Hey Product Hunt 👋 I'm Ziyad, maker of AgentWing. I kept seeing coding agents get powerful enough to do risky things — reading .env files, installing packages, and running shell commands in the wrong directory. Most teams have logs, but no enforcement before the action actually runs. AgentWing is a runtime control layer that sits between your agent and the tools it wants to execute. Bring your own agent — whether it's a coding agent, Claude Code/Cursor-style workflow, LangChain loop, or custom agent runtime — AgentWing handles the check. One API call before tool execution: POST /api/v1/check-action { "actionType": "shell_command", "tool": "terminal", "command": "npm install lodash", "description": "Agent wants to install a dependency" } → { "decision": "sandbox_required", "policy": "sandbox-node-command", "receiptId": "aw_receipt_8f2a..." } Five decisions: allow, block, sandbox_required, approval_required, restore_point_required. Every check produces an auditable receipt. A few things AgentWing is NOT: - Not an agent. Bring your own. - Not a sandbox. It routes sandbox_required actions to your BYOK E2B sandbox. - Not a prompt/guardrails filter. It checks actions, not prompts. Apache-2.0 licensed. Open-source repo + hosted dashboard for receipts, policies, and usage. Repo: https://github.com/12ziyad/agent... App: https://agentwing.gpmai.dev Question for builders here: if you're running an agent in production (or trying to), what's the one action type you'd most want blocked or sandboxed automatically? I want to make sure the default policies cover what people actually fear. Will be here all day — happy to dig into integration questions.

AgentWing

Check every AI agent action before it runs

Check every AI agent action before it runs