Chat interfaces remain the default surface for many large language model (LLM) tasks, yet the model only perceives the current dialogue and not the surrounding workspace. We introduce meta access, giving the assistant a structured, permissioned view of the interface and define a ladder of five levels: feature tutoring, metadata/settings/preferences access, retrieval over summaries and messages, confirmed mutation, and full orchestration. Meta access can collapse multiple screens into a single entry point for search, navigation, organization, and operation, while introducing risks such as grounding errors, privacy leaks, prompt injection, and trust issues. General‑purpose agentic runtimes can modify their own settings, so meta access must have a separate control‑plane boundary to remain reflexively safe. In the Meta Chat prototype we demonstrate these ideas: bounded context is built from summaries, settings, and relevant messages; references link conversations, topics, and personas; changes are confirmed and reversible; a browsable tutorial library is provided; and personas are recommended as user‑initiated conversations. By exposing application state in a structured form, meta access extends screen readers from widget descriptions to semantic workspace operations. Consequently, chat evolves from a window onto a single conversation into a control layer for the entire workspace, enabling assistants embedded in IDEs, dashboards, office suites, and operating systems to gain agentic capabilities without surrendering user control. Review