PrismAgent is a private, local, conversational screen companion that combines structured accessibility information with visual interpretation, and can expose the evidence behind its own local execution.

PrismAgent runs locally. The user's data does not leave the device. The agent observes the screen through the accessibility tree and authorized visual context. The screen is the surface. The engine is the authority.

A user asks a question about what is happening on screen. The agent reads the accessibility tree, looks at the authorized visual context, and answers. The answer is local. The receipt is local. The evidence is local.

PrismAgent is one product. Its role is to expose Prism's evidence to a user who is looking at a screen, not at a CLI or a configuration file. The diagnostics surface is a capability. The Compiler Lab access is a capability. The multimodal assistance is a capability. The product's reason for existing is the conversational screen companion.

PrismAgent observes. PrismAgent does not make compilation decisions. PrismAgent does not rewrite receipts. PrismAgent does not represent itself as the engine. PrismAgent is an instrument that reads the engine and answers the user.

When the agent exposes a Prism receipt behind an answer, the receipt is the real receipt from the engine. The agent does not synthesize a receipt. The agent surfaces what the engine produced.

PrismAgent is Unreleased in v1. The page surfaces this state explicitly: the product is being developed, the build is available from the source path, no packaged release is published. When a release record appears in releases.json, the page updates: the version, the signature, the support boundary, the installation path.

The page is not a marketing surface. The page names the product, its boundary, and its current state. A visitor who comes to the page looking for "is it ready to ship" finds the answer: the distribution state is Unreleased; the source path is reviewable; the release is a milestone in the roadmap.