Telemetry: what atrium collects and how to opt out
What atrium sends: core lifecycle and setup events by default, plus opt-in analytics and crash reports. The exact payloads, what is never collected, and how to turn it all off.
atrium has two tiers of telemetry: a small set of core events that tell us the app launches, stays in use, and gets set up successfully, and an opt-in analytics stream that helps us understand how the product is used. Core events are on by default. Analytics and crash reports stay off until you choose Share usage & crash reports during first-run calibration or in Settings → Privacy & data. Choosing Send nothing turns both off.
Config
Three flags control telemetry, stored in the active data directory's config.json under telemetry. This example leaves analytics off:
{
"enabled": true,
"analyticsOptIn": false,
"coreTelemetryDisclosed": false
}
enabled— master kill-switch. Whenfalse(the Send nothing choice), atrium sends nothing, regardless of the other two flags.analyticsOptIn— user explicitly opted into product analytics and crash reports.coreTelemetryDisclosed— user has made a telemetry choice in calibration or Settings. Analytics only flow once this istrue.
internal, marks a test machine: every event it sends carries internal: true, so dashboards can leave it out. It is not a consent choice and changes nothing about what is sent. Set it with atrium config set telemetry.internal true or the ATRIUM_TELEMETRY_INTERNAL=1 environment variable. Development builds, and installs on a machine that also has a development build (a .atrium-dev data directory), are marked automatically.
Installs that have never made a choice see a one-time prompt in the corner of the window asking for one. Closing it leaves the defaults as they are, and it does not come back.
Tier 1 — Core
On by default whenever enabled is true. Includes:
- App lifecycle — launch, version, OS, architecture, locale, and the release channel (stable or beta).
app_activeheartbeat — fires after a minute of focused use and then every 10 minutes of it, with the same payload as launch plus the seconds of focused time since the last heartbeat (at most 15 minutes' worth) and the peak number of agents running at once. This is the signal we use to measure daily-active installs without relying on launch-only events, since users keep atrium open across days. Nothing accumulates while the window is unfocused, backgrounded, or the computer is asleep.- Daily rollups — once per day, from each machine's background service (desktop or a headless remote box):
daemon_active_daily: version, OS, architecture, whether a window was open that day, and bucketed uptime, attached clients and agents started.
- hq_state_daily: how many directors, staff, enabled routines, dashboards and docks exist; how many connectors, by catalog id (any connector not in atrium's catalog counts as other); and that day's delegation count, routine fires and peak concurrent workers. It also counts tracked asks (an agent's questions to you): how many wait on you, how many you put off, how many were seen with no update since, how many agents are waiting, and the oldest waiting ask's age (bucketed), with how long atrium takes to look for asks in a finished turn and to read them back (in microseconds) and the size of an ask update (in bytes). Counts only, never names.
- crash_count_daily, only on a day with crashes: how many app panics, webview errors and native crash reports, by kind. The crashes themselves are not sent; crash reports stay under the analytics opt-in below.
- Setup progress — opening adapter setup, and whether each adapter install started, succeeded, or failed (with a fixed failure category and duration); whether the restore chooser was shown and how it was resolved.
- Activation milestones — one-shot events the first time an install reaches a step like "first agent launched" or "first room created", with the time since install. One event per install, used to understand which features users actually reach.
- Consent prompt — that the one-time prompt was shown, and whether it was answered with Share usage & crash reports or Core only, or closed. Choosing Send nothing sends nothing, including this.
- Updates — an update being offered, installed (the old and new version), or failing (a fixed stage and failure category, never the error text).
Tier 2 — Analytics & crash reports
Only collected when you chose Share usage & crash reports: enabled: true, analyticsOptIn: true, coreTelemetryDisclosed: true.
- Feature usage — which pane types are opened, which commands run, which keybindings fire. Counts only — no argument contents. The Tier 2 schema covers vault commits and restores, activity sidebar transitions, dictation start/stop, layout operations (split, snap, restore, popout, expose), code-review dispatch, launcher interactions, browser open and crash signals, room switches and pinning, file-tree multiselect actions, and task and run lifecycle transitions.
