Publication statusThis page is an implementation draft. Open operator decisions and legal review must be completed before publication.
01One customer, one machine
Every customer gets a dedicated, isolated virtual machine. Your data does not share a machine, database, or bucket with another customer.
- Each Mainframe has its own Google Cloud virtual machine, service account, storage bucket, database, and backup schedule. The machine is the trust boundary.
- We do not pool Customer Content across customers.
- Published machine tiers make the underlying infrastructure visible instead of hiding it behind an abstract usage limit.
02Your choice of region
Choose the European Union or United States at signup. Your machine, disk, database, and backups stay in that region.
- Launch regions are europe-west1 in Belgium and us-central1 in Iowa.
- Your billing country sets a suggested region, but you can override it.
- Connected services and selected AI providers may process data outside your machine region. Those flows are explained in the Privacy Policy.
03Encryption and private access
Data is encrypted in transit and at rest.
- Google Cloud default encryption protects virtual-machine disks, snapshots, buckets, and databases at rest.
- Web and API surfaces use TLS in transit.
- Administrative access runs over a private Tailscale network. Tenant machines do not expose public SSH.
04Credentials stay out of the workspace
Passwords and API keys live in a dedicated secret store. Agents use references, not raw secret values.
- OAuth tokens and API keys live in a per-customer secret namespace. Connector code resolves a reference only when it runs.
- API keys are collected through single-use secure forms that expire after 15 minutes, never through chat.
- AI subscription logins use the provider login flow on your machine. Browser-session passwords are entered directly into your machine through a live takeover window.
- Secret values are excluded from workspace files, chat history, control-plane responses, and logs, with automated checks and redaction.
05Connectors ask first
Every connector declares what it can reach, which secrets it needs, and whether it can write. You approve that manifest before installation.
- The default posture is read-only or draft-only. Sending, posting, or writing is a separate permission upgrade.
- OAuth credentials for remote connectors are injected by the authentication gateway, so models do not receive raw tokens.
- Remote tool lists are pinned at installation. A changed tool set pauses for re-approval instead of silently gaining capability.
- Installed connector code is visible in your workspace and can be inspected, edited, or removed.
06Agents draft, you approve
Email, messages, posts, and other external actions wait in Approvals. Agents cannot approve their own work.
- Approval authority is bound to your identity, not to an agent credential.
- The permission floor is enforced by the runtime. An agent cannot prompt its way around a capability it has not been given.
- Memory, concurrency, and worker budgets limit resource consumption by tier.
07You choose the AI provider
Use your own AI subscription, your own API key, or bundled credits. Relevant context goes only to the provider selected for that request.
- Claude and ChatGPT subscriptions run under your direct provider relationship from your dedicated machine.
- Bring-your-own API keys bill directly to your chosen provider.
- Bundled credits route through OpenRouter using a separate, hard-capped key for your account.
- Mainframe does not use Customer Content to train models. Provider handling remains governed by the selected provider's terms.
Before publication, Inflection Systems will document the exact OpenRouter routing and data-policy configuration used for bundled credits.
08Inspect what happened
File changes are version-controlled, sensitive actions are audit-logged, and operational logs avoid prompt text and secret values.
- Agent-made file changes are committed with provenance, so they can be inspected and reversed.
- Restores, browser takeover sessions, and secret access create audit events.
- Supervisor logs record operational metadata such as event type and byte counts, not prompt text, goal text, or secret values.
- Technical users can inspect their own machine directly.
09Backups and restore
A rolling 14-day catalog protects the workspace, database, and connector configuration as one coherent set.
- Backups live in the selected region and are encrypted at rest.
- A safety snapshot is created before a restore.
- Restore operations create a durable audit record.
10Export before deletion
Account wind-down is gradual, with export offered before deletion.
- Payment failure starts a 7-day grace period, followed by 30 days suspended and 90 days archived before deletion.
- The archived state retains a final snapshot, database export, and workspace bundle while deleting the running machine and disk.
- You can export the workspace repository, files, and database at any time.
11Incident handling
Affected customers will be notified without undue delay, in line with applicable data-protection obligations.
- The formal incident-response process and any public response-time commitments are still being finalized.
- No specific response-time guarantee will be published before the process exists operationally.
12Providers in the data path
The architecture uses a short list of infrastructure and service providers, while services you connect remain your own provider relationships.
- Google Cloud provides machines, storage, databases, backups, and secret management.
- Stripe handles payments and tax. Mainframe does not store card details.
- OpenRouter supports bundled AI credits. Telegram carries messages when you choose that surface. Tailscale provides private networking. ElevenLabs supports voice replies when enabled.
- A final live subprocessor register, including locations and provider links, is required before publication.
13Shared responsibility
We secure the platform. You control what it may touch and what leaves it.
- Inflection Systems is responsible for the isolation model, platform software, secret store, permission floor, backups, and providers it selects.
- You are responsible for approvals, connected-account scopes, third-party terms, changes made through direct machine access, and the security of your own login and Telegram account.
14The honest limitations
No system is perfectly secure, and AI agents can make mistakes.
- Mainframe does not currently hold SOC 2 or ISO 27001 certification and will not imply otherwise.
- AI output can be wrong, incomplete, or manipulated by content it reads. Review anything you approve.
- Connected apps, model providers, and communication channels have security properties outside our control.
- This page describes the current architecture and will evolve with the service.