Inventory and agent tools only read
- Your repos: the Xerg GitHub App can only read. It can’t push code, open pull requests or change settings. See GitHub.
- Your accounts: every connector lists, describes and counts. Each connector’s page has the exact permissions or scopes it uses, and how to make the narrowest key the vendor offers.
- Ask and coding agents: graph tools read the graph; production-context tools read observations and request scoped verification. They never execute external changes. An explicitly configured trusted host can consume a verification receipt before running its own fixed command. See Production context.
Secrets
- Xerg downloads example env files and keeps variable names, never values. It never downloads real
.envfiles. See What Xerg reads. - It downloads up to 300 source files per repo and scans them for imports, API hostnames and calls to the Expo push-token API. It keeps the service, the file and the line, never the code or call arguments. An owner or admin can turn this off in setup.
- Static
app.jsonconfiguration can identify Expo’s EAS Update or Heroku. Xerg does not run the configuration or keep update URLs, project IDs or query strings. - MCP configurations in the repository are downloaded to find configured agents, servers and recognized services. Xerg keeps configuration evidence, recognized package names, server names, transport, enabled state, environment variable names, and remote endpoint origins. It discards raw command lines and arguments, credential values, HTTP headers, URL paths and queries. It does not launch servers or test their access. See MCP servers on the graph.
- Connected keys and tokens are encrypted with AES-256-GCM, stored per workspace, used only on Xerg’s servers, and never sent to a browser. Disconnecting deletes them.
- Coding-agent keys are stored hashed, last 90 days and allow 600 calls an hour. Their maker or an admin can revoke them. For each key, Xerg keeps the name and version its MCP client reports (say, Claude Code 2.1.3), when it was last used, and how many calls it made each month.
Who sees what
A workspace’s graph and connections are visible to its people only, each with a role. Invites go to one email address and expire after 7 days.What’s stored
- The graph: names of apps, services, resources and configured MCP agents and servers, how they connect, and which file and line each came from. It also has the env variable names each app uses, and who owns what from
CODEOWNERS(GitHub handles and emails). File contents aren’t stored. - What connected accounts return: names, regions, status, counts and spend, user counts, Slack channel names, and the DNS targets that point at other services.
- History: a compact copy of each day’s graph (its items’ names, kinds and statuses, how they link, and saved MCP configuration metadata) for 180 days, even after the graph is cleared. See History.
- Production evidence: selected source definitions, deployment IDs/commits, timestamps, coverage gaps, checks, findings and review decisions. Opt-in collectors send allowlisted journey paths and audience labels, source hashes, runtime identities/relationships, contract results, webhook delivery/processing metadata and selected infrastructure mappings. Raw HTML, response bodies, cookies, source credentials and telemetry payloads stay off Xerg. Screenshots stay on the runner; only their hashes are submitted.
- Verification: actor and operation scope, target/artifact/command digests, policy/evidence revisions, decisions, overrides and host-reported outcomes. Observation history, event/audit records and verifications expire after 90 days; the current observation is retained even when older. Source configuration and findings remain until workspace deletion.
- Accounts: your email, name, GitHub username and profile picture from your sign-in method, your sessions, the tokens Google or GitHub issue at sign-in (encrypted), your role and when you were last seen.
- In your browser: the last 30 Ask Xerg messages, in the tab’s session storage, until the tab is closed.
- Analytics: see below.
