Tkach Security

A fail-closed security boundary for AI agents and compromised model output.

The model proposes. Tkach authorizes.

Put Tkach between model proposals and protected actions. Its Rust Core checks authority and information flow before allowing an action or releasing output. Model output remains untrusted data, even when the model is compromised.

Why Tkach

• Explicit authority: protected actions need an exact-scope, Core-issued Propusk; model text cannot create one.
• Controlled data flow: permission to read is not permission to export. Provenance and release checks remain inside the boundary.
• Fail-closed decisions: invalid, denied, replayed, cancelled, or uncertain states do not silently become permission.
• Isolated secrets and bounded evidence: broker-held secrets stay outside ordinary model context; Sled receipts do not include payloads.

These guarantees apply to paths routed through Tkach. It does not make the model trustworthy, detect every prompt injection, or protect a compromised host.

Start in minutes

Install from crates.io with Rust 1.85 or newer:

cargo install tkach-cli --version 0.1.1 --locked
tkach init my-agent
tkach check my-agent/.tkach/request.json
tkach run --demo

init creates a starter request without overwriting files; check validates its schema, not permission to execute; run --demo exercises the offline boundary without a model connection. Use tkach ui for the interactive panel.

Prefer no Rust toolchain? Download a binary below and start with tkach init.

Packages and downloads · v0.1.1