Open source Grok Bot alternative, Bring your own agent and subscription: message Claude Code, Codex, Gemini CLI, Cursor, Copilot and 40+ AI agents as bots on your own computer, from iPhone, Mac, Windows, Linux or terminal. Agent-agnostic and free. Also a Muse and Dots alternative. Built in Rust.
Usage analytics
Product analytics in PostHog (US Cloud, project "Codync App", id 649427), on by default on
every device; Share usage analytics in Settings turns it off. It exists to spot bugs sooner
(failed or stopped turns per agent and platform) and to serve users better. The public promise
lives in web/app/privacy/page.tsx: keep it true. The info button beside the switch (terminal
UI: About usage analytics…) shows the same promise in the app, in one text per client
(desktop views/settings/AnalyticsSheet.tsx, CodyncUI/AnalyticsSheet.swift, Android
AnalyticsSheet.kt, host/src/tui/view/analytics.rs), and the Windows installer sums it up
(apps/desktop/resources/installer.nsh): change all five with the policy.
Rules
- Only feature use, never content. No message text, prompts, code, file names, paths,
bot names, credentials or agent output. Properties are counts, booleans and public
catalog names (agent, model, connector catalog name; a custom connector is just
custom). Errors are categories and code locations, never an error's message (see Error tracking). - Setting per device. The computer's owner decides for that computer (host kv
analytics_enabled, on unless0); each iPhone decides for itself (UserDefaultsanalyticsEnabled, on unlessfalse; each Android phone in itscodync-analyticspreferences). Turning it off drops anything still queued. A device that said no to the earlier opt-in question stays off.codync-host analytics on|offsets the computer's choice from the command line: through the running host, or saved for its next start. - The Windows installer asks. An install run by hand (not an update, not a silent
/Sone) shows what is shared, links the privacy policy and asks before anything is sent (resources/installer.nsh): No runscodync-host analytics off; Yes leaves the setting as it is (on for a new computer), so the first launch can still reporthost_setup_failed. Known gap: a reinstall answered No over an app that last saw analytics on can still report onehost_setup_failedif the host never starts (the app'shost-prefs.jsoncopy is stale). Open source code signing (SignPath) requires this for a signed installer; the Mac and Linux packages have no install step that could ask. - Identity. Every device gets a person profile. Signed in to a Codync account: the Clerk
user id. Otherwise a random install id (the host id on a computer, a UUID on the iPhone).
A device's first signed-in event sends
$identifywith$anon_distinct_idso earlier events join the account. Only once per device: PostHog refuses to merge an id into a second account (cannot_merge_already_identified), so a later account on that device starts its own profile. - Sessions. Every event carries a
$session_id(a UUIDv7) that changes after 30 minutes without events or 24 hours in all, as in PostHog's own SDKs, and on each launch of an app or start of the host (hostanalytics/session.rs,AnalyticsSessionon the iPhone and Android), so an error's page shows the events that led up to it. Nothing is recorded for it beyond the events themselves. A crash reported on the next launch keeps the session it happened in, with the time it was reported. - Location, never the IP. The project's GeoIP transformation turns the sender's IP into an
approximate location (
$geoip_country_name,$geoip_city_name, …), then the project's Discard client IP data drops the IP before storage. envismainordev; filter dev out in PostHog. Debug host builds (cargo test,cargo run) send nothing: every e2e test starts a fresh host that would count as a new user.
Adding a feature
Every new feature or module ships with its PostHog events in the same change, so a later bug can be traced from the events around it:
- Use: an event when it is used. A host method gets it from
api::dispatch(analytics::observe_callinanalytics/observe.rs); an app-only action from its client (trackon the desktop and terminal UI, added toanalytics::Reported; the analytics client on the iPhone and Android, both alike). - Outcome: how it ended, as
turn_finisheddoes:outcome(done,failed, …) and, for a failure, its category inerror. Never the error's message. - Errors: failures that should never happen go to Error tracking below (
reportErroron the host, desktop and terminal UI); crashes are already caught. - Display preferences send nothing. A device's own look (Settings › Appearance: theme, text size, usage icons, compact sidebar, agent logos; thinking details) has no outcome or failure to trace, so switching one is not an event.
- The Rules above apply: feature use only, never content. List each new event and property in
analytics events, and check they arrive in PostHog
(
env = dev) during the change's e2e test.
Where events come from
- Host (
host/src/analytics.rs,analytics/):api::dispatchmaps successful API calls to events (analytics::observe_call), so every client is covered once; a resent message (sameclientNonce) is not counted again. Bot actors reportturn_finishedandagent_fell_back(a turn moved on to the bot's next agent: the agent and why, see agent fallbacks), the phone channel reports pairing and daily phone connections, and the host reports its own start. Each event carries its owntimestampanduuid(a resent batch counts once) and the computer's key (computer, a hash of its id); batches go to/batch/every 30 s and once more when the host stops. - Desktop and terminal UI report app-level events (opening, onboarding, sign-in, screens,
the star ask, links) through the loopback-only
trackmethod, which accepts only the catalog's events and properties. The desktop app queues them until the host answers; when it can't get a working host at all, its main process reportshost_setup_failedstraight to PostHog, only while the host's last answer said sharing is on (or on a first install, before the host stored any setting). It turns sharing on or off withsetAnalytics;hellocarriesanalytics: true | false. - iPhone (
CodyncKitanalytics client, no SDK) and Android (Analytics.kt, its port) send their own app-level events straight to PostHog. Widgets and the notification extension never send. - codync.dev counts page views, time on page, scrolling, load speed and clicks on download
and outbound links without cookies, through its own
/epath (website).