Opt-in change workflow
Workspaces that enable the change workflow retain source-attributed changes and reports for 180 days, and underlying observations for 90 days plus the current observation. Human corrections, ownership notes, review decisions and notification preferences remain until the workspace is deleted. Reports label expired evidence rather than reconstructing it. Optional repository review collection keeps PR titles/descriptions, author/reviewer/merger identities and review decisions. Agent instruction analysis keeps paths, revisions and extracted references, not full instruction files; it never executes their contents. Restricted Kubernetes collection keeps allowlisted resource metadata and relationships, not Secrets, environment values, ConfigMap bodies or arbitrary custom-resource specifications. Gateway analysis keeps explicit alias-to-deployment mappings and model/version/region metadata. Retirement reviews store the submitter’s scoped usage checks, owner, evidence links and uncertainties. Report feedback stores usefulness, estimated time saved and corrections. Subscribed alerts and digests include selected findings and report links; delivery rechecks workspace membership and subscription preferences. Email delivery can be paused independently of collection. Clearing the current graph does not delete retained observations, reports or review history; deleting the workspace removes those records. The independently exported local core works without a Xerg account or telemetry. Local scans, history and collector exports remain on the local machine unless the operator explicitly uploads them. Publishing the local core does not publish the hosted application’s source or customer data. Public repos scanned without signing in are stored, and scanned again when someone opens one more than 24 hours later.Analytics and cookies
CookieYes requests consent in the EU and UK: Analytics cookies and Advertising are off until allowed. Outside those regions, Analytics cookies are on, and Advertising is on unless you opt out or send Global Privacy Control. The round privacy icon at the bottom left of xerg.ai opens cookie preferences where the banner applies, or a sale/sharing opt-out panel elsewhere. Advertising opt-outs are shared with the app. If CookieYes cannot confirm the applicable choices, the site uses cookieless analytics and keeps visitor identification off.- PostHog counts visits and pages. When Analytics cookies are not allowed, it sets no cookies and keeps nothing in your browser. With Analytics, a cookie keeps an anonymous ID across xerg.ai and the app.
- In the app, once you sign in and Analytics is allowed, events carry your Xerg account’s ID and your workspace’s ID, never your email, name or GitHub username. They record steps such as installing the GitHub App, your first graph, connecting an account and asking a question, never the question, your repos or what’s on your graph. Page addresses are sent without their query string. No session recordings.
- Journey and usage events also record sign-up method and invitation context, membership changes, remaps, successful public-repo explores (featured or not), checkout and subscription changes, and failed payments. They contain IDs, categories and counts, never repo names, service accounts, questions or payment details. An internal-account flag excludes Xerg’s own activity without sending email addresses.
- Without a browser, agent use is counted once per key per day, under a separate agent ID. Xerg stores the signed-in person’s latest Analytics choice: agent events require the key creator’s permission, and billing webhooks require the checkout initiator’s permission. Unknown or denied consent sends neither. Changing Analytics preferences in the app updates that choice and stops future events; it does not remove events already sent.
- Cardinal (run by RB2B, which also loads LiveIntent) identifies business visitors to xerg.ai, only with Advertising allowed and never with Global Privacy Control on.
- When you sign up, Xerg sends Cardinal your email, and your first and last name if your account has them, so Xerg can follow up. Not if your browser sends Global Privacy Control or you turned Advertising off.
Email sign-in
Email sign-in uses a necessary, 15-minute browser cookie containing a one-way hash of the requested link. It allows that browser to finish signing in automatically; other browsers require confirmation. The cookie is cleared when the link is verified.Who processes data
- Vercel hosts Xerg and runs its servers.
- Neon hosts its Postgres database (workspaces, graphs, encrypted connections).
- GitHub provides sign-in and repo access; Google provides sign-in if you choose it.
- Resend sends sign-in, invite and explicitly subscribed report emails.
- Stripe handles payment for paid plans, once billing is live: it receives the owner’s email, the card details they enter on Stripe’s own page (Xerg never sees them), the workspace’s ID and how many members and agent seats are billed.
- Anthropic runs the model behind Ask Xerg: it receives your question, the conversation’s earlier messages, the workspace’s name and what Ask looks up on the graph.
- PostHog runs Xerg’s analytics; CookieYes shows the cookie banner and keeps your choice; Cardinal, RB2B and LiveIntent support business-visitor identification on xerg.ai with Advertising allowed, and Cardinal receives a new account’s email and name (above).
- Brandfetch and the Simple Icons CDN serve service logos to your browser.