Melt Studio took Agent G Runtime — an open-source egress firewall for AI agents — from a working prototype to an MVP that runs behind corporate VPNs, in five weeks, as part of its pro-bono prototype program, so founder Arjun Padmajan could start testing with design partners.

Agent G Runtime: an open-source firewall for AI agents

Arjun Padmajan is building Agent G, deterministic guardrails for AI agents. Agent G Runtime sits between an agent — Claude Code, Codex, OpenClaw, Hermes, or a custom one — and the internet. It sees every network call the agent makes, blocks destructive actions like deleting a GitHub repo or dropping a database, sends risky actions to a human for approval, and keeps a tamper-evident audit log.

The challenge: an AI agent firewall that had to work behind a VPN

Arjun came to the program with a different product and pivoted to Agent G while in it, with a prototype already built. The deciding question was technical: most companies run a VPN, and a firewall that stops working when the VPN is on can't be tested with real teams. For a solo founder splitting time between code, product, and fundraising, that question was go or no-go.

The solution: VPN coexistence and safe defaults, before anything else

We spent the first week studying the existing Go codebase and the VPN conflict before writing any code, then scoped the MVP around two things:

• VPN coexistence: the firewall runs on macOS and Linux with a VPN active, with Linux brought to parity with macOS.

• Safe defaults: dangerous actions like repository deletion are blocked from the first run, with no configuration.

• Source attribution: every request is tied to the agent and application that made it, shown in plain language in the dashboard.

• A kill switch to pause all agent traffic, or a single agent, at any time.

An enterprise console, Kubernetes support, and Windows capture were left out of month one on purpose. The goal was a product one team could install and trust, not a platform.

How a dedicated team shipped it in five weeks

A developer, a project manager, and our CTO, working with Arjun in daily standups: a discovery week, then four weekly sprints. The last sprints focused on stability and testing, so the MVP would hold up on real machines, not just in demos.

Results: an MVP live with design partners

• VPN-compatible capture on macOS and Linux — the go/no-go requirement for testing with companies.

• An MVP live and in use by design partners, workflow by workflow.

• A working product to show investors as Arjun's fundraise begins.

"Nowadays it's so easy to vibe code anything you want, to have a POC version. But it's a completely different beast if you want to have a production-ready solution." — Arjun Padmajan, Founder, Agent G

Got a prototype that works for you, but not yet for the teams you need to sell to? That's the gap we fill.

Schedule a call today!

Thx for reading. With Love, MS.