'
npm --prefix apps/studio ci --no-audit --no-fund
ollama pull qwen3:0.6b

Keep Ollama running with cloud features disabled:

OLLAMA_NO_CLOUD=1 OLLAMA_HOST=127.0.0.1:11434 ollama serve

In a second terminal:

DRAKULIS_MODEL=qwen3:0.6b .venv/bin/python scripts/local_stack.py

Open http://127.0.0.1:5173. The supervisor creates a private .data/local-settings.json containing distinct operator and worker tokens. Enter the operator token in Studio. Never share this file or commit it. Ctrl-C stops the supervisor's own services.

Native CLI

DRAKULIS_MODEL=qwen3:0.6b .venv/bin/drakulis \
  --expected-number 576 'Use the calculator to calculate (128+64)*3'

The expected-number check requires both a real calculator result and the same number in the answer. A plausible answer without a tool result fails verification. Ordinary chat without success criteria is explicitly unverified.

Read a document

Upload a small UTF-8 document in Studio, select Read + cite source, and enter its exact filename. Ask the agent to read it. Approve the specific call when requested. The initial check confirms a successful read and filename citation; it does not prove that every generated sentence is entailed by the document.

Privacy and providers

Local operation defaults to local-only task privacy. Switching the environment to online mode alone does not permit a task to send its context to a remote model. Cloud execution additionally needs explicit per-task cloud_allowed consent (CLI --allow-cloud). Studio currently submits local-only tasks.

See offline mode, model gateway, verification, and the security model.