Nytka for Android

The phone half of Nytka, a companion app and self-hosted server for the Omi AI necklace. The app takes the pendant's audio over Bluetooth, keeps it in a queue on the phone and uploads it to your own Nytka server, which transcribes it and serves the conversations back. It shows the titles, summaries and tasks your server's language model writes, the memories it keeps and a search over all of it, and it copies the audio the pendant stored while the phone was away. Omi's cloud sees none of it.

Status: v0.4 of the roadmap is shipped: capture and transcription (v0.1), the AI layer (v0.2), offline sync (v0.3) and memory and search (v0.4). Each spec says what its version does and does not do. Roadmap milestones are not release numbers: the app and the server each have their own.

How Nytka compares

As of 2026-10, against the official Omi app and two other pendant/voice-journal apps. Nytka's and Omi's rows come from this project and Omi's own published free tier; Chronicle's and LocusTether's are not independently verified here, so confirm them yourself before relying on this table.

	Nytka	Omi (official app)	Chronicle	LocusTether
Self-host effort	Your own server, Docker Compose, about 15 minutes	None — cloud only, no self-host option	Not verified	Not verified
Choose your models	Yes — any transcription and LLM endpoint you set in the server	No — Omi's own cloud models	Not verified	Not verified
Usage quotas	None — your server, your limits	300 minutes and 30 questions a month free, then paid	Not verified	Not verified
Offline sync	Yes — reads audio the pendant stored while the phone was away (firmware 3.0.20+)	Not verified	Not verified	Not verified
Tags	Yes, on conversations and people, up to 20 each	Not verified	Not verified	Not verified
Raw audio export	No — playback only, while the server keeps the audio	Not verified	Not verified	Not verified
TV and media kept out of people	Yes — marks TV, video and call audio and leaves it out of names, facts and voice groups	Not verified	Not verified	Not verified
iOS	No	Yes	Not verified	Not verified

What you need

A phone on Android 12 or later, with Bluetooth on.
An Omi pendant on consumer firmware 3.0.x. Nytka checks the pendant's audio codec when it connects and tells you if the firmware needs an update. Offline sync needs firmware 3.0.20 or later; live capture works without it.
A Nytka server the phone can reach, with its address and an admin token: Nytka__AdminToken in the server's .env, or an admin token made in the app. A read token is refused ("The app needs an admin token."). The server README sets one up with Docker Compose. Keep the address and the token where you can paste them on the phone: the token is a long random string. The tabs, the sync and the server settings below need server 0.4.0 or later. An older server still takes audio, and a screen it lacks says "This server needs an update".

Install

Uninstall the official Omi app, or at least force-stop it: two apps cannot hold the pendant at once. To update the pendant's firmware later, install it again, update, and uninstall it.
On the phone, download nytka-<version>.apk from Releases and open it. Android asks once whether your browser may install apps; allow it, then tap Install. From a computer: adb install nytka-<version>.apk.

A newer APK installs over the old one the same way and keeps the queue and the settings.

Install with Obtainium

Obtainium installs and updates Android apps straight from their GitHub releases. Add this repository, nytka-app/android, as an app in Obtainium: it finds nytka-<version>.apk under Releases, the same file the manual install step above uses, and offers the next one as it is published.

First run (about 5 minutes)

Charge the pendant and keep it next to the phone. Open Nytka: the setup has four steps, shown as four segments at the top with the step in words.

Server. Enter the server's address (https://…) and an admin token, and tap Connect — or tap Scan QR code (offered first) and point the camera at the QR code the server's setup page shows, which fills both fields and runs the test for you. The address and token are set in a monospace font, and Show reveals the token. Scanning asks for the camera permission the first time; Nytka never uses it for anything else. The next step opens once the server answers and accepts the token. A read token stops here with "The app needs an admin token."
Permissions. Two rows say what each permission is for and show a check and Allowed once it is granted. Tap Allow. Android shows a prompt for Nearby devices, so Nytka can reach the pendant, and on Android 13 and later a second one for Notifications, for the recording notification and the alerts. There are no others, and Nytka never uses the phone's microphone or location.
Consent. Read the four points: people near you are recorded too, how to stop and delete, that recording people without their consent is illegal in some places and you answer for following the law (and the pendant also records while the phone is away), and where your recordings live. Tick I understand, which starts unticked, and tap Continue.
Pairing. Turn the pendant on, tap Pair and pick it in the list that opens. Set up later skips this; the Device screen has the same Pair pendant button.

Nytka now opens on Conversations. The capture status at the top reads Waiting for pendant, with a broken ring turning slowly, while the phone connects to the pendant, then Recording, with a green disc and a solid line under the top bar: the first run is done. Say something: within a minute the status card shows "Server: Last upload …", and the speech shows up under Conversations a few minutes after it is said. Pull the list down to refresh if a new conversation hasn't appeared.

