Launch HN: Prized (YC S26) – Let non-engineer staff build secure internal tools

Marinos and Hudson just launched Prized on Hacker News with a pitch that hits a familiar nerve: every company has a pile of half-built internal tools trapped in notebooks and spreadsheets because engineering never has time to formalize them. Prized wants to be the thing that turns those scripts into real, deployed apps, without an engineer touching the keyboard and without handing sensitive credentials to whoever's typing the prompt.

The security model is the actual product here. The sandbox where the AI agent works never sees real API keys or connector secrets. It only gets scoped session tokens that act as placeholders, and the real credentials get swapped in later at an egress proxy. When a tool is wired to production data, network access defaults to deny, so the proxy is the only door out. Every outbound call to a connector also gets reviewed by an LLM judge before it's allowed through, which is the founders' answer to the obvious question of what stops an agent from doing something dumb or dangerous with company data.

The example they gave is a good illustration of the whole pitch: a data scientist at one customer pasted in a personal fraud-detection notebook, hardcoded thresholds and all, and after a few prompts it became a published risk console with those thresholds turned into adjustable UI controls. The founders say most of that company is now using it, which they learned on a call earlier the same day they posted this launch.

What makes Prized different from just being another AI app builder is the permissions layer sitting underneath. Admins can scope data access to specific users or teams, and every access gets logged. Each tool runs on its own Postgres schema and role, with queries going through an authenticated SQL gateway rather than open database access. Tools also aren't private artifacts — anyone in a workspace can see, fork, and extend what colleagues have built. One example: a marketing lead built a promotional analytics tool, and a data scientist at the same company forked it to add confidence intervals, turning the workspace into something closer to a shared library than a pile of one-off scripts.

The founders position Prized between Lovable, which is great at generating and hosting software but not built around permissioned distribution, and Retool, which generally assumes a technical builder is making something for someone else. Pricing is straightforward for now — a free tier with two tool builds a month, a $100/month Teams tier, and a custom Enterprise tier that includes things like on-prem deployment. The team says they're still figuring out where the line should sit between giving people freedom to build and keeping unaudited access away from company systems, which is honestly the right thing to still be unsure about at this stage.