- Adapter launches — which adapters are started and their versions, and agent runs: which surface started them (launcher, chat, HQ, task, restore, terminal) and the model family (Opus, Sonnet, GPT, Grok, Gemini, and so on, or
other). - HQ — a director created or a staff member hired (with the harness, and the library role or
custom), a delegation started (from a director, staff member, person or routine, to staff or a worker) and how it ended (done, failed, cancelled, or needing input, with a bucketed duration), routine fires (scheduled or manual, and the outcome), connectors connected or disconnected (catalog id orother), and which HQ surfaces are opened. - Tracked asks — an agent's question to you being created (its kind, where it came from, the adapter, whether it is a director's, and whether it has a default, a deadline, or acts outside your machine), answered (with an option, your own words, a note, a tick, a form, or the default going ahead; on which device and from which surface; and the time to answer, bucketed), put off for later and until when, brought back, declined (whether you gave a reason), reopened, marked done by the agent (from which state, and how long after it saw your answer, bucketed), closed (why, and by whom), or seen with no update. Never the question, your answer, or any id.
- Remote use — a remote location added (box, SSH or tailnet), a phone or another computer attaching (at most once per device a day), and Walk Mode sessions (device type, bucketed length and number of panes visited).
- X post notices — once a day, from each machine: how many times the opt-in card was shown, turned on or closed, and how many post notices were shown, opened on X or turned off (and whether a switch came from Settings). Counts only, never which post. To know whether to offer them, atrium checks once at launch whether getatrium.dev's public file of atrium's recent posts is available, and every few hours while it isn't; with the notices on, it reads that file about once a minute while a window is visible. These requests carry no cookies and no account or install id, and the same file goes to everyone (like any web request, the server sees the address it comes from).
- Accounts — a Claude account switch (manual, or automatic after a usage limit, with a bucketed count of configured accounts) and usage-limit hits (the limit's kind, and whether atrium moved the run to another account).
- Performance — mosaic render times, pane switch latency, terminal throughput (aggregate, not per-session).
- Crashes & errors — Rust panics and uncaught webview exceptions, with stack traces and release/commit attribution but without user content. Captured via Sentry and gated on the same opt-in as analytics, so users who want lifecycle data without crash collection can keep
analyticsOptIn: false.
chat/:id/send), never with the pane, chat or agent ids in it.
Each machine is its own anonymous install: atrium does not link one person's computers, phones or remote boxes to each other.
What we never collect
- Terminal scrollback.
- Editor or markdown buffer contents.
- Browser pane URLs, cookies, or form inputs.
- Task card contents, comments, or descriptions.
- Files from the project directory.
- AI agent prompts or responses.
- Keybindings or search query text.
How to opt out
- Master switch: choose Send nothing in Settings → Privacy & data, or set
telemetry.enabledtofalseinconfig.json. atrium sends nothing — neither PostHog nor Sentry. - Analytics + crash reports off: choose Core only in Settings → Privacy & data, or set
analyticsOptIntofalse. Tier 1 events still flow, but Sentry is not initialized at all and PostHog feature-usage events are suppressed.
Local diagnostics and support
Settings → Diagnostics records local performance evidence; it does not provide an audit of every outbound telemetry payload. The recorder is off by default. You can choose a 30-minute, 2-hour, or 8-hour recording window, or stop it yourself; every recording also stops at 500 MB. atrium retains the newest five recorder traces within a combined 2 GB budget.
Creating a performance bundle does not upload it. Manually captured bundles and profiles are outside the recorder's automatic retention limit. See Diagnostics for the files included, and review logs and bundles before attaching them to a public issue.
Licensing note
Licensing and paid-tier gating are deferred past the 1.0 release. Nothing in the telemetry path is tied to a license check today.
Contact
If you see behavior that surprises you — a request that does not look like it should be happening, a payload that includes more than expected — open an issue and we will investigate.