AI safety that runs in process,
built for the .NET stack.

No Python sidecars. No per call scanning fees. Our libraries run inside your ASP.NET Core process, so prompts never leave your servers to be scanned. The audit trail stays there too.

What we build
Available today

Three libraries that compose.

Guard screens what goes into the model and comes back out. Gate decides what an agent is allowed to do. Evidence proves what was decided. They answer different questions and all run inside your process.

Products

What we build.

Runtime tooling that installs from NuGet and runs inside your .NET application. Your data stays on your hardware.

Invarix.Guard

AI safety middleware for .NET

Scanner that screens inputs for prompt injection, PII leakage, toxicity, and 19 harm categories before they reach your LLM. ASP.NET Core middleware or standalone engine. Runs in process.

  • Free Community tier on NuGet
  • 19 harm categories, including hate speech, self harm, and malicious code
  • Hot reloadable custom blocklist (Pro tier)
  • Lifetime license, no phone home

Invarix.Guard.Evidence

Sealed audit records for AI decisions

Tamper evident records for the AI decisions you instrument. You seal batches under your own Ed25519 key, and an auditor can prove that nothing changed after sealing.

  • One JSON file plus your public key: RFC 6962 inclusion proofs, verified offline
  • Batches sealed with your key, which stays on your hardware
  • No prompt fields in the record, by construction
  • CloudEvents 1.0 on the wire, a documented format you can reimplement
Learn more v1.0.0-rc.1

Invarix.Gate

Action firewall for AI agents

Reads every tool call routed through it and flags the ones that cannot be undone. Deterministic rules and counters, no second model in the decision path, and a report of what enforce mode would have stopped.

  • Destructive SQL, shell and git commands, cloud teardown, credential egress
  • Call and token ceilings evaluated before the call, not read off an invoice
  • Observe mode is the default and blocks nothing, including runaway loops
  • A run report that names what it could not see
Learn more v1.0.0-rc.1