ShinrAI Overview Live demo Benchmarks Platform Connector Partners Open beta
Desktop app

On-Device Protection
for Your AI Tools

The ShinrAI Connector is a small tray app: it routes local AI tools through on-device privacy protection before anything leaves your machine — and shows you, request by request, exactly what it did.

Claude Code · Codex · SDK tools
100% on-device
Monitor before you protect
Loopback only — never LAN
The app

The App at a Glance

Green means the daemon runs and your tools route through it. Every counter is inspectable — down to a per-request transcript.

ShinrAI Connector protecting
Off
Monitor
Protect
Requests today247
Entities protected1,032
DetectionShinrAI 1.1 · on-device
Routed toolsClaude Code · Codex
Transcript: last 50 requests, in memory only — gone when the connector stops. Replacements are scoped per conversation.
How it works

How It Works — Fully Local

No cloud component. Four steps, all on your machine:

Sit on loopback

A native daemon opens listeners on 127.0.0.1 only — the Anthropic and OpenAI base URLs of your tools point at it. No LAN binding exists in the code.

Detect on device

Three engine tiers: built-in patterns, patterns + name pools (no download), or the full ShinrAI neural model (1.2 GB, fp32) — a compressed build is available for smaller machines.

Protect outbound

In Protect mode, names, places and organizations are pseudonymized before the request leaves your machine — and restored in the streamed answer as it returns.

Show the receipts

Live counters, per-request transcript, and a diagnostics copy button. If the daemon dies, the icon says so — a green light over a dead daemon is treated as the worst possible bug.

Compatibility

What It Routes Today

We publish what works and what does not — you should not have to discover it.

ToolRoutableWhy
Claude Code ✓ yes Respects the Anthropic base-URL override — full protect + restore.
Codex ✓ yes Respects the OpenAI base-URL override — full protect + restore.
Anything on the vendor SDKs ✓ yes Custom agents, scripts and internal tools built on the Anthropic/OpenAI SDKs route cleanly.
Cursor not yet Sends traffic to its own cloud, assembled server-side — nothing local to intercept.
Claude Desktop / ChatGPT app not yet Ignore environment overrides; routing them needs local TLS interception, which the app deliberately does not do yet.

Routing uses system environment injection for GUI and terminal launches. Already-running processes keep their old destination until restarted — the app tells you that too.

On-device models

Choose Your Model Size

Measured on Apple Silicon — resident memory after one inference, agreement scored against the full-precision model on 300 sentences across four languages.

Detection tierDownloadMemory in useQuality
Patterns (built-in) 22 MB Structured PII: emails, phones, cards, IDs — format-driven, language-independent
Patterns + names (built-in) ~25 MB Adds German + English name pools — instant, no download
ShinrAI compact (INT8) 1,014 MB 93.8% agreement with the full model · 138 ms per 512-token prompt
ShinrAI full (fp32) 1.2 GB 2,846 MB Reference quality · 159 ms per 512-token prompt

Compact vs. full — the real difference

The compact model is not the full model, and we say so in the app, in numbers: in field tests it finds 57% of street addresses where the full model finds 98%. Roughly one entity in sixteen that the full model would catch goes out as written. If your texts carry addresses, download the full model — that is exactly the kind of decision the app is built to let you make consciously.

Security

Security Properties

Loopback only, ever

Control API and both listeners bind to 127.0.0.1. The UI window never opens a socket itself and holds no control token — the main process pushes it snapshots.

Metrics that cannot leak

The statistics store is schema-constrained to counts — text columns are allow-listed at build time. Request bodies live in a RAM-only ring buffer with no serialization path.

Conversations stay separate

Each conversation keeps its own replacement map — a name from one chat can never surface in another. Restores stay lossless per thread.

Updates on your terms

Signed and notarized builds; a weekly background check downloads quietly and installs only on your consent. Models pull from Hugging Face with hash verification and resume.

Platforms

macOS (Apple Silicon, notarized) today. Windows and Linux builds are in preparation on the same codebase — the beta form is where to raise your hand for them.

Honest failure states

Three restarts with backoff, then the tray goes red and says the daemon is down. Clear words for sensitive contexts: patient or case material gets named as such, with the choice yours.