Trust and privacy
What leaves your computer
Every feature of Growth OX, what it sends out, to whom and when, and what happens to it in Only this computer mode. Generated from the audit file the app is checked against.
How to read this page#
This list is generated from the same audit file that the app's privacy checks use (audited 2026-10-31, code state 0bf89ffa). A feature that is not in the table sends nothing that was found. A row marked "Not verified yet" says so instead of promising.
You decide what runs: a feature that uses your own AI account or key sends the text of the job to that provider, and Growth OX never receives it. Your provider's own terms apply.
What each feature sends#
| Feature | What leaves | To whom | When | In Only this computer mode |
|---|---|---|---|---|
| Chats (Asistan) on the local model | nothing; the turn goes to Ollama/llama.cpp/LM Studio on 127.0.0.1 | 127.0.0.1:11434 (Ollama); 127.0.0.1:1234 (LM Studio, probed); 127.0.0.1:8888 (Unsloth, probed) | each message | Allowed: it stays on this computer. |
| Chats (Asistan) on a cloud model (Claude Code / Codex subscription CLI) | task text + inputs embedded in the prompt (up to 200 KB) and copied to input_dir: brain-context.md (cloud never gets local-only notes; secret-checked), project.json, status.md, items-*.jsonl, tracker.md, documents-index.txt or projects.json/awaiting-replies.jsonl, conversation.md (recent turns). PM-1 masking (TCKN, VKN, IBAN, card, phone, e-mail, people-list names) on by default. The CLI may also Read/Grep/Glob files on disk outside a deny list. | Anthropic (claude CLI) or OpenAI (codex CLI), through the person's own subscription login | each message where the chosen provider is cloud and decideRoute says cloud | Refused, with one line that says why. |
| Second brain / memory | nothing by itself; brain-context.md is part of a cloud call's inputs (DF-02) minus local-only notes | Nowhere | - | Allowed: it stays on this computer. |
| Brain daily sources and 'Kaynak ekle' (web, social, remote repos) | GET of the URL/feed/repo the person added; GitHub users/repos API | the added site; api.github.com; youtube.com/feeds | 06:30 scheduled feed (built-in scheduler) and on 'Kaynak ekle' | Refused, with one line that says why. |
| Document and PDF analysis / project from a file (Dosyadan proje) | nothing; readers run locally (PDFKit via osascript JXA, office readers), embeddings and extraction on local Ollama | 127.0.0.1:11434 | on upload/apply; watched files re-read on a slow tick (WORKSPACE_INGEST_WATCH) | Allowed: it stays on this computer. |
| Local Git analysis and the data repository | local workspace: nothing (origin is a bare folder on this computer). GitHub workspace: commits pushed to the private repo; records autopush every 15 min from the responsible Mac | (local) none; (github) github.com | on Paylaş; records autopush every 15 min (github workspaces) | Allow for local, block for github sync. |
| File create/edit (documents, changes, drafts) | nothing (git commit to the data repository) | Nowhere | - | Allowed: it stays on this computer. |
| Computer use | nothing: screenshots go through the shared model wire, which refuses an image for any route that is not local and loopback (allow_remote_images is never set outside tests); screenshot bytes are not stored | 127.0.0.1 (local vision model) | on an approved computer-use run | Allowed: it stays on this computer. |
| Local agent web access (web pack) | page fetches and search queries of a local-model run | any public host the run fetches; 127.0.0.1 SearXNG container, which queries its default engines | only for jobs with web=true (default false); refused when local-only context is present | Refused, with one line that says why. |
| Browser routes (AX-B fallback: Gmail, Outlook, Sheets, Drive, Paraşüt, Logo, Medium, Substack…) | nothing in production today: the browser route has no page unless a fixture sets one | Nowhere | - | Refused, with one line that says why. |
| Gmail | (a) mail content via claude -p and the claude.ai Gmail connector → Anthropic → Google; (b) metadata-scope reads | Anthropic (claude CLI); gmail.googleapis.com | scheduled browser collect at 09 and 18 for projects with a gmail source; discovery on user start (owner edition) | Refused, with one line that says why. |
| Outlook / Microsoft, Google Sheets, Google Drive, Paraşüt, Logo | nothing: no API code exists; only browser recipes, inert (DF-10). Logo also has a computer-use route that needs an approval | Nowhere | - | Refused, with one line that says why. |
| GitHub (device login, collectors, Actions reader, track, skills sync) | token + API reads (repos, stars, workflows, README badges); skills list fetched from public repos | github.com; api.github.com; raw.githubusercontent.com; registry.npmjs.org; pypi.org; api.securityscorecards.dev | skills sync 30 s after start then every 6 h; collectors on the 06:30 slot; hourly track run; device flow on user action; gh auth status/gh api user at start and when Bağlantılar renders | Refused, with one line that says why. |
| YouTube | uploads (approved), comment reads, video metrics; public feed reads | googleapis.com (youtube/v3, youtubeanalytics, upload); youtube.com/feeds | metrics hourly timer; uploads on an approved action or enabled automation | Refused, with one line that says why. |
| TikTok, Instagram, Facebook/Meta (incl. ads) | metrics reads; Meta ads spend reads and pauses; page comments | open.tiktokapis.com; graph.facebook.com | metrics hourly; Meta ads enforcer every 15 min | Refused, with one line that says why. |
