Security
1. Two tiers, and the boundary between them
Every account you connect starts on the core tier. Desktop talks to Gmail with an authorization-code grant, PKCE, and a loopback redirect; the refresh token lives in the OS credential vault. The web app uses authorization-code + PKCE too: Google returns a one-time code to this origin, the browser proves the verifier, and only the short-lived access token is held in browser memory and tab-scoped session storage (cleared when the tab closes or you sign out) — our server does not store these Gmail tokens. Sync does not upload the local mailbox or search index. AI is a separate permission: after you enable it for a mailbox, the client sends relevant thread content or bounded batches through our AI proxy, even when that mailbox has no cloud grant. Content is processed in memory and never persisted on our servers.
The cloud tier is an explicit opt-in you make one mailbox at a time. Enabling it issues our backend its own separate server-side OAuth token — never a copy of your client's token — so you can revoke cloud access for a single mailbox without touching the app on your desk. Only cloud-tier features may use that token, and only for the mailboxes that opted in.
No passwords, anywhere
We do not collect, process or store user passwords. There is no password field in Wingman Mail, no password database, and no reset flow, because authentication is handled exclusively through native Google and Google Workspace OAuth. You sign in on Google's own screens, and all we ever receive is the token that grant returns. A credential we never hold is a credential we cannot leak.
2. Where your mail actually lives
On the desktop app, your mailbox is mirrored into a local database on your own machine — threads, messages, drafts and the search index. Reads never touch the network, which is why the app is fast and why it keeps working when your connection does not.
Our servers store derived state only. Concretely, that means memories, snippets, automation specs, run records, reminders, synced settings, and the account roster. It does not include message bodies, subjects, headers, or attachments — not in databases, not in queues, not in caches, not in logs, metrics, or crash reports.
Cloud jobs carry account-scoped message identifiers and nothing else. A worker picks up an ID, re-fetches the content from Gmail, processes it in memory, writes its output to the proper home, and discards it. Generated drafts are written back through the Gmail drafts API rather than stored with us; Gmail is the content store, our database is not.
3. Credentials and keys
Desktop token custody
Refresh tokens live in your operating system's credential vault — DPAPI on Windows, Keychain on macOS, Secret Service on Linux — one grant per account, each revocable on its own. If the platform cannot provide encrypted storage, we keep the token in memory for that session only rather than writing it to disk in the clear.
The grant broker
Cloud-tier refresh tokens do not live with the application. They sit in a separate broker process with its own database file, envelope- encrypted with a per-grant data key that is itself wrapped by Google Cloud KMS. The key-encryption key never leaves KMS.
The application talks to the broker over a unix socket and may only ask for the specific Gmail operations automations need — reading history, fetching and modifying threads, managing labels, creating drafts, sending. The broker never hands a token back; it performs the operation or refuses. Deleting a message is not in the allowlist, so an automation cannot destroy your mail even if it tries.
Sensitive control operations — exchanging or revoking a grant, changing what a grant may do, pausing the pipeline — additionally require an application-only secret that the automation worker never receives. Compromised automation code cannot widen its own permissions. Disabling the KMS service account is the break-glass: it renders every cloud grant unusable without touching the database.
Account isolation and shared services
Requests and stored records are authorized for the signed-in user and owning mailbox. MCP apps receive only the verified cloud mailboxes you select, with separate permissions and revocation. These are application authorization boundaries, not a separate server or encryption key for every customer. The application, worker, database, and broker serve multiple customers; a compromise of a trusted service or its credentials could affect more than one customer. Broker separation limits token exposure but does not prevent an authorized compromised service from misusing its permitted provider operations.
Encryption at rest
Memories, snippets and automation specs are encrypted with AES-256-GCM before they are written, keyed from a secret held only in the server environment and never in source, logs or backups. Every external connection requires modern TLS.
4. How AI requests are handled
Model API keys never ship in the client. Every model call goes through our own AI proxy, which sends the minimum context the task needs — a relevant thread or a bounded batch — and enforces rate limits. Request bodies are parsed, forwarded, and dropped; when a provider call fails we log the status code, never the payload.
AI starts only after you enable processing for the relevant mailbox. Automatic summaries send up to eight threads per request; Trips sends up to four likely booking emails; memory learning sends up to 60 sent messages initially and 25 on later passes; scheduled reports use up to 60 inbox snippets. Memories are derived facts stored encrypted for your use, with separate controls to stop learning and delete what was learned.
On the provider side we set zero-data-retention and no-training flags explicitly in code on every completion, rather than relying on an account-wide setting that somebody could change later. The honest claim is that we do not store your email. It is not that we never see it — enabled AI and cloud features see content in flight, and we would rather say so plainly than hide behind wording.
AI output is treated as untrusted. Automations parse model output against a closed vocabulary of permitted actions, on the server and again in the client engine; the engine executes validated actions and never free-form model text. Model output cannot grant a permission or select a more privileged tool.
External AI apps
Connecting an MCP app requires separate consent for selected cloud-enabled mailboxes and optional reply drafts. The initial tools cannot send mail, read existing drafts, or download attachments. Apps identify themselves; that name is not proof of their identity or security. They must honor Google's Limited Use restrictions, including no advertising, resale, or generalized AI training. Their retention rules are separate from our OpenRouter settings. Revocation stops future access; request deletion of already received content from the app itself.
5. Untrusted content
Every message is hostile input until proven otherwise. Email HTML is sanitized before display, rendered in a sandboxed frame under a restrictive content-security policy, and message scripts never execute.
Remote images and the proxy
Remote images from senders you don't know are blocked by default; people you already correspond with and verified brands load automatically, and you can block every sender in Settings. When images do load, the client rewrites each approved URL to an HMAC-signed request on a dedicated image host, separate from the app origin, so a sanitizer mistake cannot reach the application or its cookies. The session cookie is host-only on the apex and is never sent to the image host.
The proxy fetches with no cookies, no referrer and no user-agent, so senders see our infrastructure rather than you. It allowlists image content types — SVG is refused — caps size and duration, resolves DNS itself and rejects private, link-local and carrier-grade NAT ranges, and re-validates on every redirect. Image URLs are neither logged nor persisted.
We will not oversell this: it is a privacy proxy, not an anti-tracking product. A unique tracking pixel still fires once, from our address rather than yours, on the images you choose to load. We do not prefetch, cache-bust, or silently drop pixels and claim you were never counted.
6. Transparency, pause, and deletion
You can see what the cloud tier has been doing. The Transparency view in Settings reports counts of provider activity by kind — reads, writes, drafts, sends, control operations, failures — with a per-day breakdown and the most recent operations, alongside a tally of every category of data we hold for you.
From the same place you can pause the pipeline, which is checked before every worker tick, or delete your cloud data outright. Deletion wipes the derived-state tables and records an opt-out marker, leaving only what identity and billing require. Deletion covers local replicas, caches, queued jobs, memories and model inputs — not just the primary row.
7. Reporting a vulnerability
If you find something, we want to hear it before anyone else does. Write to [email protected] with enough detail to reproduce the issue. We will acknowledge it, keep you updated while we fix it, and credit you if you would like us to. Please give us a reasonable window before disclosing publicly, and please do not test against other people's mailboxes.
The same contact details are published in machine-readable form at /.well-known/security.txt, per RFC 9116.