A server on a VPN such as Tailscale or WireGuard, without HTTPS, needs the Private network (allow plain HTTP) switch on the server step and an address like http://<address on the VPN>:8080. It allows plain http://, which carries your audio unencrypted: use it only on a network you trust.

The tabs

Every switch shows a check in its knob when it is on, so the state does not rest on color alone. Addresses, tokens and signing secrets are set in a monospaced font.

Conversations. A status card, a search icon and your conversations by day; the day's name stays at the top while its cards scroll. A card shows the title, two lines of summary, and the time range and length. A thin moving line and Summarizing show while the server writes a summary; Summary failed with a red border shows when it could not. While the list loads it shows three grey placeholder cards. When your server cannot be reached the list says Offline and that your recordings stay on the phone; pull down to try again. Other failures say what failed and that your recordings are not affected, with Retry. Tasks, People, Memories and Ask use the same states. Open a conversation for its summary, its tasks and the transcript, with the speaker above each run of words when the transcription provider names speakers. Speakers differ by shape, not colour: you are a You tag, a named person is a name led by one of seven marks (disc, square, triangle, diamond, ring, plus, bar) given in order of first appearance, and a voice nobody has named is a saffron Voice 2 tag with Who is this?, which opens the naming dialog. The paragraph being heard gets a darker ground and a pause glyph. Its ⋮ menu has Rename (an empty title restores the generated one), Regenerate summary (not while the conversation is open) and Delete. When the server kept the audio (14 days by default, and none for imported conversations), a player is docked at the bottom: a plain track with a tick wherever the speaker changes, the position, length and who is talking, Follow (the transcript scrolls to the paragraph being heard; on by default), −10 and +10 seconds, play or pause, and the speed (1×, 1.25×, 1.5×, 2×). Tap the time beside a paragraph to play from there. Pauses between speech are not stored, so they are skipped. If the server cannot be reached the player says so in one line and stays. Playback stops when you leave the screen or the app. Touch and hold a paragraph to say whose it is: This is me, This is not me, or Clear (only where you marked it). Your mark beats the server's guess and shows as "Me · marked"; a line the server's voice match or the provider calls yours shows "Me · voice" or "Me · provider". Needs server 0.12.0 and an admin token. Speech kind. On a server that lists speech-kind, a line the server took for a TV, a video or a call shows a chip: Media (muted text) or Call. While the server only watches, a guess shows as an outlined Media? or Call?. Touch and hold a line for Person here, Media, Call or Clear kind; it shows at once and comes back, with a note, if the server refuses. This works on your own lines too. The ⋮ menu has Other voices are media and Other voices are people, which mark every line that is not yours and read the conversation again. Needs an admin token; without the feature there is no chip and no menu entry. When the server has a name suggestion for a voice in the conversation ("A voice may be Olena"), a saffron card above the transcript shows the line that carries it with Accept, Not Olena and Later, which only hides the card on this screen (Undo in the snackbar brings it back). Accept names the voice, so every line of it shows the name; Not Olena keeps the name from being suggested again for that voice. Nothing is applied until you tap. Tap a person's name above a run to open their page in the People tab. Needs server 0.14.0 and an admin token. When the same name is suggested for several voices (the server sends sameName, 0.21.0 and later), a third button Accept all 16 accepts every pending suggestion of that name at once, in the banner and in the inbox, and says "Added 16 voices to Olena", with "2 skipped" when some no longer applied. It sits beside Accept and never runs by itself. If the pending ones disagree, nothing changes and the app says to accept them one by one. A server without the route answers 404: the app says it needs an update and hides the button until the screen is opened again. On a server that lists roles the suggestion may be a role: someone known by what they do before you know their name. A role alone reads "A voice may be the repairman" with Add as repairman and Not that; a name and a role read "A voice may be Mykola, the repairman". The role is shown as the server sends it. Add as repairman creates a person called "Repairman" (or "Repairman 2") and tags them repairman; nothing is created until you tap. A server without roles shows the name only. Titles, summaries and tasks come from a language model on your server: without one, a row shows the start of the transcript instead.
Tasks. What your conversations left you to do, newest first, with the conversation's title and day. Tick a task to complete it, tap it to open its conversation, and use ⋮ to edit or delete it (delete asks first). Done tasks sit below, collapsed, in a dimmer color and never struck through. The model finds the tasks; there is no add button.
Memories. Lasting facts about you that the server took from your conversations, grouped by how long ago it learned them (Today, This week, This month, Earlier). Each is set in the reading type with its source, a link that opens the conversation (a memory you add yourself has none). + adds one, the pencil edits one, and the trash icon forgets one: it leaves the list at once and an Undo snackbar stays for a few seconds; when it goes away the server is asked to delete it, and a memory forgotten does not come back from later conversations. If you close the app before the snackbar is gone, the memory stays.