| LinkedIn, X | API clients refuse before any request (UNVERIFIED); reads happen only through claude --chrome in the person's Chrome (page text → Anthropic) | Anthropic (claude CLI); linkedin.com / x.com in the person's browser | scheduled browser collect for projects with these sources | Refused, with one line that says why. |
| Other collectors (website, mentions, Etsy, Printful, RSS) and link-click tracking | GET reads with the person's keys; click pulls from the person's own redirect worker | the configured site; hn.algolia.com; dev.to; reddit.com; bsky.social; openapi.etsy.com; api.printful.com; <worker_base> | 06:30 collect; hourly track tick (daily pull) | Refused, with one line that says why. |
| Automations (n8n, Node-RED, Postiz) | to local services by default; user-created flows may call anything | 127.0.0.1:5678 (n8n); 127.0.0.1:5000 (Postiz); whatever a flow calls | 1-min tick of enabled automations; Node-RED starts at boot if enabled before | Allow loopback services; block flows' outbound (needs the MCP/tool boundary, decision 3). |
| Cloud model calls with an API key (model routes: Anthropic, OpenAI, Gemini, DeepSeek, Mistral, OpenRouter, xAI, Groq, Together, Moonshot, Z.ai, MiniMax, Qwen) | model gate: fixed test prompts only; key test calls on 'Test' and self-check 'retest keys' | the preset base_url per provider (presets.ts) | user action (gate run needs confirm_cost; key test) | Refused, with one line that says why. |
| Provider sign-in checks (Claude Code / Codex CLIs) | claude auth status --json, claude mcp get "claude.ai Gmail" and the CLIs' own start-up calls reach Anthropic/OpenAI even with no account connected | Anthropic (observed 160.79.104.10:443, 2607:6bc0::10:443 in a trial run); OpenAI (codex) | at start-up, when Bağlantılar/Model pages render, daily login check (WORKSPACE_CONNECTION_CHECK) | Refused, with one line that says why. |
| Voice / speech-to-text | nothing for transcription (whisper.cpp; network denied by sandbox-exec on macOS and by a network namespace on Linux; on Windows the shipped program is hash-checked against config/runtime/whisper.json before every start, plus an optional firewall rule where an administrator installed it); the transcript becomes a chat turn (DF-01/DF-02) | 127.0.0.1 (warm whisper-server, loopback TCP; a unix socket on Linux) | on push-to-talk | Allowed: it stays on this computer. |
| Phone remote | (a) phone ↔ API over Tailscale Serve (tailnet only); (b) Web Push {kind,key}, encrypted, VAPID sub carries the tailnet host name. The first pairing may also publish the phone page through a short-lived Tailscale Funnel (DF-29) | the person's tailnet; fcm.googleapis.com / push.apple.com / push.services.mozilla.com / notify.windows.com | listener only after pairing; push on attention changes (debounced) + 60 s poll | Refused, with one line that says why. |
| Licence and trial | activate/refresh/deactivate: raw licence key + device_hash (sha256 of machine/hardware id) + app_version; trial: device_hash; recover: e-mail + order ref | license_service_url (null by design: the public edition builds and runs with no licence service; a release build refuses a company or Cloudflare host here, by name; amended 12 Nov 2026: a Pages static mirror host is allowed only in the feed fields, see docs/release/cloudflare-pages-mirror.md) | refresh when 7 days passed, on GET /workspace/license/status; trial sync at most every 6 h; activate/recover on user action | Waits until you leave this mode. |
| Updates (app bundle and desktop shell) | GET of the manifest and archive (no identity); owner edition: GitHub releases with the keychain token; shell: nothing today (no feed is configured); only if the owner sets a feed host, one Tauri updater check at start (version/target/arch) | update_manifest_url (null by design; a release build refuses a company or Cloudflare host here, by name; amended 12 Nov 2026: a Pages static mirror host is allowed only in the feed fields, see docs/release/cloudflare-pages-mirror.md); api.github.com (owner edition); shell updater endpoint (none configured; the release overlay has no endpoints, and the build refuses a company, Cloudflare or workers/pages/r2/railway host by name, in the base and the per-OS overlays; a custom domain that Cloudflare serves cannot be recognised by name; amended 12 Nov 2026: one Pages static mirror host listed in trust.json static_mirror_hosts is allowed in the feed fields only, never a Worker, R2 or tunnel name) | API: when GET /workspace/update/status runs and 24 h passed (check_enabled and auto_download default true); shell: once per start, only when an owner-set feed host exists (update.rs configured needs a non-empty https endpoint list) | Waits until you leave this mode. |
| Crash reports and support bundle | nothing; the person sends the file themselves | Nowhere | - | Allowed: it stays on this computer. |
| Site and console analytics | console: nothing (CSP connect-src 'self' + 127.0.0.1:8790; no analytics SDK); site: no analytics script found | Nowhere | - | Allowed: it stays on this computer. |
| Backups | nothing from the app; a sync folder (iCloud/Dropbox/Drive/OneDrive) would upload the encrypted file, and the app warns | Nowhere | first snapshot 90 s after start, then hourly | Allow (warn when the copy folder is a sync folder). |
