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

Privacy

Your chats never leave your device. Replies, pictures, voices and speech recognition are all generated inside the browser tab. No account is needed and there is no chat server, so there is nowhere for a conversation to be sent. The one exception is a choice you make: cloud replies while the model downloads, which are off unless you turn them on.

What the app downloads

• Its own files (HTML, CSS, JavaScript, character portraits) from local-character.skillsafe.ai.
• Model files from models.skillsafe.ai, only after you agree to each download. The requests name a file by its hash and carry nothing about you or your chats.

What the app stores on your device

• Characters, chats and personas in the browser's IndexedDB, and the documents you add to a character's Knowledge (their text only, read from the file in the page).
• Model files in the browser's Cache Storage, so they download only once.
• A few settings in local storage (for example which chat model is selected and its context length), all prefixed oa_.

Clearing this site's data in your browser deletes all of it. AI Models inside the app removes the downloaded models one by one, and has Export, which saves your characters and chats to a file on your own disk.

Microphone and voice clips

While the microphone is on, audio is recorded in the tab, transcribed by Whisper on the device and then discarded. A voice clip you load to match a character's voice is analysed the same way. The browser's built-in speech recognition is not used, because in Chrome it sends audio to a server.

The AI helpers run here too

Auto-fill on the Create page and Find with AI in the voice settings are written by the same chat model on your device by default. They send nothing and need no sign-in. Auto-fill also has a Model menu: if you pick a server model (Gemma 4 26B on Cloudflare, or GPT-Sol from OpenAI, both through SkillSafe), the request you typed is sent to that server to write the card - free, paid from the app's sponsored budget. The note under the box says which one is in use.

Cloud replies: only while the model downloads, only if you turn them on

The chat model is a 1.5 GB download. Once the download is running and the app has measured that it will take more than about 10 seconds (and less than 2 hours), the chat asks whether to use the cloud while waiting. If you say yes, replies come from a larger model, Gemma 4 26B, running on Cloudflare Workers AI through SkillSafe, until the download finishes; then the same chat continues on your device, with no further cloud requests. It is free for you: the app pays for those replies from its sponsored budget, and no sign-in is needed (an anonymous guest session is started).

For each cloud reply, the app sends that turn only: the character's instructions (name, description and the chat's rules), the conversation shown so far - in a long chat, also that chat's story-so-far notes and the earlier lines of it that match your message - and your new message. Memories, personas, a character's documents, other chats, pictures and voice are never sent. Cloud replies are marked "cloud" in the chat instead of "on device". While the on-device voices are still downloading, a cloud reply in English, Spanish, Chinese, Japanese or French is read aloud by a cloud voice (Deepgram Aura or MeloTTS on Cloudflare, through SkillSafe; also free) - only text that already went to the cloud is read this way. A reply written on your device is sent out to be read aloud only if you say yes when the app asks ("Read replies with a cloud voice while the voices download?", once per page); otherwise it is read by Kokoro as soon as its download finishes. The browser's built-in speech is never used. SkillSafe processes the request to produce the reply and Cloudflare runs the model; as with any API, the request passes through their logs. You can stop using the cloud at any time from the same card, and follow the download in AI Models. The choice lasts for that page only; nothing is sent until you say yes.

Phones and browsers without WebGPU

The on-device models need a computer with WebGPU (Chrome or Edge). On a phone, or in a browser that cannot run them, the app does not offer the downloads: replies come from the cloud model (Gemma 4 26B on Cloudflare, through SkillSafe), voices from the cloud voices, and Auto-fill from the cloud model too - free, paid from the app's sponsored budget, with no sign-in. The page says so ("Cloud replies on this device"). Each turn sends the same things as the cloud replies above. Pictures, speech input and voice-clip matching need the on-device models, so they are not shown there. For chat that never leaves your device, open Local Character on a computer with WebGPU.

Backup: only if you log in

Logging in to SkillSafe is optional. It only backs up your characters - their settings and portraits, including starters you edited - and your personas and narrator voices, to one private file in your own SkillSafe account (apps/local-character/characters.json in your SkillSafe drive). Only your account can read it, and you can delete it there at any time. Your chats, memories and characters' documents are never in it and never sent.

When you log in, characters and personas in the backup that are missing from this browser are added back (where both have one, the copy changed last wins), and after that each change to a character, a persona or the narrator voices is copied up. The Mine page shows where each character is saved (this browser, your account, Community) and lets you remove it from your account or delete it everywhere. Log out on this device from the same button; the backup stays in your account until you delete it.

Community: only when you use it

Community is the other part of the app that talks to SkillSafe's API, and only when you open it, search for a character on the home page (the list of shared characters is downloaded once and searched on your device; what you type is never sent) or press Share. Once you are logged in, the Mine tab also lists the characters your account has shared.

• Sharing a character needs a SkillSafe sign-in and sends only that character's settings file: name, tagline, description, greeting, appearance, category, voice, portrait, and the display name you choose (or none). Your chats, memories and personas are never sent. A shared character is public and is linked to your SkillSafe account so that only you can update or remove it. Removing it does not remove copies other people already added.
• Browsing Community downloads the list of shared characters. Adding one copies its settings file to this device and records one load, which is how characters are ranked. Loads are recorded under an anonymous identity kept in this browser, never your account, so the public load counts do not say who added what.

If you never log in, open Community, search on the home page, press Share or turn on cloud replies, the app makes no request to SkillSafe's API at all.

Hosting

The bundle contains no analytics, advertising or tracking code, and sets no cookies. The site is hosted by SkillSafe behind Cloudflare. The hosting layer keeps standard request logs (IP address, time, path, browser) and Cloudflare adds a bot-protection script to each page, which may set its own cookies. None of this can see your chats, because on-device chats are never sent.

Contact

Questions or corrections: local-character@skillsafe.ai. Changes are recorded in the change log.

FAQ

Open Local Character →