Running AI Drafting on Hardware You Own

If sending findings to a hosted model is not acceptable to your clients, or not to you, the platform will talk to a local model instead. It is not a reduced mode: the same features work, on your own hardware, with no rate limits and nothing leaving the machine.

Understand this first

The portal runs in the browser and calls the model directly. Nothing is proxied through a Pental server, which is the whole point, but it means the model has to accept a cross-origin call from your portal’s address. Almost every failure below is that one fact showing up in a different disguise.

One machine per tester

• Install Ollama and pull a model.
• Set OLLAMA_ORIGINS to your portal’s origin. The AI settings screen prints the exact value for your hostname, so copy it rather than typing it.
• On Windows, set it in PowerShell and then restart Ollama FULLY from the tray. Closing the window is not restarting it, and this is the most common reason it still will not connect.
• On Linux, add it as an Environment line in the systemd unit and restart the service.

Or one shared machine

A single well-specified box for the team is often better value than a model on every laptop. The catch is that a browser will not make a plain HTTP request from a page served over HTTPS, so the shared machine has to be reachable over HTTPS as well. Tailscale will expose it with a single command, which the AI screen gives you; any other way of terminating TLS in front of Ollama works just as well.

When it will not connect, check in this order: is Ollama running, was it fully restarted after the origins change, is the origin exactly right including the scheme and port, and is the connection HTTPS if your portal is.

Choosing a model

Report writing rewards instruction-following more than breadth of knowledge, because every field is generated against a prompt you can read and edit. Start with a mid-sized instruction-tuned model, set the token limit to around two thousand, and switch on keeping the model resident on the GPU if it is fourteen billion parameters or more. Then generate a handful of findings from real evidence and compare them with what your team would have written. That afternoon settles the question better than any benchmark.