Choose an installation path

ParseHawk has two supported bundled-runtime paths. Both expose the same Web UI, CLI, and REST API.

Host	Bundled model runtime	Recommended hardware	Setup guide
macOS on Apple Silicon	vLLM Metal on the host	16 GB unified memory minimum; 32 GB or more for larger contexts	Install on macOS
Linux x86_64 with NVIDIA	vLLM in Docker Compose	16 GB VRAM minimum; 24 GB or more for larger contexts	Install on Linux

Windows and Intel Macs are not currently supported for the bundled runtime.

Use another model server

You can run the ParseHawk API, worker, and Web UI without launching the bundled runtime:

parsehawk start -x runtime

Then point the openai_compatible_api provider at Ollama, LM Studio, a separate vLLM server, or another compatible endpoint. You can also configure OpenAI or Microsoft Foundry. See model providers for the exact setup.

What starts locally

A default parsehawk start exposes four services on loopback:

Service	Address	Purpose
Web UI	http://127.0.0.1:5173	Human workflow for files, extractors, and jobs
REST API	http://127.0.0.1:8000	Data plane used by the UI and CLI
Phoenix	http://127.0.0.1:6006	Local model-call tracing
Model runtime	http://127.0.0.1:8080	Bundled OpenAI-compatible inference server

Run parsehawk doctor after installation to check the host prerequisites and service health.

Next step

Once the stack is healthy, continue to your first extraction.