agent-sec

Kernora Agent Security

A zero-install security baseline for AI coding agents. Point Claude Code, Cursor, or any MCP-capable agent at one URL and it reads a curated, cited security baseline every session (via a minimal JSON-RPC-over-HTTP MCP subset — works with Claude Code today) — so it stops shipping the mistakes that cause incidents (a hardcoded secret, a curl | bash, a prompt-injected "send this to that URL"). When it catches one, it cites the exact rule.

Advisory grounding, free and open. Real-time blocking against your organization's own decisions — plus a tamper-evident audit ledger for EU AI Act / SOC-2 — is the paid Kernora Axiora Integrity Plane.

Live: https://agentsec.kernora.ai

Connect in one line

Add it as an MCP server in your agent's config:

{
  "mcpServers": {
    "agentsec": { "url": "https://agentsec.kernora.ai/mcp" }
  }
}

That's it — no install, no signup. It's read-only: get_security_baseline sends nothing, and check_action sends only the short action text you choose to pass it (never your files, repo, or environment).

Tools exposed:

• get_security_baseline — the full known-good rule set.
• check_action — pass an action/command; get back the baseline rules that apply, so the agent can self-correct. Advisory only — it does not block.

Or read the rules directly: /baseline.json · /baseline.yaml

What's inside

15 rules across categories including secrets, injection, supply-chain, destructive ops, permissions, transport, data protection, exfiltration, authz, and VCS safety. Every rule cites a real source — OWASP (incl. the LLM Top 10), CWE identifiers, and regulations (EU AI Act, PCI-DSS, HIPAA).

Self-host

It's a single Cloudflare Worker with no dependencies, no storage, and no secrets.

npm install -g wrangler   # if needed
npm test                  # runs the local test suite (node test.mjs)
wrangler deploy           # deploy to your own Cloudflare account

Air-gapped or privacy-strict? Run it on your own infrastructure and point your agents at it.

Free grounds. Paid blocks.

Talk to us about blocking + attestation: hello@kernora.ai

How it relates to endpoint monitors

Endpoint agent monitors (e.g. Perplexity's Numbat) watch for generically suspicious behavior. Kernora Agent Security supplies what they lack: what's known-good for your codebase. The two are complementary — this project can even export its rules for a monitor to enforce.

Honesty note

This endpoint provides advisory grounding only. A matched rule is guidance; an unmatched action is not an assurance of safety. Real-time enforcement and audit are the paid Integrity Plane.

Claude Code plugin (optional, one command)

Instead of the manual MCP config, install the plugin — it bundles the MCP server, a skill that tells your agent to consult the baseline, an advisory PreToolUse hook (warns before risky commands), and helper commands:

/plugin marketplace add kernora-ai/agent-sec
/plugin install kernora-agent-security

Commands: /agentsec-rules (show every rule), /agentsec-check <action>, /agentsec-report (a LOCAL view of how many risky actions it caught — nothing is sent to us). Disable the hook anytime with AGENTSEC_HOOK=0.

See every rule / how effective / how to remove

• See the rules: readable at https://agentsec.kernora.ai/rules · raw at /baseline.json (or .yaml) · or the get_security_baseline MCP tool. Nothing is hidden.
• Effectiveness / reporting: the free tier is stateless and private — we store nothing. The plugin logs flags LOCALLY (~/.agentsec/flags.log); /agentsec-report summarizes them. Full per-action reporting with a tamper-evident audit ledger is the paid Kernora Axiora plane.
• Remove it: Claude Code — claude mcp remove agentsec (+ /plugin uninstall kernora-agent-security if installed). Cursor — delete the agentsec entry from ~/.cursor/mcp.json and restart. It's read-only and stores nothing, so nothing is left behind.

About

Kernora Agent Security is built by Kernora on the open Factlet Protocol. Licensed Apache-2.0 (see LICENSE).