Privacy
This page is what the nanoMuse apps (the phone, the web and desktop apps, the dsh plugin) and nanoMuse Cloud do with your data. It is short because they do little with it. Where it matters, the one thing you decide is a switch: Settings → Data controls.
Everything stays on your device
The agent runs inside the app: on the phone, in a Linux root file system in the app's private storage; on a computer, in the runtime you installed. Conversations, memory, files the agent writes, scheduled tasks, skills and settings are stored there and, where you mount them, in folders you chose. There is no analytics, no crash reporting to us and no telemetry of any kind. The only nanoMuse server is the optional relay described below, and the apps work fully without it.
What leaves the device
- Your model provider. Messages, attached images and tool results are sent to the model endpoint you configured (Alibaba Cloud Bailian, DeepSeek, OpenAI, OpenRouter, your own vLLM or Ollama, …), under that provider's terms, with the API key you entered. The key is stored on your device only.
- nanoMuse Cloud, if you signed in. Sign in (free) on the first screen (or Settings → nanoMuse Cloud) signs you up with a phone number or an e-mail address and a code, and the app then uses our relay as a model provider: free, ¥10 of model use per account (community members: no limit; each new person who signs up with your invite code and makes a first request adds ¥5 to your allowance and ¥5 to theirs), paid by the developer; nanoMuse is a non-profit community project and never charges. The relay forwards chat, attached images and screen-operation requests to the operator's
Qwen/Qwen3.6-27B-FP8deployment on Tencent Cloud. It does not offer hosted image or video generation; those requests go only to a provider you configure with your own key. What it keeps about the account: a salted hash of your address, the address itself encrypted so the person running the relay can see whose account it is (the database file alone shows nothing), a masked hint such ass***@example.com, the hash of the key issued to this device, the kind, model and token count of each request (and, since relay 0.28, how each chat request went: how long it waited for a slot at the model, the time to the first token and to the end, how it ended, how many messages and pictures it carried, the names of the tools the app offered and the reply called, whether it was a screen-operation turn, and the client platform; never a word of the conversation; these rows are kept 90 days and the operator's pages read them in aggregate), your invite code with the ids of accounts that used it, and, so your devices wear the same one, the agent's name and look (which face; for one drawn in the avatar studio, its five small pictures and, from 1.0.7, the short clips a device made of it, so the others need not render their own). With each sign-in, request, timeline event and device it also records the network address the request came from and the client software (theUser-Agent: the Android app and its version, the runtime on Windows, macOS or Linux, a browser), and the account keeps its first and last address and when it was last seen; the person running the relay sees these on the operator's page, per account and per address, to tell one person on two accounts from two people on one address and to see a problem's shape. Of the conversations themselves it keeps, since relay 0.29, the record: every chat turn it answered, in full: the request as it sent it to the model (every message: the system prompt with your memory, SOUL and instructions, what you wrote, what tools returned, the screenshots and pictures) and the answer as the model gave it. It is kept as your account's own data, so that your devices can show the same history, screenshots included, and the person running the relay can see a problem whole. Pictures are stored once each as files named by their hash; the turn says which device and conversation they belong to when the app names them (1.0.7 on). Your devices read the record back through the relay; the operator can read it too; nobody else can, and a file opens only for the account whose turn it belongs to. It stays until you delete it (one tap under Settings → Data controls removes the whole record with the training turns described next) or delete the account; a relay you run yourself can set a retention period (RECORD_KEEP_DAYS) and keeps no record at all withRECORD_TURNS=0. The pictures have a ceiling in all (MEDIA_MAX_GB): past it the least recently used are removed first and the words stay. From 1.0.7 an app that ran a turn on a key of your own sends that turn up as well, so the record is whole wherever the model was. From relay 0.31 the apps also say, with each turn, where it came from: whether a person started it or a scheduled task, the feed's routine, a goal's check or another of your devices; which turn of the conversation it is and which model call within the turn; which screen-operation task and step it belongs to; and the relay notes one word for the script of what you wrote (zh,latin,ja, …), never the words. And the apps report, as short codes and numbers tied to the turn and never as text, what happened around it: what the Sentinel decided about a tool call and at which risk level, how an approval went, why you stopped a reply when you chose to say, a thumb up or down you gave an answer, how a screen-operation step went. Those reports are deleted with the turns and with the account. Two switches, described next, decide what else is made of the conversations. It does not receive your location: the apps never send it. Next to an address, the operator's page shows the country, province and city that address belongs to, read from an offline copy of ip2region's public database on the relay's own disk when the page is drawn; no third party is asked and the place is not stored. Signing out revokes the key; deleting the account removes all of it, addresses included. The relay's source is in the repository undercloud/, and docs/cloud.md says how to run your own. Nothing about the relay applies when you use your own key. - The project site (nanomuse.cn), if you visit it. Its web server keeps an access log for seven days and then deletes it; a script turns it into counts per day: pages viewed, how many visitors, downloads per file, which sites linked here. A visitor is counted once a day through a hash of the address and browser string under a random salt made for that day and discarded after two; no address is stored beyond the week the raw log lives, nothing is shared with anyone, there are no cookies and no third-party analytics. The relay's own operator page shows counts of sign-ins and accounts per day, from the same records described above; nothing more is collected for it.