People. Everyone Nytka knows, most recently heard first and never-heard last, each with their first letter in a circle (a voice without a name gets a ? on a different colour), when they were last heard and how many facts the server holds about them. People a voice match asks about ("Is this Olena?") come first with Needs you. A server that does not send those two fields yet gets the list sorted by name with the number of lines each person said. Below it, Unnamed voices: tap one to name it. Where the server has no person page, tap a person for Rename, Merge into another person and Delete (with or without the stored voice model). Pull down to refresh. Needs server 0.7.0 or later; the list is not kept on the phone. Tap a person on a server with person pages (0.14.0 or later) to open their person page: a large monogram, when they were last heard, a line with a lock saying the voice print lives on your server (and from how many lines), a Rename button (Add name for a person known by a role), your note (up to 500 characters; only you write it, the model never reads or changes it; empty it and save to remove it), their facts (newest first, 50 at a time, Show more for older ones) each with why it is there ("They said", "Said about them", "Mentioned" or "Added by you") and the conversation it came from, Open tasks owed to them and their ten newest Conversations; tap a fact, task or conversation to open the conversation. Add a fact adds one; a fact has the same pencil and trash as a memory, with Undo (a deleted fact does not come back from later conversations). The page's ⋮ has Merge into…. At the very bottom, Forget this person asks first and names what goes: the voice print (also in the transcription service), the facts and the name. Adding a fact that exists, or taking a name someone has, says so under the field. A server without the page says it needs an update. Facts and notes need an admin token to change. Review inbox: when the server lists the review feature and something waits, an inbox icon with a count sits in the top bar. It lists name suggestions ("Is this Olena?" with the line that carries the name), voice matches (with the similarity) and lines Nytka is unsure are yours ("Did you say this?"), each with Yes and No. Nothing is applied until you tap; the row goes at once and comes back if the server refuses. Tap a row to open its conversation. Needs an admin token. A role reads "Is this the repairman?" or "Is this Mykola, the repairman?". Known by a role. A person who has a role and no name yet shows their role as the name with a second line "Name not known yet", in the list and on their page, whose ⋮ then has Add name in place of Rename. It opens the usual rename dialog, empty; saving sets the name. A name someone else has is refused ("taken") as for any rename. Later the server may suggest the name it hears for them: accepting that, when the name already belongs to another person, merges the role-only person into them, and the app says "Merged with {name}". Needs roles; an older server never sends the flag. A server that lists tag-suggestions adds proposed tags to the inbox ("Tag this conversation work?", "Tag Anna repairman?"), each with Yes and No; a person's tag opens that person. A tag that the item has no room for (20 tags) stays in the list with "This item has 20 tags." A server that lists speech-kind adds stretches of speech it is unsure about ("Was this a TV or a video?", or "Was this a call?" when it guesses a call), with the stretch's lines. Yes keeps the guess as your mark (a TV or video when it was unsure), No marks the stretch as a person. A stretch you or someone else already marked leaves the list without a notice. Who is this? cards sit above the list when your server groups voices (it needs people.voiceMatching on and keeps audio for at least a day). A card shows the conversation, up to three lines and a clip of at most 10 seconds, played from your server with your token and never stored on the phone. On a group, type a name or pick someone you know and tap Save, or tap Skip (hidden for 7 days) or Not a person. On a match ("Is this Olena?") tap Yes, Not them or Skip. Nothing is named until you tap. Needs an admin token. ⋮ → People settings edits the server's people settings (people.suggestNames, people.facts, people.voiceMatching, people.voiceThreshold) without developer mode; it needs an admin token. Under people.voiceMatching it says that recording people without their consent is illegal in some places, and the server's refusal (such as "Enroll your voice first.") shows under the switch in the server's English. The same screen has Delete all voice groups and voiceprints, which asks first and then removes every voice group, every voiceprint of a named person and every pending voice match from your server. It does not reach the voice models the transcription service keeps: those go with Delete and remove voice model on one person.
Ask. Type a question (up to 500 characters) and tap send for an answer drawn from your conversations and memories, with its sources numbered below it. Tap a source row, or its [n] in the answer, to open the conversation it came from; a memory with none opens nothing. "The server has no model set up, or is restarting.", "The model could not answer. Try again." and "The server took too long to answer. Try a shorter question." cover the server's own failures. Needs server 0.8.0 or later; an older server says it needs an update.

The bar at the bottom holds five tabs, each with its label: Conversations, Tasks, Ask, People and Memories (Памʼять in Ukrainian). People carries a saffron count of the name suggestions and other review items waiting for your answer; the review inbox icon on the People screen uses the same badge.

