Tailscale Builds 'Aperture', a Customer-Facing Model Router Based on Vercel AI Gateway

Tailscale built Aperture, a solution for launching hundreds of AI models within customer products, based on Vercel AI Gateway and Vercel Sandbox. As an existing infrastructure company, Tailscale initially attempted to develop its own routing layer but adopted Vercel's solutions due to engineering complexity and security issues.

Background on Choosing Vercel and the Tech Stack

Tailscale controls model access permissions via tailnet network IDs instead of individual API keys. AI Gateway provides a single API for hundreds of models and returns cost and usage data for every request, offering customers transparent visibility into their spending. Sandbox provides an isolated environment for agent execution, addressing the security risk of the 'lethal trifecta' where private data and public internet access are combined.

Zero Data Retention and Cost Structure

To support customers who prioritize data protection, Zero Data Retention (ZDR) is supported. AI Gateway has no data retention of its own, and routing can be restricted to ZDR-compliant providers only via global settings or per-request flags. Regarding costs, there is no markup on token costs for any provider or model, and BYOK (Bring Your Own Key) is also supported.

Internal Migration and Future Goals

Tailscale also migrated its internal AI layer to AI Gateway, consolidating the management of dozens of provider accounts. This process was completed without downtime and serves as a playbook for enterprise customers. Currently, Aperture supports Chat UI and MCP connectors, with a future goal of achieving a 'Time to first app' of under 10 minutes, from sign-up to building and sharing an app.