Local vs Cloud AI Companions

Most people meet AI companions as apps. A smaller community runs them at home, with open models and free software. This guide compares the two approaches on the points that matter to companion users: privacy, cost, hardware, features and effort. It is part of our AI companion privacy section. Facts about the local tools come from their official docs and repositories; we did not benchmark models.

What "Local" and "Cloud" Mean for AI Companions

A cloud companion runs the language model on the company's servers. Your phone or browser sends each message there and gets a reply back. That is how every app we review works, and why Nomi's wiki says end-to-end encryption "is not possible with an AI chatbot": the server has to read your message to answer it.

A local companion runs the model on your own computer. Three kinds of software are involved, and the difference matters for privacy:

• Front-ends are the chat interface where characters, personas and lorebooks live. SillyTavern is the best known.
• Runtimes load a model file and generate text, often with their own simple UI and a local API: KoboldCpp, LM Studio, Ollama and TextGen.
• Engines do the low-level computation inside those runtimes. llama.cpp is used by KoboldCpp and LM Studio and is a backend option in TextGen.

The key point, which SillyTavern's own README makes clear: a front-end is not a local model. SillyTavern can connect to local backends (KoboldAI/KoboldCpp, Ooba, Tabby) or to cloud APIs such as OpenAI, OpenRouter, Claude and Mistral. Chats only stay on your machine if the model runs there too.

The Main Local Tools

A few details from the docs are worth knowing. KoboldCpp ships as a single executable with no installation, and its README warns that koboldcpp.com is a fake site: download only from GitHub releases or koboldcpp.net. LM Studio's docs say "Nothing you enter into LM Studio when chatting with LLMs leaves your device", but the app needs the internet to find and download models and runtimes and to check for updates; it is now branded LM Studio Bionic and its homepage focuses on work agents rather than companions. Ollama's FAQ says "We don't see your prompts or data when you run locally", and its optional cloud models can be switched off with the OLLAMA_NO_CLOUD=1 setting. TextGen says it has "zero telemetry, external resources, or remote update requests"; older tutorials call it text-generation-webui, its former repository name, which now redirects to the new one.

Local vs Cloud AI Companions Compared

Privacy: Where Local Wins, and Where It Does Not

With a local model, the cloud risks covered in the rest of our privacy section simply do not apply: no retention policy, no moderators, no outside model providers, no breach of a company database. That is a real difference when you compare it with what cloud apps document. Kindroid, the most careful app we checked, still states that it is "not end-to-end encrypted" and lets authorized staff see flagged excerpts. Our guide on who can read companion messages covers the cloud side.

Local setups have their own leaks, though:

• Cloud backends: connect SillyTavern to OpenAI, OpenRouter or another API and your chats go to that provider. Ollama and LM Studio also offer optional cloud models; LM Studio's paid cloud plans claim zero data retention.
• Downloads: models, runtimes and updates come from the internet, and fake download sites exist (koboldcpp.com).
• Your own device: chats sit unencrypted on your disk unless you protect it, and a shared or synced computer is not private.
• Public character cards can contain adult or unlawful content; treat downloaded cards like any other file from a stranger.

Hybrid services sit in between. Janitor AI is a cloud platform, but you can connect your own model through an API proxy; its privacy policy says requests then go to that third-party provider, outside Janitor's own policy. That changes who processes your chats, not whether they leave your device.

What Cloud Apps Still Do Better

Convenience is the obvious advantage: a cloud app works on a phone in seconds. The feature gap is just as real. Kindroid documents voice calls and live video calls with an animated avatar, Nomi documents Mind Map memory and image-to-video, and Replika offers a 3D avatar with voice and video calls. Local tools can do voice and images too (SillyTavern supports TTS, image generation and Live2D/VRM through extensions, and KoboldCpp supports image and speech generation with compatible models), but you assemble and maintain it yourself.

Memory is another difference. A local model sees what fits in its context window; Ollama's default is 4,096 tokens unless you raise it. SillyTavern and KoboldCpp add World Info, lorebooks and author's notes, which work well but are manual. Cloud apps automate this, with varying success; our AI companion memory comparison shows what each documents.

On the other hand, local files cannot be taken away. When Soulmate shut down in 2023, users had about a week; Dot gave users until October 5, 2025 to export; and Character.AI's future depends on a bankruptcy sale. A model and a character card on your disk keep working regardless.

Character Cards: Moving Characters Between Tools

Local front-ends share a portable format, the character card: a JSON description of a character (name, description, personality, scenario, first message, example dialogue and more), usually embedded in a PNG image so one file carries both portrait and definition. Community-written V2 and V3 specifications define the fields; SillyTavern is built around cards, and KoboldCpp imports Tavern cards from files or external sites. Support for individual fields varies by tool, and these are volunteer specs, not an official standard. If you want to design your own character, our guide to building an AI character covers the structure.

Which Should You Choose?

Choose local if privacy is your first concern, you have a reasonably recent computer (the official notes point to 16GB of RAM or a GPU with 4 to 6GB of VRAM as recommended starting points), and you enjoy configuring software. Start with an all-in-one runtime such as KoboldCpp or LM Studio, then add SillyTavern if you want character cards and group chats.

Choose a cloud app if you want to chat on your phone, want voice calls, images or avatars without setup, or do not have suitable hardware. Then pick one with clear terms; our ranking of private AI chat apps and the list of AI companion apps for desktop are good next steps.