Approvals & privacy
The interface reflects enforceable permissions.
Approval requests
The ChatGPT agent runs with the official engine’s untrusted approval policy and workspace sandbox. Supported command and file-change approval requests ask you to Approve once or Decline. Unknown interactive requests are declined, rather than automatically granted.
Local and remote data
API mode sends the recent conversation and explicitly selected text attachments to the provider. ChatGPT agent mode can read the chosen workspace as required for your task. Review the task and approvals before sharing sensitive material.
Stored data
Chat history, settings, metadata, and local usage records are ordinary JSON on this PC. API keys and vault values use Windows safeStorage. ChatGPT and MCP sign-in use the official engine with OS-keyring credential storage. These are different storage mechanisms.
MCP authorization
Notion, Linear, and Figma connectors open their official browser consent flow. Review the provider permissions yourself. A catalog entry or an installed app does not automatically grant service access. The 74-provider catalog includes setup guides, preset provider endpoints and registry-listed options identified by publisher. Account permissions and compatible authorization are still required; some services need a custom adapter or registered OAuth application.
Security boundaries
The app isolates its renderer, disables Node integration and direct renderer network calls, and uses a narrow IPC bridge. A local profile is not an OS security boundary. No paid OneX plan, unlimited provider access, or bypass of account limits is granted by installing the app.
Documentation reviewed November 20, 2026.
OpenAI billing and API accessAnthropic credential guidanceGemini model list APIElectron safeStorage security boundaries