| Local model downloads and model finder | search terms and model ids; HF_TOKEN to huggingface.co only | huggingface.co / hf.co CDNs; registry.ollama.ai (by the Ollama daemon); raw.githubusercontent.com, github.com (publisher checks) | on search; downloads need approval | Refused, with one line that says why. |
| Phone page through Funnel (first pairing); signed phone updates are off | While a phone pairs and no relay is ready, the phone page (public build files only, GET and HEAD, /phone/ only; never the API) is reachable on the internet at https://<this computer's Tailscale name>.ts.net:8443. The computer's Tailscale name is in public DNS and certificate logs for that time. Funnel closes 120 s after a phone paired, 15 s after an unclaimed offer ended, and at the latest after 10 minutes; a marker survives a crash and the next start switches it off first. Signed phone updates are OFF in this release (one constant, PHONE_BUNDLE_ENABLED = false): the computer registers no phone-bundle route, creates no bundle key and puts no update key (bk) in the QR or the link; the phone ignores a bk, adopts and stores no key, shows no pin tap or short code, checks for no update and applies none. Until a later release, a page update comes from the person's own host when Funnel is on, and first-load and update trust rests on that host: the phone cannot verify the page it loads, and neither the page nor the service worker file sw.js is signed. The phone keeps its normal cache so the installed page opens with the host closed. | the public internet (Tailscale Funnel) for the page only, while pairing; public Nostr relays for the pairing exchange when no relay Worker is ready (DF-30) | only while a phone pairs; off otherwise | Refused, with one line that says why. |
| Computer presence and pairing over public Nostr relays | Signed presence beats (an event signed with a daily key derived from this computer's identity key, with a claim counter) so a second computer with the same identity is noticed; and the pairing exchange when no relay Worker address is in the offer. The relay sees the IP address of this computer and the daily key. It does not see the pairing code. | public Nostr relays from the signed pool (config/remote/nx-pool.json) | while the relay service runs (beats) and during a pairing (exchange) | Not verified yet. |
| Superpowers brainstorming picture (logo shown by the Superpowers visual companion inside a provider CLI) | Nothing, because the opt-out is set. Without it the companion would load one logo image from its vendor, and the address carries the Superpowers version (not the project, prompt or agent, per the vendor's README). | Nowhere | never: SUPERPOWERS_DISABLE_TELEMETRY=1 is set in the environment of every Claude Code and Codex child that Growth OX starts | Allow (no connection is made; the opt-out is a fixed value in the child environment, also present outside Strict Local). |
| PageIndex (advice rows: Python SDK, Cloud) | SDK: the model-price list is downloaded once per process from raw.githubusercontent.com (override LITELLM_MODEL_COST_MAP_URL) and document text goes to the model provider the person chooses. Cloud: the documents go to the vendor. All of it happens only inside the tool the person downloads or the account the person opens; Growth OX never fetches or runs the package. | raw.githubusercontent.com (SDK, price list); api.pageindex.ai (Cloud, vendor account) | only when the person installs the SDK or signs up for Cloud themselves | Allow (the app makes no connection; the tool is the person's own download or account). |
| treg (advice row: hosted pay-per-call data route) | Queries, including competitor names the person researches, go to treg and to its upstream data vendors; the vendor's CLI sends anonymous usage (command name, exit code, duration, version, OS, random install id) unless TREG_TELEMETRY=0 or DO_NOT_TRACK=1. Growth OX never fetches or runs install.sh and stores no treg key. | treg.to; treg's upstream data vendors | only when the person opens a treg account and uses it themselves | Allow (the app makes no connection; the tool is the person's own download or account). |
| Project safety (known problems in the packages a project uses) | package names and installed versions from the project's lockfiles go to the OSV advisory service; no file content, no path, no project name. A package update the person allowed runs in a sandboxed copy whose registry downloads (npm, PyPI, crates.io, Go proxy) go through the tool proxy. | api.osv.dev; registry.npmjs.org, pypi.org, files.pythonhosted.org, index.crates.io, static.crates.io, proxy.golang.org, sum.golang.org (only inside an update the person allowed) | when the person opens Project safety and presses Check, and once a day when the person turned the daily check on | Block (the check shows 'not checked'; nothing is applied). |
| Signed intelligence feed (recipes, platform rules, checks, advice and model lists) | a plain GET of static files: registry.json, registry.sig.json, revocations.signed.json, feed.manifest.json and the files it lists; no identity, no key, no device code, no body, no cookie | feed_url (null by design: with no feed address the app reads only the baseline inside it and nothing leaves; a release build never calls an address the owner has not set) | 30 s after start, then a daily check that fetches only when 7 days passed since the last success (1 day after a failure), or when the person presses refresh | Waits until you leave this mode. |
The calls the app makes itself#
These are the outward connections the app starts on its own account, each with the moment it happens and what a run showed.
