The XDA writeup on turning an old laptop into a private AI assistant for the whole house is the case study a lot of us needed. A 2020-era ThinkPad with 32 GB of RAM and an integrated GPU is enough to run a modern 7B or 13B model at conversational speed, host a chat UI for the household, and stop paying $20 a month per person for cloud AI. These are the best desktop apps for building a whole-house private AI assistant on an old laptop in 2026, chosen for the combination of runtime, chat UI, and household-friendly access that makes local AI actually useful outside the operator’s own terminal.
What to look for in a household AI stack
A private AI assistant that the whole house actually uses needs to cover:
- Efficient inference on modest hardware. A runtime that squeezes acceptable tokens/sec out of integrated graphics or an older discrete GPU.
- Multi-user chat UI. A web interface where anyone in the house can log in, keep a separate history, and not step on each other’s runs.
- Model library management. Downloading, updating, and switching models without a command line.
- API-compatible server. OpenAI-compatible API endpoints so other household apps (email, notes, home automation) can hook in.
- Voice input and output. Optional, but the household adoption jump when voice works is real.
- Home automation integration. Hooks into Home Assistant, HomeKit, or Google Home for spoken queries.
- Privacy defaults. Nothing leaves the laptop unless the operator explicitly turns something on.
Quick comparison
| App | Best for | Runtime | Chat UI | API server | Cost |
|---|---|---|---|---|---|
| Ollama | The base runtime everyone else builds on | Ollama | CLI + community UIs | Yes | Free |
| LM Studio | GUI-first local models | llama.cpp | Yes | Yes | Free |
| Open WebUI | The household chat UI in front of Ollama | Any (Ollama, OpenAI) | Yes (multi-user) | Proxies | Free |
| GPT4All | Simplest single-user local chat | llama.cpp | Yes | Optional | Free / Pro $5/mo |
| Jan | Open-source ChatGPT replacement | llama.cpp / Ollama | Yes | Yes | Free |
| Text Generation WebUI | Power-user model tuning and format support | Multiple | Yes | Yes | Free |
| Home Assistant with Local LLM | Voice-driven home AI over Ollama | Ollama backend | Voice + Chat | via Ollama | Free |
The 7 best apps for private household AI on an old laptop
1. Ollama, best as the base runtime everyone else builds on
Ollama is the runtime that made local LLMs a one-command install. ollama run llama3.1 downloads, loads, and serves a model with an OpenAI-compatible API on port 11434. Model library is curated (Llama, Mistral, Phi, Qwen, Gemma, and community fine-tunes), and the runtime auto-detects and uses NVIDIA CUDA, AMD ROCm, or Apple Metal.
On a laptop with an RTX 3060 (or Apple M2 and up), a 13B model runs at conversational speed. On CPU-only, 3B-class models like Phi-3 Mini stay usable.
Where it falls short: Command-line only by itself; needs a chat UI on top for household use. Curated model list is narrower than the raw GGUF universe. Docker install is preferred on Windows.
Pricing:
- Free and open source (MIT)
Platforms: Windows, macOS, Linux
Download: ollama.com
Bottom line: The default runtime for any private-AI household setup in 2026.
2. LM Studio, best for GUI-first local models
LM Studio is Ollama’s GUI competitor. Point-and-click model download from Hugging Face, a chat panel with per-conversation model selection, and an OpenAI-compatible server that starts with a checkbox. LM Studio ships with per-model recommendations for context length, quantization, and GPU offload count.
For an operator who prefers a Windows-style app over a terminal, LM Studio is the shortest path from a fresh install to a working local model.
Where it falls short: Single-user by design; not a household chat UI. Model management is per-machine, not centralised. Not fully open source (proprietary desktop app around open runtimes).
Pricing:
- Free for personal use
- Business licensing on request
Platforms: Windows, macOS, Linux
Download: lmstudio.ai
Bottom line: The right pick for someone learning local models before setting up a multi-user server.
3. Open WebUI, best as the household chat UI in front of Ollama
Open WebUI is the piece that makes an Ollama server usable by the whole house. Multi-user accounts, per-user chat history, model-switcher, RAG document upload, prompt library, and role-based permissions all live in a self-hosted web UI. Point it at an Ollama endpoint (local or across the LAN) and every family member gets a ChatGPT-style experience without sharing history.
Runs as a Docker container on the same laptop as Ollama, or on a separate machine talking to it over the LAN.
Where it falls short: Docker install is the recommended path (bare Node is possible but harder). Voice input is via Whisper add-on, not native. Storage of chat history grows quickly with heavy use.
Pricing:
- Free and open source (MIT)
Platforms: Docker, Linux, Windows (via Docker Desktop), macOS (via Docker Desktop)
Download: docs.openwebui.com
Bottom line: The single most important add-on for turning Ollama into a household service.
4. GPT4All, best for the simplest single-user local chat
GPT4All is Nomic’s local-first chat app: download the installer, pick a model, chat. The UI is built for someone who has never touched a terminal, and the model library covers Llama, Mistral, Phi, and Falcon variants tuned by the community.
