agentzt CLAUDE.md

CLAUDE.md

Guidance for Claude Code (and other agents) working in this repository.

What this is

agentzt — Zero Trust for AI Agents. An AI agent reaches enterprise model APIs and tools only through a zero-trust proxy, the same way a human reaches enterprise apps through a ZTNA client + gateway. It does not replace ZTNA for humans; it secures the agent→model and agent→tool path.

• agentzt-client (src/client) — local proxy + per-agent identity. The agent points its model SDK at it (ANTHROPIC_BASE_URL=http://localhost:8787); the client attaches a short-lived token to every request and forwards to the gateway.
• agentzt-gateway (src/gateway) — PDP+PEP. Authenticates the agent, issues scoped short-lived tokens, enforces deny-by-default RBAC + ABAC + JIT, runs input/output guardrails, audits everything, then forwards to the real model API / tools.

Open source, Apache 2.0, repo: https://github.com/openguardrails/agentzt

Hard constraints — read before editing

• Zero runtime dependencies. No build step. Everything runs on Node ≥ 22.6 via native TypeScript type-stripping (node src/...ts). Do not add npm dependencies or a bundler. package.json deps stay empty (only @types/node + typescript as devDeps for editor/typecheck).
• Because of type-stripping, write erasable TypeScript only: no enum, no namespace, no constructor parameter properties, no decorators. Use import type for type-only imports and .ts extensions on every relative import (ESM).
• The only allowed external process is openssl, and only in src/cli/tls.ts for opt-in mTLS PKI setup. The runtime never shells out.
• Use Node built-ins only: node:http, node:https, node:crypto, node:tls, global fetch. Note global fetch can't carry client certs — use src/client/transport.ts (which falls back to node:https) for outbound client calls.

Commands

npm run demo        # end-to-end: enroll, start gateway+client, drive 8 control checks, audit
npm run demo:mtls   # same, over mutual TLS (needs openssl)
npm test            # 28 checks (auth, RBAC, guardrails, ABAC, JIT, audit chain, mTLS)
npm run typecheck   # tsc --noEmit (needs `npm i` once for typescript)

npm run gateway                                  # start gateway (PDP+PEP)
AGENTZT_AGENT_ID=<id> npm run client             # start client proxy for an agent
node src/cli/index.ts enroll --agent <id> --role <role> [--mtls]
node src/cli/index.ts roles | agents | audit [--verify] | tls init | tls issue --agent <id>