KEKKAI結界

GASLIGHT breaks agents. KEKKAI stops it. 🥷

The AI-Agent Firewall — a runtime ward around your agent's I/O. Like a WAF for web apps, an AIAF sits inline and screens every input, tool-call, and response for prompt injection, exfiltration, secret leakage, and tool abuse. Low-footprint, drop-in, one policy to manage. Below: the same agent takes the same 9 GASLIGHT kills — ward off, then on.

📦 Library

3 lines, zero dependencies. Wrap your agent's I/O in-process — nothing to run.

pip install kekkai

🧩 Sidecar gateway

One tiny container in front of any agent, any language. HTTP in, verdict out.

docker run kekkai

🖥️ Managed console

Edit policy, watch blocked events live, flip monitor→enforce.

open the console →

watch each kill chain break at input, tool-call, or output.

kekkai@ninja: ~/ward --demo unprotected · warded

// hit RUN — see 9 kills land, then get warded off. 🛡️

9/9

UNPROTECTED agent — KILLS

0/9

KEKKAI-warded — KILLS

Screen your own text through the ward

Wrap your agent in one policy. pip install kekkai · guard.inspect_input(text) · check_toolcall(tool,args) · inspect_output(reply)
Find the kills first with ⚔ GASLIGHT, close them with KEKKAI, prove it with the attestation. · join the clan 🥷