AgentRaaS - Approvals, spend caps and undo for AI agent actions

Hi Product Hunt, I'm Sumedh, and I build AgentRaaS on my own. AI agents are good at deciding what to do. The risky part is when they do it: an agent calls Stripe, the request times out, the agent retries, and the customer is charged twice. Or a loop sends the same Slack message fifty times. Or an agent moves money at 3am with nobody watching. AgentRaaS sits between your agent and the real APIs and checks every call first: - Same call twice? The second one gets the first result back instead of running again. - Risky action? It waits for a person to approve it in Slack, Teams, Telegram, Discord, email or the dashboard. - Spend caps, daily money limits and allowed hours per agent. - A circuit breaker when a service is failing, and a dead-letter queue you can replay. - A tamper-evident audit log of every action, and undo for refunds and messages. - Works with plain webhooks, the Python and JS SDKs, LangChain, CrewAI, the Vercel AI SDK and MCP. 31 services built in, plus your own endpoints. It's open source and free to self-host, approvals, SSO and team features included. It runs on Linux, macOS or Windows (WSL2) with one ./install.sh. Your API keys stay on your own server, and can stay in your own Vault or cloud KMS. There are also free tools that need no server: pip install agentraas gives you a chaos tester for your agent's retries and agentraas scan for checking an MCP server before you connect it. I'd love to hear what your agents do that worries you, and what you'd want checked before it happens.