Custos — The bouncer for your AI agents - Stop AI agents from doing things they shouldn't.

Custos is a safety and control layer for AI agents. It sits between an agent and the tools it uses—like APIs, files, databases, shell commands, and MCP servers—and checks every action before it runs. You define what the agent is allowed to do, Custos blocks anything outside those rules, and every action is recorded in a tamper-evident audit trail. It’s open source, works with Python and TypeScript, and is built for teams running AI agents in real-world environments.

Hey PH! Maker here. I built this after seeing enterprise teams blocked on AI agent deployment because they couldn't answer "what did the agent actually do and can you prove it?" The MCP ecosystem is growing fast but the governance layer is missing. Custos fills that gap — policy evaluation + cryptographic audit trail in a single library. The cross-runtime wire compat (Python ↔ Node verified) was the hardest technical piece — had to get byte-identical canonical JSON and Ed25519 wrapping right on both sides. Would genuinely love feedback on: - The policy DSL — too simple? too complex? - What evidence bundle formats regulators actually ask for - Use cases I haven't thought of Ask me anything 👇 Gallery assets to create: 1. Hero: dark terminal showing gate.call() → signed ledger record JSON 2. Flow diagram: Agent → Gate (policy check) → Tool → Ledger 3. Cross-runtime diagram: Python sign ↔ Node verify 4. Policy YAML snippet with syntax highlighting

Custos — The bouncer for your AI agents - Stop AI agents from doing things they shouldn't.

Replies