Capture status. Every main screen shows what the pendant's audio is doing at the top, in four states that differ in shape and words, not only in colour, so none can be mistaken for another:

State	Glyph	Line under the top bar
Recording	Filled green disc	Solid, with a stitch travelling along it
Muted	Two bars	Dashed
Waiting for pendant	Broken ring, turning once in 4 s	Dotted
Recording off	Ring with a slash	None

Green means only that audio is being captured now. With Remove animations on in Android's settings the ring and the stitch stand still; the shapes and words stay. TalkBack reads the state when it changes. The pendant's own light shows its link to the phone, not recording. Tap the capture status to open the Device screen: the pendant, its storage and settings, the server, the settings, Phone context, Your voice, Appearance and the version.

Appearance. Nytka uses its own linen-and-ink colours in light and dark, with Onest for the interface, Literata for reading and JetBrains Mono for addresses and tokens, all bundled in the app. Match wallpaper colours (off by default) lets buttons and surfaces follow your wallpaper; recording green, the saffron of questions that need you and the red of failures keep their colours. The font licences (SIL Open Font License 1.1) are under Device → About → Font licences.

Search opens from the magnifier above the conversation list. It looks through transcripts, titles, summaries and memories, in Ukrainian and English, while you type: from one letter or digit, after a short pause. All, Conversations and Memories narrow it; All also finds people by name (a server with the People tab), and a person opens their page in the People tab. A hit shows the title or the memory and a snippet with the matches in bold, and opens its conversation when it has one. Every word must match, exactly or as the start of a word. Endings of Ukrainian words need the server's optional dictionary: see the server README.

Tag chips. Tapping a tag on a conversation opens the Conversations list with only the conversations that have it, and tapping one on a person page does the same in People (the unnamed voices, which carry no tags, drop out). The filter lasts until you leave the app; a server without the tags feature keeps the full list.

Offline sync

While the phone is out of range, the pendant stores what it hears on its own card. When the phone comes back, Nytka reads that storage and uploads it with the times the audio was recorded, so it lands in the right conversations. It needs pendant firmware 3.0.20 or later and server 0.4.0 or later. On anything else the Pendant storage card on the Device screen says why, or is not there, and live capture is not affected.

A sync starts by itself whenever the pendant connects. The card shows it, and the recording notification adds "· syncing 42%". Stop holds it until the next connection or Sync now.

The card says	What it means
"Nothing stored", or "About 12 min stored. It syncs when connected."	That much was waiting at the last check: the line shows "Nothing stored" until the phone has read the pendant. Sync now starts a sync when the pendant is connected.
"Checking the pendant…"	Nytka is asking the pendant what it holds. Stop cancels.
"Syncing 42%, 38 KB/s, about 3 min left"	Under way. The rate depends on the Bluetooth link.
"About 2 h stored. Waiting for your answer."	The first sync found a large backlog. Tap Import or discard… and answer the question below.
"Waiting for uploads: the queue is 61% full."	The phone's queue is over 50% full. The sync waits and goes on below 40%, so it never drops audio.
"Paused: the link dropped. It resumes when the pendant reconnects."	The pendant went out of range. The next connection carries on from where it stopped.
"Paused: live audio is losing packets. It resumes shortly."	Live audio lost over 5% of its packets in one read window (Nytka needs at least 50 live packets to judge). The sync pauses for 30 seconds; the third pause in a row ends it until the next connection or Sync now.
"The pendant did not answer (status 9). Trying again in 30 s."	The pendant's storage failed to answer. Status 9 means its card is not ready yet, and -1 is a timeout. Nytka tries again after 30 seconds, then after 2 minutes, then every 10 minutes.
"The server did not answer. Trying again in 30 s."	The phone could not reach the server. Same tries.
"Stopped. Sync now continues where it left off."	You pressed Stop.
"Offline sync needs firmware 3.0.20 or later; this pendant has 3.0.19. Update it with the official Omi app."	Update the firmware, then uninstall the official app again.
"This pendant does not offer offline storage." (or "…did not report its firmware", "…its features")	The pendant cannot store audio for Nytka to read. Live capture is not affected.

Under the status line, the card shows Synced and Lost once they are above zero. They count audio since Nytka's recording service last started. Lost is audio the pendant no longer held when Nytka came to read it: its storage filled and overwrote the oldest audio, or the firmware freed it early after an interrupted transfer. "Pendant clock was off by 6 min, times corrected" appears when the pendant's clock had run wrong (after a crash, say) and Nytka fixed the times. The first sync of a pendant that holds more than 2 hours asks "Import the stored audio?": Import reads it all, which takes a while, and Discard… frees it on the pendant unread, after a confirmation.

