OneCLI: a firewall for your AI agents that also hides your API keys
A Docker gateway that scopes credentials per request and rate-limits or blocks risky calls across MCP, CLI and agent-written code. Free for 2 agents.
Copy markdown
Keys your agent never actually holds
Point Claude Code, Codex or Cursor Agent at OneCLI's gateway and give it a placeholder key; it matches each outbound request by host and path, then injects a scoped real credential (AES-256-GCM at rest) so the model, its logs and its context never touch a live secret. Revoke one agent without rotating your root keys.
A firewall, not just a vault
OneCLI also enforces policy at the network layer across four agent paths — MCP tool calls, CLI commands, curl and agent-written code: block specific endpoints, rate-limit per agent, scope access per project, and gate sensitive operations behind approval.
Run it this weekend
Apache-2.0 and self-hosted in one Docker container, with a dashboard to manage agents and secrets; free forever for up to 2 agents, no card. Timely: this week's mrmustard PyPI stealer hunted exactly the SSH, AWS and Kubernetes keys a gateway like this keeps out of an agent's reach.
One gateway, every agent
It sits in front of anything that makes HTTP calls or speaks MCP, so one policy covers a Claude Code session, a Codex run and a background cron agent alike. Point them at the gateway once and every new tool they pick up inherits the rules — no per-tool secret wiring.