- The showcase (demo.nanomuse.dev), if you try it. The page sets no cookie and its web server keeps no access log. Starting a Muse there means signing in to nanoMuse Cloud first, through the showcase's gateway: the relay records that sign-in as described above, with your address, which the gateway passes on. The gateway's own database then keeps, per account: the account's id, the relay's masked hint, whether it was a phone number or an e-mail address, how many times you signed in, and the address and browser string of your first and your latest sign-in or demo; and one row per demo started: when, from which address, with which browser, whether you brought your own key, how it ended, and how many model calls, tokens, pictures and clips it used. Your browser holds a sign-in ticket good for thirty days; the gateway stores its hash. What you say to the Muse goes through the gateway to the model the operator configured (Alibaba Cloud Model Studio on the public showcase); the gateway counts calls and tokens and keeps no message content, and a key of your own lives in its memory for that session only and is never written down. The Muse runs in a container made for your session and removed with it, and everything in it, after thirty minutes or ten quiet ones; your browser's time zone name travels with the start so the Muse's clock is yours, and is not kept. The gateway's log names sessions by id and visitors by the masked hint, never by address, and Docker keeps at most 100 MB of it. The account and demo rows stay until the operator deletes them; there is no automatic deletion. The person running the relay sees them on its operator page, beside the account. The source is under
demo/showcase/in the repository, and its README says what each limit is. - Data controls: "Help improve nanoMuse's AI models". Settings → Data controls on every app holds one switch. While it is on, the text of your chats with the nanoMuse Cloud models may train the community's own open model: what you wrote, what the model answered, and the tool calls it chose, together with the model, the token counts, and the app's platform and language from the request headers. The training set does not take the system prompt (your memory, SOUL and instructions), what tools returned, pictures, audio or clips (those are replaced by a marker), and its turns are tied to your account id only; exports for training carry no account id, address or hint. Beside each kept turn the relay stores its shape (since relay 0.28): the number of messages, character counts, tool calls and pictures, whether the prompt was trivially short, and a hash of what you wrote, so repeats can be counted without reading the turns; deleting the turns deletes the shapes. From relay 0.31 a turn the app ran on a key of your own and sent up to the record joins the training set in the same way while the switch is on, and each kept turn says where it came from (a person or a routine, which turn and step, on the relay's model or your own key), so the thumbs and stop reasons above can be matched to it. Exports for training name your account by a pseudonym (a keyed hash the relay alone can make, which recovers nothing without its secret); when you delete your kept data or your account, the relay writes that pseudonym and the time to a deletions ledger and nothing else about you stays, so a set built from an earlier export drops your turns in its next build. The switch decides what joins the training set, not what the record above holds. On
cloud.nanomuse.cnthe switch is on for accounts created from relay 0.9 on; accounts that existed before keep whatever they had chosen. The switch says so right next to it, and a relay you run yourself chooses withIMPROVE_DEFAULT. Turn it off at any time (nothing more joins the set from then on), delete everything kept so far, the training turns and the record, with one tap on the same page, and deleting the account deletes it too. Nothing about the allowance depends on this switch. Apps older than 1.0.7 describe the switch in the earlier words, with the record not yet mentioned; this page is current. - Synced conversations: on by default when you are signed in (relay 0.19 on). The text of your conversations is stored on nanoMuse Cloud so every device of your account shows the same chats. Files and images are not uploaded by this switch (the record above holds the pictures of each turn). Settings → Data controls turns it off and deletes what is stored; deleting a chat on one device deletes it on all of them. The text is stored as it is, so, as with the record, your devices and the person running the relay can read it and nobody else; the operator's page shows counts, not text. (What is stored, exactly: each conversation's title and which device started it, each message's role, text, time and device, and the names and sizes of attached files; never the files, never a picture, never what a tool returned. Chats addressed to another device, or run on this one for another, are not synced. The relay keeps at most 20 000 messages per account and cuts a message at 16 KB. docs/cloud.md has the details.) Since 0.1.38 only the main conversation is synced by default: side chats stay on the device that made them, and other devices' side chats do not arrive, unless you turn on Also sync side chats on that device. While one of your devices is answering, the others are told so (kwai is working…); that note goes through the relay's memory, is never stored, and is forgotten after ten minutes.