Storage is freed after your server has the audio. Nytka tells the pendant to free stored audio only after the server accepted it; the exceptions are audio dropped for a mute and a backlog you discard. If you kill Nytka or switch Bluetooth off in the middle of a sync, the next sync carries on without repeating text. Firmware 3.0.21 also frees a little on its own during a transfer, so an interruption can cost a few seconds, which the card counts as Lost. Stored audio your server refuses (400, 409 or 413) stays parked on the phone, and the pendant keeps its copy.
Muted stretches are dropped. The pendant keeps no mute state, so it records while the phone is away whatever you chose. Nytka logs every mute change and, before it uploads anything, drops each stored frame within 2 seconds of a muted stretch. A mute stays open until you unmute, so it covers a whole absence, and muted audio never reaches the server.
The pendant's clock. At each connection to a pendant that supports offline sync, Nytka sets its clock from the phone. The pendant stamps stored audio with that clock.
What sync cannot recover. Audio that a stalled Bluetooth link dropped while the phone was connected (the pendant never stores it), audio the pendant overwrote or freed early (shown as Lost), a backlog you discard, and muted stretches, which are dropped on purpose.

The Device screen

The Device screen groups settings by where they live, in four cards: Pendant, On this phone, Your server and Developer (the last only while developer mode is on), then About.

Pendant shows the pendant's state, its battery as a meter (a neutral colour; below 10 % it turns red and says to charge the pendant), the firmware, the pendant storage and the pendant settings. Pendant details opens a screen with its name and address, where Forget asks first and the dialog names the pendant that is removed. Forgetting a pendant is only possible there.
How people nearby know (Device → Pendant, and the People near you step of the first run) says what the Omi pendant shows the people around you, on firmware 3.0.x. It never shows that the pendant is recording or muted: the light means connected (blue), not connected (red), clock not set (blinking red), charging (blinking green, alternating with blue or red), charged (steady green) or off. Under a red light the pendant still records to its own memory while the phone is away. The button: a double-tap mutes or unmutes with one long buzz, a single tap marks the moment with one short buzz, and holding it for 3 seconds turns the pendant off. It is read from the pendant's firmware (see the page's last line), and the colours are written in words, not drawn.
About → About recording shows the four points of the first run's People near you step again, to read; there is nothing to tick.
Server addresses, tokens and webhook URLs are set in a monospace font. A token shows as dots and its last four characters; Revoke is on the token's own screen (Developer mode → Server → Access tokens → tap a token) and asks first, naming the token. Deleting a webhook asks first and names its address.

Pendant settings

While the pendant is connected, the Device screen has a Pendant settings card with two sliders, for pendants that report the feature (firmware 3.0.21 does). LED brightness runs from 0 to 100 % and dims every status light; at 0 the link and charging lights stay dark too, and the card says so. Microphone gain has levels 0 to 8: 0 is mute, then -20, -10, 0, +6, +10, +20 (the default, level 6), +30 and +40 dB, as the firmware maps them. Level 0 records silence, and the card warns about it. The value is written when you let go of the slider, at most once every 2 seconds for each slider, because the pendant saves it to its flash memory and keeps it after a restart. Nytka reads the current values on every connection and stores nothing on the server. A failed write shows a message and puts the slider back.

Pendant firmware notice

The Pendant card on the Device screen shows the firmware the connected pendant reports. Nytka also tells you when Omi has published a newer one: the card then says "Firmware 3.0.21 is available (this pendant has 3.0.20)" with a How to update link to Omi's own instructions. Nytka only tells you. It never downloads or installs firmware; the official Omi app does that (install it, update, uninstall it again).

When it asks. At most once every 24 hours, and only while the Device screen shows a connected pendant, Nytka asks GitHub for the newest release of BasedHardware/omi in the pendant's firmware stream (three or four requests, no token). Offline, it tries again an hour later at the earliest.
Which pendants. Only the consumer pendant, which reports the model "Omi CV 1" and a 3.x firmware, and whose releases are tagged Omi_CV1_v<version>. Omi's other hardware (DevKit 1 and 2, EVT, Glass) has its own tags and is not checked; a pendant that reports anything else gets no notice, because a wrong guess would send you to the wrong firmware. Pre-releases are ignored.
What GitHub learns. Your IP address, like any website you open. No account, token, cookie, pendant data, audio, or server address leaves the phone for this. Switch it off with Check for new pendant firmware on the Device screen (on by default); then Nytka never contacts GitHub.

Your voice

With server 0.12 or later and its speaker model (/api/v1/info lists voice), the Device screen has a Your voice card. It teaches the server your voice once, so it marks the lines you said as yours whatever the transcription service says. It needs an admin token.