| Call | When | Seen in a run | In Only this computer mode |
|---|---|---|---|
| skills sync (api.github.com contents) | 30 s after start, every 6 h | strict_sim + default; run:strict none; run:default (31 Oct, PRIV-M) none | Waits until you leave this mode. |
| claude auth status / mcp get, codex, gh auth status / gh api user (child processes that reach Anthropic, OpenAI, GitHub) | start-up, pages Bağlantılar/Model, daily check | strict_sim + default; run:strict none; run:default (31 Oct, PRIV-M) none | Refused, with one line that says why. |
| tailscale status/serve | start-up probe (default profile), pairing | default | Refused, with one line that says why. |
| connection probe: a DNS lookup of example.com | when the outbox list is read (console polls it) | default: one lookup of example.com; strict_sim: none (GROWTH_OX_FORCE_OFFLINE=1); run:strict none; run:default (31 Oct, PRIV-M) none | Changed so that no connection is made. |
| API-key model routes | user action | capture only | Refused, with one line that says why. |
| cloud CLI runs (claude -p / codex exec) | cloud runs | capture only | Refused, with one line that says why. |
| collectors | 06:30 slot, 09/18 browser hours | no (nothing configured) | Refused, with one line that says why. |
| studio: ads, metrics, caption model download, media fetch | timers (15 min, hourly) and user action | no | Refused, with one line that says why. |
| publishing, tracking, growth readers | 1-min route tick (approved posts), hourly track, user action | no | Refused, with one line that says why. |
| updates | status route (24 h), shell start | no (no host configured) | Waits until you leave this mode. |
| licence, trial, sales | status route (7 d refresh, 6 h trial sync), user action | no (no host configured) | Waits until you leave this mode. |
| model search and downloads | user action | no | Refused, with one line that says why. |
| local agent web pack, SearXNG | web=true jobs | no | Refused, with one line that says why. |
| Web Push | paired phone | no | Refused, with one line that says why. |
| discovery (Gmail metadata, registry) | user action | no | Refused, with one line that says why. |
| installer: GitHub workspace, owner release, bundle download | wizard / user action | no (local workspace) | Refused, with one line that says why. |
| automation readers (GitHub Actions public API, n8n) | automation list refresh, 1-min tick | no | Allow loopback, block GitHub. |
| brain sources | 06:30 feed, Kaynak ekle | no | Refused, with one line that says why. |
| local engines (Ollama, llama.cpp, LM Studio, Unsloth) | runs, probes | strict_sim: 127.0.0.1:11434, :1234, :8888 only | Allow (loopback; OLLAMA_URL pinned to loopback unless GROWTH_OX_ENGINE_LAN=1, localmodel/src/ollama.ts). |
| own loopback health checks | start-up | loopback | Allowed: it stays on this computer. |
| build/dev-time downloads | release build / developer | n/a (not runtime) | Not verified yet. |
| local tools | start-up and commits | strict_sim child_processes | Allow (docker inspect talks to the local daemon only). |
Questions#
- Does anything leave my computer when I use a local model?
- The text of your jobs does not. The rows above list the other connections (for example downloads and updates) with when they happen.
- What does Only this computer mode do?
- It refuses or pauses every connection to the internet. The last column shows the outcome for each feature.
- How do I know this page is true?
- It is generated from the audit file that the app's own network checks read. A test fails if the page and the file differ.