Short answer: My research agent is a self-hosted harness on my own machine, driven from the terminal, pointed at whichever model fits the task — a cloud model for hard reasoning, a local one through LM Studio for anything sensitive. It reads sources, extracts and summarises, and drafts, while the prompts, transcripts, and files stay with me. It’s a workflow habit, not magic: the value is that it’s always there, shaped how I work, and nothing leaves the box unless I send it.

The own-the-agent argument is easy to nod along to and harder to picture in daily use. So here’s the concrete version: how a self-hosted agent actually fits into my week, and what it does for research specifically.

The setup, in plain terms

I run an agent harness locally — Pi is the one I reshape most — driven from the terminal. Two deliberate choices make it mine:

• Bring-your-own-model. The harness points at whatever model suits the task. Hard reasoning gets a strong cloud model; anything with sensitive content runs against a local model through LM Studio, so the text never leaves my machine.
• My tools, my rules. I give the agent exactly the tools I want it to have — read these files, search the web, run this script — and nothing else. The permissions are mine to set, not a product’s to decide.

That’s the whole trick. No dashboard, no account, no history sitting on someone else’s server.

What it does for research

Research is where an agent earns its keep, because most of it is gather-then-condense:

• Sweep and summarise. Point it at a set of sources and get back a condensed, cited digest instead of twenty open tabs.
• Extract from documents. Pull the specific facts, figures, or clauses I need out of long PDFs and reports, into a structure I can use.
• Draft against my notes. Turn a pile of rough notes into a first draft I then edit, in my own voice, grounded in what I actually wrote.
• Interrogate a corpus. Ask questions across a folder of documents and get answers with the sources attached.

None of that is unique to a self-hosted agent. What is unique is that the sources, the questions, and the outputs never become someone else’s training data or someone else’s analytics.

Why self-hosted specifically

Two reasons, both practical:

1. Sensitive material stays put. Client documents, draft strategy, anything under NDA — that goes to a local model on my own machine, not to a third-party product. The own-your-stack case is strongest exactly here.
2. It’s shaped to me. Because I own the harness, the agent behaves the way I want after months of small adjustments. A rented product resets to its defaults; mine accumulates my preferences.

The honest part

This is a habit, not a magic button. It’s worth setting up if you do enough research or drafting that an always-there, owned agent pays back the effort of running it. If you ask a chatbot something twice a month, don’t bother — a rented tab is fine.

If you want the fuller picture, Pi vs OpenClaw covers the terminal agent versus the chat-channel assistant. And if you’d like an owned agent or automation built for your business and handed over, that’s what I do at Leinss Consulting.