Hi Product Hunt 👋 AgentLock is a security gateway for AI agents. The problem it solves: for an agent to actually be useful, it needs access to APIs, databases, and servers — but handing it permanent API keys is a disaster waiting to happen. One prompt injection or a single bug, and the agent wipes production data or leaks secrets. Most existing solutions in this space are enterprise-first: SSO, RBAC, compliance sales cycles. Built for big orgs, not for the people actually shipping agents into production today. AgentLock works differently: → The agent never sees secrets. It only holds an Ed25519 keypair. Every action is signed and sent to the gateway, which verifies the signature, checks for replay, and evaluates the policy → ALLOW, REQUIRE_APPROVAL, or BLOCK. Real credentials live encrypted in a vault (envelope encryption: master key → workspace DEK → libsodium secretbox) and are only decrypted in memory on the runner, never returned to the agent. → Approvals come to your phone. When an action needs review, you get a push notification. Action preview, raw payload drawer, hold-to-approve with Face ID. Two-person rule optional. The whole loop takes seconds — no need to be at your laptop. → One-click undo. Connectors that support it (e.g. DB writes) emit an undo payload on execution. Made a mistake? Click undo, approve the reversal, the runner rolls it back. The action becomes a transaction. → MVP connectors: HTTP proxy with domain allowlists, SSRF protection and DNS pinning against rebinding; SSH with per-command policy rules (glob or regex) and optional per-session WireGuard VPN routing; an MCP server for use inside Claude Desktop and other MCP clients. Everything lands in an append-only audit log enforced at the database level via triggers — even admins can't tamper with history. Stack note for the curious: Next.js 15 on Vercel handles the gateway and non-VPN connectors inline; a separate Fastify runner handles WireGuard tunnels and long-lived SSH sessions (outbound-only, no inbound port). Postgres + Realtime via Supabase, RLS on every table. Node and Python SDKs. Would love feedback — especially from anyone building agents in production: What's missing for your setup? Which connectors should ship next? Is anything in the trust model unclear? Happy to answer everything in the comments today.

getviktor.com — An AI coworker that actually does the work

An AI coworker that actually does the work

Promoted

Maker

📌

4d ago

RiteKit Company Logo API

@appiq This is a thoughtful approach to a real problem. The distinction you're making between enterprise-heavy solutions and what developers actually need right now is important — most teams shipping agents today don't have the infrastructure for complex compliance workflows. The push notification approval model seems like it'd catch the 80% case without the overhead.

2d ago

Agentlock

Your AI agent asks. You decide. From your phone.

Your AI agent asks. You decide. From your phone.