In development. This page describes the intended architecture and planned workflows. Organization administration and policy authoring are implemented; managed client delivery and regulated workflows are not available end to end. See current availability.

Architecture

Glyph Trust

Glyph Trust is a security boundary between your content and everything outside your device: models, tools, connectors, and the web. Two planes keep it fast and private: the Fortaify control plane manages identity, policy, and audit metadata; the target architecture enforces policy locally in Glyph Desktop and Glyph Mobile. Managed client integration remains unfinished.

Target request architecture

  1. Normalize — prompts, retrieved context, files, and tool structures are collected with provenance.
  2. Inspect locally — secret, sensitive-data, file-type, and prompt-injection detectors run on device.
  3. Evaluate policy — a signed organization bundle (or a built-in offline-safe baseline) decides.
  4. Enforce — allow, warn, confirm, redact, tokenize, reroute, or block before anything is transmitted.
  5. Quarantine output — model responses are buffered and inspected before display, logging, or persistence.
  6. Record metadata — decisions, categories, and destinations are audited; content is not.

Intended integration coverage

  • Cloud model requests (chat, vision, agent runs) — inspected before any provider transport.
  • Agent tool execution — arguments and results are gated independently of user approval mode.
  • MCP connectors and direct tool calls — same enforcement pipeline, no bypass route.
  • Web search and page fetch — queries are scanned before sending; fetched pages are scanned before use.
  • Attachments and pasted files — sensitive file types are blocked before upload or extraction.
  • Knowledge-base retrieval — classifications propagate to chunks and gate retrieval.

Design goals and limits

Organization policy cannot be weakened by a prompt, a retrieved document, model output, or an agent mode. External flows fail closed when policy is missing, expired, or invalidly signed. Detection is layered and continuously improved, but no detector is perfect: Glyph Trust is a control that reduces risk, not a guarantee that every sensitive value is caught. Purely local inference continues to work offline under the built-in baseline.

Managed deployment is in development for Glyph Enterprise. Healthcare and legal capabilities are described in healthcare and legal workflows. Operational posture is documented in the trust center.