Every event, its sender and its properties: analytics events.
Error tracking
Crashes and uncaught errors go to PostHog Error Tracking as $exception events with the same
switch, identity and common properties as everything else, plus process. No SDK: each client
builds PostHog frames itself (oldest first, the failing one last): generic custom frames, or
apple frames on the iPhone.
- Host (
host/src/analytics/errors.rs): its own panics (a panic hook queues them for the analytics task), andreportError, loopback only, for the computer's apps. Processhost. - Terminal UI: its panics, through
reportErrorfrom the panic hook. Processtui. - Desktop (
apps/desktop/src/main/error-reports.ts,shared/error-report.ts): uncaught errors and rejections in the main process and every window, errors a view threw while rendering (mechanismreact: the chat and pairing windows show Something went wrong with Try again instead of going blank), and windows or helpers that crashed (RendererGone,GPUGone, … with Electron's reason), throughreportError. A window whose page crashed loads it again (renderer-recovery.ts), at most twice in a row before it has stayed up a minute, and says what came of it (window_recovery). Not reported: error events that carry no error (the ResizeObserver loop notice), and helpers that exited normally or were ended from outside (clean-exit,killed: logout, shutdown, an update), and a GPU process that couldn't launch (the app goes on with software rendering). Processesdesktop-main,desktop-renderer. The renderer is built unminified so function names stay readable. - iPhone (
CodyncKitCrashReports): MetricKit's crash diagnostics, delivered on the next launch, straight to PostHog with the version that crashed and the session it ended. Frames are code addresses with the binaries they ran in ($debug_images: each binary's UUID and load address). PostHog symbolicates them with the release's dSYMs, which the iOS dSYMs workflow uploads from its Xcode Cloud archive after Submit iOS (tools/upload-dsyms.py, secretPOSTHOG_CLI_API_KEY: a personal API key with error tracking write on this project). A frame it can't resolve, such as the system's, showsbinary + offset. Processios. Reading one, older versions and fixes: iPhone crash tracing. - Android (
CrashReports.kt): an uncaught exception is written down as it happens and sent on the next launch: its type and frames (class, method, file name, line), never the message. Processandroid.
What leaves the device: the error's type (Panic, TypeError, EXC_BAD_ACCESS), how it was
caught, function names, and file names and line numbers inside Codync's code (on the iPhone,
code addresses and binary UUIDs instead); an absolute path keeps only its file name. Never the
error's message: only a panic message written as a literal in the code, an Electron crash reason
or a signal name. Release host builds are stripped, so a
host panic is its location only.
Alerts email the owner, from insights on the Codync users dashboard (production only): Any
crash or error in a day ($exception > 0, checked daily) and Failed agent turns over 15 in an
hour (turn_finished with outcome = failed, checked hourly; a normal hour has a few). A turn
that only needs the user to sign in or install the agent reports needs_setup and isn't counted;
older hosts still report it as failed with error sign_in/not_installed. Error
tracking's own issue alerts (new, reopened, spiking) need a Slack or Discord destination.
Errors traced to a root cause are written up in investigations; ways to record more detail are proposed in error tracking detail.
Metrics
Shared filters: env = 'main', internal accounts excluded, UTC days. A user is a PostHog
person: the account when signed in, otherwise one install, so an anonymous iPhone and an
anonymous computer are two users until both sign in. Messages are recorded by the computer,
so an anonymous iPhone's chats count for its computer. To count people rather than devices,
join an iPhone's computer_paired to the computer's events on computer in SQL.
| Metric | Definition |
|---|---|
| Active user (DAU/WAU/MAU) | Sent ≥1 message_sent in the period. Any event at all counts as present, not active (app_opened on the iPhone fires on every return to the foreground). |
| Present (frequency) | Days in the period with app_opened, window_opened or phone_connected (stickiness). A computer whose only events are turn_finished/routine_run ran unattended. |
| New user | First event of any kind in the period. A new computer also sends host_started with first_run. |
| Onboarded | Desktop: onboarding_completed. Phones: computer_paired (or onboarding_completed with via = skipped, which is not onboarded). Funnels: analytics events. |
| Star rate | star_prompt_answered with starred ÷ all star_prompt_answered, per ask. |
| Activated | Sent ≥1 message from a phone (message_sent, client = remote) within 7 days of first seen. In the first week of data, 4 of 5 such users came back after 24 h and none of the 6 computer-only ones did. |
| Successful turn rate | turn_finished with outcome = done ÷ all turn_finished with source = user, per agent and per $os. |
| Frustration rate | (turn_stopped + session_reset) ÷ message_sent, per $os. |
| Retained (D1, D7) | Active again on day 1 / any of days 7–13 after their first active day (PostHog retention, message_sent both ways). |
| Power user | Active on ≥4 of the last 7 days (stickiness). |
| Feature adoption | Share of WAU with ≥1 of: voice_call_finished, remote_screen_opened, connector_installed, routine_saved, bot_created with kind = group. |
| Up to date | Share of computers active in the period whose last app_version is the latest release. |
Turning it off
No client asks: sharing starts with the first launch. Every client has the same switch,
Share usage analytics (iPhone, Android and desktop Settings, terminal UI ^k). The App
Store privacy label declares Product Interaction and Coarse Location, linked to the user, not used
for tracking (the same two in apps/ios/Resources/PrivacyInfo.xcprivacy).