Choose the languages you speak (Ukrainian, Russian, English) and tap Enroll. The screen shows three short sentences per language: read them aloud with the pendant on, somewhere quiet, for at least 30 seconds in all (with one language, read them twice). One sentence shows at a time in large type; Next sentence moves on. A timer and a bar count the seconds against the server's 120-second limit, with a mark at 30 seconds; recording stops by itself at the limit. Send stays off until 30 seconds, and the screen says how many more to read. Send gives the reading to the server, which answers with how much speech it used. It refuses a reading with less than 20 seconds of speech, one too short to make 3 samples, or one whose samples do not sound like one voice (another speaker or noise), and the screen says which and what to do.

The reading comes from the pendant, never the phone's microphone, because the server matches pendant audio later. While you read, the pendant's live audio goes only to the enrollment and is not queued for upload, so the reading never becomes a conversation; capture goes back to normal as soon as you send, cancel, reach the limit or leave the screen. The audio of the reading is not kept on the phone, and the server keeps only the voiceprint.

Once enrolled, the card shows when, from how many samples, and how many of your segments the voiceprint learned from since. Re-enroll starts over, Add more blends a new reading in (for example another language), Forget what it learned goes back to the enrolled voiceprint, and Forget my voice (asked first) deletes the voiceprint, every segment fingerprint and the labels made from them; your own marks stay. See the server README.

Phone context

When the server lists context-ranges in /api/v1/info, the Device screen has a Phone context switch, on by default: "Tell your server when this phone plays sound through its speaker or is in a call. Only times are sent, never what played." While capture runs, the app notes two kinds of stretch: media, when the phone plays music or video through its built-in speaker at a volume above 0 (it ends 2 seconds after the sound stops, or at once when the sound moves to headphones), and call, when the phone is in a call or a voice or video chat, with the route it held longest: speaker, earpiece, headset, Bluetooth or other. The server uses them to tell a conversation in the room from a television, a video or a call.

What the app sends is the kind, the route and two times for each stretch, in batches of up to 500 once a minute. Never an app, a title, a contact or a number, and nothing from the microphone: the app listens to Android's playback and call-state callbacks, which need no permission, and does not poll. A stretch under 3 seconds is dropped. Stretches wait on the phone until the server has them and are deleted after 7 days if they never go out. With the switch off, or on a server without context-ranges, nothing is recorded or sent.

Meeting briefs

With a server that has a calendar feed (Nytka__Calendar__IcsUrl, see the server README), the server writes a short brief before a meeting with someone you named: who they are to you, what to remember, what is still open. Switch on Meeting briefs in the Device screen (off by default) and Nytka posts it as a notification. On Android 13 and later it asks to show notifications first; if you refuse twice, the switch offers Open settings. Under the switch, "Your server has no calendar feed" means the feed is not set, and "This server needs an update" means the server predates it.

When. Nytka asks your server every 15 minutes, with a network, and posts each brief once, for meetings not over. The server makes a brief 30 minutes before the meeting by default (Nytka__Calendar__BriefMinutes), so it arrives 15 to 30 minutes before; with 5 it may arrive after the start.
The lock screen shows only "A meeting brief is ready": no title, name or fact. Unlock to read it. The notification channel is called "Meeting briefs".
A failed check (no network, an older server) posts nothing and waits for the next one.
Nothing about a brief goes into the app's logs or the debug report.

