Local Character › About · Last reviewed 2026-10-09

Local character AI, with nothing to install

Local Character is a character chat app whose AI runs on your own device, inside a browser tab. You talk to role-play characters the way you would on character.ai, but the replies, pictures, voices and speech recognition are all generated locally on your computer's GPU. There is no account, no install and no server that reads your chats.

What "local" means here

Most character AI sites send every message to a cloud model and store the conversation on their servers. Local character AI usually means the opposite: a model you run yourself, typically by installing a desktop program such as Ollama, llama.cpp or SillyTavern and downloading model files.

Local Character gives you the second without the setup. The browser downloads the model once, keeps it in its own storage, and runs it with WebGPU, the browser's direct access to your graphics card. Your messages are processed in the tab and written to the browser's own database, so they stay on the device you typed them on.

What it does

• Chat with characters - 55 ready-made characters on the Discover page, or create your own with a name, greeting, personality and voice.
• Group chats - add up to three more characters to any chat with the + next to the character's name at the top. One character answers each message: the one you name, the one you tap, or the one the model judges should answer. Each speaks in their own voice, and what you tell the group, every member remembers. Press Let them talk for 1 to 8 rounds, or tap a member with the message box empty, and the characters talk to each other without you.
• Replies by Google's Gemma 4 E2B, a 2-billion-parameter model, running on LiteRT-LM over WebGPU.
• Scene pictures by DreamShaper 8 (a Stable Diffusion 1.5 fine-tune), drawn on the device.
• Spoken replies by Kokoro-82M, with 28 English voices you can mix.
• Eight languages. English, Simplified and Traditional Chinese, Japanese, Spanish, French, Italian and Portuguese: characters reply in the one you write or speak, Auto-fill writes the card in it, speech input recognises it, and Kokoro speaks it with that language's own voices. These are the languages all three on-device models handle well; anything else is answered in English. A language's pronunciation data (1 to 2.4 MB) downloads only the first time a reply in it is spoken. Pick the interface language with the globe at the bottom of the sidebar.
• Talk instead of type - turn the microphone on and just talk: Whisper transcribes you on the device, the character answers aloud, and it keeps listening until you turn the mic off.
• Edits and personas - edit any message, including the character's replies, to see how a different reply changes where the story goes; start a new chat from any point, and tell characters who you are.
• Export and import - save all characters and chats to a file and load them in another browser.

How to start

1. Open local-character.skillsafe.ai in Chrome or Edge 137 or newer.
2. Pick a character. The first time, the app asks before downloading the chat model (about 1.5 GB, once).
3. Chat. Pictures, voices and speech input each ask for their own one-time download when you first use them.

Requirements

All four together take about 2.9 GB of browser storage. Safari and Firefox cannot run the chat model yet, because it needs WebAssembly JSPI and relaxed SIMD; the app says so on the Discover page instead of failing later.

What it gives up

A 2B model that fits in a browser is much smaller than the cloud models big character sites use. It writes shorter, simpler replies, and in very long scenes it can lose track of earlier details. There is no community library of user-made characters, and no group chat. See local vs cloud character chat for the full comparison.

FAQ

Open Local Character →