A self-hosted AI assistant runs its brain on infrastructure you manage. A hosted service runs that brain for you. With Jarvis, either approach can connect to companions on your computers. The choice changes who manages the runtime and where its working data is processed.
Choose who runs the brain.
Self-hosted: install the source-available runtime on your computer, home server or VPS. You choose the model providers and manage updates, access, storage and recovery. There is no Jarvis software subscription, but hardware, hosting and model usage can cost money.
Hosted: we run the brain and include an AI allowance. Plans currently start at $29.99/month. You still install the desktop companion, connect the computers you want to use, and configure permissions and workflows.
Compare the current hosted plans or follow the self-hosting installation guide.
Follow the data, not the label.
Self-hosting places the brain and its knowledge database on infrastructure you control. Your configured model provider can still receive prompts and context. Screenshot analysis and speech may use external providers too. A locally hosted brain does not, by itself, make the whole system offline.
On a hosted plan, the companion sends observations and relevant work context to the hosted brain. Its memory and conversation state live on the hosting servers. Configured AI providers may process prompts and screenshots. The apps run on your computer; that does not mean all information stays there.
If local processing is a requirement, check model routing, vision, speech, telemetry and every connected tool. The privacy policy explains data flows and controls; the provider documentation explains model choices.
Hosting keeps the brain available.
An always-on server can keep the brain running when your laptop is closed. A workflow that needs your laptop’s browser, files or desktop still needs that machine to be awake, reachable and authorised. Hosting does not wake or operate an unavailable computer.
The same distinction applies when you self-host on a VPS: the brain may be online while one of its connected machines is not. Build failure handling around the actual source or action the workflow uses.
Count the maintenance as well as the bill.
- Self-hosted: plan for runtime updates, provider usage, machine uptime, access control and backups you arrange and test.
- Hosted: consider the monthly plan and included allowance, what happens at its limits, and the permissions required by your connected computers.
- Either way: someone owns workflow testing, failed runs, changed apps and reviewing AI output.
There is no universal break-even point. Use the same task and approximate monthly frequency to estimate your model and infrastructure costs. Avoid treating a self-hosted software price of zero as a total cost of zero.
Choose a first setup you can maintain.
Hosted is a useful starting point if you want to work on the routine without running a server. Self-hosting fits when choosing infrastructure, providers and runtime configuration is part of the value you need, and you can maintain them.
For self-hosting, start with the supported installation path and confirm your model settings before connecting real accounts. If you later change deployment, plan the data transfer with the supported export or support process. Do not assume migration is automatic or has no downtime.
Test one routine before expanding access.
Try preparing a meeting brief or drafting a follow-up from information you supply. Check the result, missing-data behaviour and required permissions. Then add the sources and triggers that make it repeatable.
By the Jarvis team. This article was first published on 4 August 2026 and revised on 30 September to clarify data flows, costs and connected-machine requirements.
Sources: installation, desktop control, model providers, privacy and pricing.