Everyday use

Mute with a double tap on the pendant: one long buzz means muted, two short buzzes mean recording again. Mute in the notification, on the status card and on the home-screen widget does the same. The notification's title is the capture state, and its status-bar icon is the same glyph as in the app, so the shape tells the state even in the status bar; Open opens the app. While muted, Nytka hears nothing and the pendant discards its audio, and Nytka stays muted when the pendant reconnects. What the pendant records while the phone is away is handled by Offline sync.
A home-screen widget shows the capture state with its glyph (Recording, Muted, Waiting for pendant or Recording off; only Recording is green), pendant battery and the upload queue length. It follows your launcher's light or dark theme and your wallpaper's colours, and keeps the recording green; tap it to open the app, or tap Mute/Unmute to toggle recording without opening it.
Bookmark with a single tap on the pendant: one short buzz after about a second means it is kept. A double tap (or two quick single taps) mutes instead and keeps no bookmark. Bookmark in the notification does the same from the phone. Bookmarks upload like audio, also while muted. A conversation shows a star beside the nearest paragraph, where you can tap to add a note. The tap works only while the pendant is connected.
Offline, audio waits on the phone and uploads once the server answers again, with the times it was said. The queue holds up to 1 GiB, about 60 hours. Past 80% Nytka warns you; when it is full, the oldest live audio goes first.
Alerts come when the pendant has been away for 5 minutes, the server has been unreachable for 15 minutes, the pendant battery reaches 20%, or the queue is 80% full.
Consent chime (Device screen, off by default): while the pendant records, the phone plays a short chime every 5 to 60 minutes (15 by default), at once when recording starts and not on each reconnect, so people nearby know. It is silent while muted, disconnected or during voice enrollment. It uses no microphone and no permission, and follows the notification volume: on silent, or when Do Not Disturb blocks notifications, you will not hear it.
Language of summaries (Device screen): English, Ukrainian or Follow the conversation, editing the server's llm.outputLanguage. Shown locked, "Set by the server's environment", when the server's environment sets it.
Delete a conversation from its ⋮ menu: its transcript, audio, tasks and memories go from your server.
After a reboot, recording starts once the phone has been unlocked and the pendant is in range; Nytka need not be opened.
The Device tab shows the pendant and the server, and holds the settings: server address, token and the private-network switch.
Diagnostics (Device → Developer mode) record link quality, queue and upload health every 10 seconds while Nytka records, and keep the last 7 days. A sample also holds the pendant's signal strength (rssi, dBm), the status of the last link loss (lastDisconnectStatus; 147 is a link that went silent), and the phone's own battery level and charging state. The app's log names that same status (and the other GATT and link statuses Android reports) next to the number, for example onConnectionStateChange status=147 (LINK_WENT_SILENT) newState=0. Export diagnostics shares them as a CSV file through Android's share sheet. Send diagnostics to my server is off by default; on, it uploads the samples to your server every minute (needs server 0.2.0 or later). The app's own log events (tags such as OmiPendant, CaptureController) are kept and uploaded with them while Nytka records; they hold no audio, transcripts, server address or token. Read them on the server with select at, payload->>'level', payload->>'tag', payload->>'message' from diagnostics where payload->>'kind' = 'log' order by at. The export and the sample count leave them out.

When something is off

What you see	What to do
"The server refused the token." on Test connection	The token differs from `Nytka__AdminToken` in the server's `.env`. Paste it again, whole.
Another error on Test connection, such as "Failed to connect" or "Unable to resolve host"	The phone cannot reach that address; an unreachable one takes 15 seconds to fail. Open the address plus `/healthz` in the phone's browser: `{"status":"healthy"}` means the server is up, so recheck the address and the private-network switch in Nytka. Anything else is the network: the VPN is off, a firewall is in the way, or `NYTKA_BIND` on the server points elsewhere.
"Without nearby devices Nytka cannot reach the pendant."	Android stops asking after two refusals. Open Nytka's app info (long-press its icon), allow Nearby devices under Permissions, come back and tap Allow again. Nytka resumes first run at the step you reached.
"Pendant not supported", or a message about the audio codec	Update the pendant's firmware with the official Omi app, then uninstall that app again.
The chip stays on "Waiting", or the pairing list is empty	The official Omi app (or another phone) still holds the pendant, Bluetooth is off, or the pendant is out of range or flat.
"Paused: The server refused the token."	The server's token changed. Enter the new one under Device → Settings; nothing queued is lost.
"The app needs an admin token."	The token is a `read` token. Enter the server's `Nytka__AdminToken`, or an `admin` token made under Developer mode → Access tokens.
"This server needs an update"	The server has no endpoint for that screen. Update it to 0.4.0 or later; capture keeps working meanwhile.
"Unreachable since …"	The phone cannot reach the server. Audio waits in the queue and uploads by itself.
A conversation shows Summary failed	The server's language model failed: a wrong key, an endpoint that is down, a model name it does not know. The server's admin finds the reason at `GET /api/v1/status`, under `ai.lastError`. After the fix, use ⋮ → Regenerate summary.
No conversation gets a title or a summary	The server has no language model set up. Set its base URL and model in the server's `.env`, or under Developer mode → Server settings; see the server README.
Regenerate summary says "No summary can be made now: the conversation is still open, or the server has no language model set up."	Wait until the conversation ends (by default after two minutes without speech), or set up the model.
The Device tab has no Pendant storage card	The card shows when a pendant is paired and the server lists `offline-sync` (server 0.4.0 or later). Its other states are in Offline sync.
Conversations appear but stay empty after you pull down to refresh	The server could not transcribe the speech. Turn on developer mode (below) and tap Check status: it shows the server's last error, usually the transcription URL, key or model in the server's `.env`.
The recording notification disappears on its own	Some phones stop background apps. Set Nytka's battery use to Unrestricted in its system settings; dontkillmyapp.com has the steps for each maker.

What stays on the phone

Audio waiting to upload, in the app's private storage, until the server has it; the settings; the token, encrypted with a key that never leaves the Android Keystore; how far the pendant's storage has been read; the time and result of the last firmware check; the ids (not the text) of the meeting briefs already posted, at most 50; the phone-context stretches not yet sent (kind, route, two times); and a log of when you muted and unmuted (times and on or off, nothing else). Android backups and phone-to-phone transfers skip all of it, so a new phone starts with first run. The pendant keeps the audio it stored until your server has it. The app talks to your server and nothing else, apart from the once-a-day firmware notice (GitHub, switchable): no analytics, no crash reporting, no Omi cloud.

