NHNH Applied
Home / Projects / NH Applied MCP

NH Applied MCP / User-controlled AI infrastructure

A bridge to the user's computer without making the relay the computer.

NH Applied MCP is a public Model Context Protocol bridge designed around a paired Agent on a user-controlled Windows computer. The relay routes authenticated requests; the Agent and local MCP policy remain the point where local authority is exercised.

0.40
Windows relay engineering RC
43 / 43
extended regression matrix pass
0
reported dependency vulnerabilities at freeze
WSS
outbound Agent transport

Architecture

The current relay candidate composes four public surfaces on one HTTPS service: MCP, device control, pairing control and an outbound Agent WebSocket endpoint. A Windows Agent connects outward, establishes identity and possession, and projects only the capabilities that the local policy layer allows.

Authority boundary: a public relay route an authenticated request, but local filesystem and process authority is exercised through the paired Agent and local MCP policy. The relay is not intended to become a remote filesystem in its own right.

What has been qualified

  • OAuth/JWT authenticated MCP plus separately scoped device and pairing controls.
  • First-run pairing with account binding and Ed25519 possession proof before attachment.
  • Windows CurrentUser DPAPI for protected local integrity-key material.
  • Persistent SQLite authority, session, pairing, revocation and device-selection state outside MCP-readable roots.
  • A production-style local TLS qualification using a genuinely separated Agent and an authorised file read routed through that Agent.
  • An extended 43/43 regression matrix and zero reported vulnerabilities in both production and full dependency audits at the 0.40 freeze.

What is deliberately not claimed yet

The engineering RC is not presented as a finished hosted product. Stable public VPS/DNS deployment, production human identity and account recovery, Windows service-account operation, cross-host high availability and OpenAI/ChatGPT end-to-end qualification remain separate gates.

Why build it

The practical objective is straightforward: let an AI system work with a computer the user owns while keeping the user's device, identity and local policy at the centre of the trust model. The design is intended to support self-hosting and bounded delegation rather than require every user's machine to route through one operator's private server.