- Your devices, when they talk to each other (relay 0.9 on). With Devices on, each signed-in app keeps one connection to the relay, so your phone can ask your computer to run something and the other way round. The relay keeps a row per device: the id the app chose for itself, the name you gave it, whether it is a phone or a computer, its operating system, the app's version, the list of actions it offers, when it was first and last seen, and the address it connected from; your other devices see the same list (name, kind, version, online or not, address) under Settings → Devices. From 1.0.7 the row also carries two ids for what answers that device's chats:
cloudwhen the account's models do, otherwise the provider and model you configured (openrouter,bailian, a model id), never the key, the base URL or any other setting; the operator's page reads them to see which providers people configure for themselves. What one device asks another (a command, a file, a screen to operate, a provider request your phone sends through your computer) passes through the relay inside your account and is not stored; the relay's log notes the device kind and the first characters of the device and account ids, never the content. Forget removes an offline device's row, and signing out takes the device off the hub. - The web, when the agent uses it. Web search, page fetches, the in-app browser, MCP servers and command-line tools reach the sites and services they are for. What the agent sends is what you asked it to do.
- Update check. Phones, the standalone harness and older desktop shells ask the project's download index at
nanomuse.cn/dl/index.json, then the GitHub API when it does not answer. A runtime you run yourself (nanomuse serve, Docker) checks once a day;server.update_check = falsestops that check. The Windows x64 desktop app checks Windows update metadata after startup and every six hours, first on the download mirror and then on GitHub, with a manual refresh in Settings. Its update library sends a random local staging identifier with metadata requests, and theUser-Agentnames the app and its version. Automatic downloads are on by default and can be switched off in Settings. Windows' network cost is read locally; metered or unknown connections wait for a manual download, and this classification is not sent to the update sources. No conversations, account credentials or settings are sent to the update sources. - Backups. If you back up to a remote destination (SMB, WebDAV, SFTP, S3, FTP), the backup goes there, encrypted with the password you chose.
Permissions
Each permission is asked for when a feature needs it and is used for that feature only: notifications for the agent's status and reminders; accessibility for operating other apps' screens, which is off until you turn it on; storage folders you mount; the microphone for voice input; contacts, calendar and location for the tools of the same names, each callable only after you granted them. Nothing is read in the background.
What the agent may do without asking
Before anything it cannot take back (deleting your files, sending a message or data out, paying), the agent stops and a card asks you, in the shell, in the browser and on the phone's screen alike. What you may remember from that card comes in three tiers, and Settings → Permissions → Remembered approvals lists everything you remembered, grouped the same way, one line each, tap to revoke:
- Runs, then tells you: installing software. Never asks; the agent says so afterwards.
- Asks first; you may remember it: deleting and sending. Allow for this chat or Always allow for X (one folder, one recipient, one site or app).
- Highest: asks at the moment of paying, every time: buying, ordering, booking, trading, transferring. There is no "for this chat". Remember and run next time in X exists, for one app or site only, and asks for your screen lock (fingerprint, face, PIN) before it takes; it is listed first on the permissions page, and every payment that runs on it is said out loud in the chat.
Passwords, verification codes and card numbers are never typed by the agent; those screens are handed to you. Anything alarming in a command (wiping a disk, force-pushing history, a download piped into a shell) is asked about every time and cannot be remembered. Shopping and trading are fine when you asked for exactly that; nothing unlawful or harmful is done, however it is worded; the agent refuses and says why.
Feedback
Bug reports go to GitHub Issues from Settings → Feedback; the report is pre-filled with the app version and the device model and nothing else. Do not paste API keys or private conversations into an issue.
Changes
This page changes when the apps' behaviour changes; the history is in the repository. The same text is at nanomuse.cn/privacy, in English and Chinese.