Good option for a parent or non-technical partner who wants their own local AI without setting up a server.
Where it falls short: Not a shared server; each machine has its own model and history. Model library is smaller than Ollama’s. Pro tier adds cloud features that undermine the “fully local” premise.
Pricing:
- Free for personal use
- Pro: $5/mo for cloud features (optional)
Platforms: Windows, macOS, Linux
Download: gpt4all.io
Bottom line: The right pick for a single laptop where one person wants local AI and nothing else.
5. Jan, best for an open-source ChatGPT replacement
Jan is a fully open-source desktop chat app that runs local models via llama.cpp or connects to remote endpoints (OpenAI, Anthropic, Groq, or a local Ollama). The interface is close to ChatGPT’s, with chats, folders, and system-prompt presets, and Jan’s philosophy is “your data stays on your machine.”
For a household where different family members want different back ends (some local, some remote), Jan handles both cleanly.
Where it falls short: Newer project; some rough edges on Windows install. Multi-user support is thinner than Open WebUI. Some plug-ins are experimental.
Pricing:
- Free and open source (AGPLv3)
Platforms: Windows, macOS, Linux
Download: jan.ai
Bottom line: The right pick when a household wants a ChatGPT-style app that can be local or hybrid.
6. Text Generation WebUI, best for power-user model tuning and format support
Text Generation WebUI (oobabooga) is the tool for power users. It supports Transformers, ExLlamaV2, llama.cpp, and MLX back ends, exposes every sampler and inference setting, and handles GGUF, GPTQ, EXL2, and AWQ formats. LoRA training and character-card RP are built in.
The right pick when the operator wants to squeeze every token/sec out of an old GPU or run niche fine-tunes not in Ollama’s library.
Where it falls short: Steep learning curve. UI has years of stacked features that overwhelm new users. Multi-user support is limited.
Pricing:
- Free and open source (AGPLv3)
Platforms: Windows, macOS, Linux
Download: github.com/oobabooga/text-generation-webui
Bottom line: The right pick when the goal is control over every inference detail.
7. Home Assistant with Local LLM, best for voice-driven home AI over Ollama
Home Assistant with the Local LLM integration turns an Ollama server into the voice-first household AI: ask about the weather, control lights, query the calendar, or hold a normal conversation through a smart speaker or the HA app. Assist Pipeline handles Whisper transcription, LLM inference on Ollama, and TTS back to the speaker, all inside the home network.
For a family where “hey house, what’s on the schedule today” is the desired UX, this is the layer that ties it all together.
Where it falls short: Setup crosses two projects (Home Assistant + Ollama) and needs both healthy. Speaker choice matters (Wyoming-compatible mics are the smoothest). Latency depends on the model; smaller models feel snappier.
Pricing:
- Free and open source
Platforms: Home Assistant OS, Docker, Linux, Windows, macOS
Download: www.home-assistant.io and Ollama integration
Bottom line: The right pick when voice-first household AI is the goal, not just a chat window.
How to pick the right one
For a fully household-friendly setup, install Ollama as the runtime and Open WebUI as the multi-user chat UI in front of it. This is the correct stack for a laptop-turned-server.
For the operator’s own single-machine use, LM Studio is the easiest install and covers most day-to-day needs.
For a non-technical family member with a laptop of their own, GPT4All is the friendliest single-user install.
For a household split between local and remote AI back ends, Jan is the correct chat UI.
For power-user tuning and niche model formats, Text Generation WebUI is the deep tool.
For voice-first home AI, Home Assistant with the Local LLM integration ties Ollama into the household voice ecosystem.
FAQ
What old laptop is enough for a household AI server? Any 2020-or-newer laptop with 16 GB of RAM or more can run a 7B model at usable speed on CPU. Add a discrete GPU (even a 2018 GTX 1660) and 13B runs at conversational speed.
Does local AI beat ChatGPT on quality? Not at the top of the model class. A well-quantized 13B model is roughly GPT-3.5-class; a 70B model approaches GPT-4-class but needs 24 GB or more of VRAM. For most household tasks (drafting emails, summarising documents, brainstorming), the local models are more than good enough.
Do local models stay updated? The runtimes update frequently (weekly release cadence on Ollama). The models themselves are versioned; updating means downloading a new tag.
How much bandwidth does the initial setup use? A 7B model is 4–6 GB; a 13B model is 8–12 GB. Plan for 20 GB of downloads if the operator wants a small library.
Can the household use this without VPN when travelling? Yes, but only if the operator exposes the server to the internet. Tailscale is the cleanest way to reach the home AI server from outside the LAN without opening ports.
Is my chat history really private? Only if the runtime is set up correctly. Ollama, LM Studio, GPT4All, and Jan all default to fully local; Open WebUI stores history in its own database on the same machine. Turning on any cloud connector (Pro tiers, remote model back ends) sends data outside the box.