Diagnostics samples hold counters and short status words: no audio, no words, no server address, no token. They stay on the phone for 7 days and leave it only when you export them or switch on Send diagnostics to my server, which sends them to your own server and nowhere else.

Developer mode

Tap the version under Device → About seven times. Developer mode shows the Bluetooth packet rate and loss, the queue, the last upload and the server's status. Under Pendant storage, when the pendant can store and the server supports it, it shows the pendant's ring (read, write, capacity and dropped), the last transfer status, the measured rate, the loss of live audio during a sync, the frames dropped for a mute, the records dropped for a bad stamp, the pendant's clock skew and the number of clock segments. It has a fake pendant that replays a bundled text-to-speech recording, a Request high-priority connection switch (off by default; capture restarts, and each connection then asks Android for a 11.25 to 15 ms interval, to compare how often the link drops with and without it), a button that saves the next 60 seconds of frames as a test fixture, and Copy debug report for bug reports, which never includes the token. A saved fixture holds the voices of whoever spoke: keep it private. Its Server section holds three screens.

Server settings edits the settings your server keeps in its database, grouped by prefix: stt, conversations, audio, llm, memories and search. A setting that the server's .env sets is locked: "Set by the server's environment". API keys show only Set or Not set, because they live in the server's environment and the app can neither read nor change them. Save sends only what you changed, all or nothing, and an emptied field restores the default. The server uses a new value from its next job, with no restart. The people keys are also reachable from People → ⋮ → People settings, and llm.outputLanguage from the Device screen's Language of summaries. What each setting does is in the server README.

Access tokens lists the server's named tokens: name, scope, when it was made and last used, and whether it is revoked. Create token asks for a name and a scope, Read (conversations, tasks, memories and search over the API, and MCP for agents) or Admin (everything, this app included). The token shows once, with Copy; Nytka cannot show it again. Revoke ends a token at once. Give MCP clients a read token.

Webhooks lists the server's webhooks, each with an on and off switch and the status of its last delivery. + adds one: a URL, the events (a conversation is summarized, a task is created, a task is completed, a memory is created, or all of them) and a description. The signing secret shows once, with Copy. A webhook's own screen has Send test, its last 30 deliveries and Delete. How a receiver checks the signature is in the server README.

Build from source

JDK 21 and the Android SDK with platform 37.0; see CONTRIBUTING.md.

./gradlew assembleOssDebug
adb install app/build/outputs/apk/oss/debug/app-oss-debug.apk

oss is the flavor Releases publishes; play is kept for a later store listing and only has to compile.

The UI is in English and in Ukrainian: app/src/main/res/values/strings.xml holds the English, values-uk/strings.xml the Ukrainian. A new language is a new values-<locale> folder whose <string> and <plurals> names match the English ones. Add the language to app/src/main/res/xml/locales_config.xml as well, or Android 13's per-app language setting will not list it.

License

Apache-2.0, see LICENSE and NOTICE.

Nytka is an independent project, not affiliated with or endorsed by Based Hardware; "Omi" is a trademark of its owner. Nytka records the people around the wearer: you are responsible for following the recording and privacy laws where you use it.

Name		Name	Last commit message	Last commit date
.githooks		.githooks
.github		.github
app		app
config		config
core		core
docs		docs
fastlane/metadata/android		fastlane/metadata/android
gradle		gradle
pendant		pendant
.editorconfig		.editorconfig
.gitignore		.gitignore
.private-terms.example		.private-terms.example
.release-please-manifest.json		.release-please-manifest.json
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md
SECURITY.md		SECURITY.md
build.gradle.kts		build.gradle.kts
gradle.properties		gradle.properties
gradlew		gradlew
gradlew.bat		gradlew.bat
release-please-config.json		release-please-config.json
settings.gradle.kts		settings.gradle.kts

Navigation Menu

Repository navigation

About

Repository files navigation

Nytka for Android

How Nytka compares

What you need

Install

Install with Obtainium

First run (about 5 minutes)

The tabs

Tags

Offline sync

The Device screen

Pendant settings

Pendant firmware notice

Your voice

Phone context

Meeting briefs

Everyday use

When something is off

What stays on the phone

Developer mode

Build from source

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Nytka for Android

How Nytka compares

What you need

Install

Install with Obtainium

First run (about 5 minutes)

The tabs

Tags

Offline sync

The Device screen

Pendant settings

Pendant firmware notice

Your voice

Phone context

Meeting briefs

Everyday use

When something is off

What stays on the phone

Developer mode

Build from source

License

About

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages