Tinyfish builds enterprise-grade AI web agents. Their customers need to run hundreds of thousands of browser sessions in short time windows, and when working with enterprises, the conversation always comes back to security. Tinyfish uses Unikraft to run each browser as a fully isolated microVM, meeting enterprise security requirements without having to build the infrastructure themselves.

The Opportunity

Every component in an enterprise automation pipeline can become a failure point during security reviews. For Tinyfish, browser isolation is one of those critical components. The alternative (process-level isolation) provides weaker guarantees that don’t hold up in enterprise procurement conversations.

Security and isolation are top concerns when you work with enterprises. At some point, the conversation always turns to the same question: What’s the security level of the end-to-end pipeline? Every cog in that chain can become a vulnerability.

— Pasha Dudka, Head of Engineering, Tinyfish

The Solution

Each Tinyfish browser instance runs inside its own Unikraft microVM, consisting of a headless browser and a custom proxy application that regulates access and ensures confidentiality. With optimizations, each instance cold starts in approximately 10 milliseconds. Further, the platform supports large scale, and is set up to have tens of thousands, scaled-to-zero, ready-to-go browsers that can be spun up in a matter of milliseconds Similar to Docker-based workflows, each instance can be customized via environment variables, creating unique variants from the same base image.

Production runs on bare metal servers for maximum performance, with EC2 instances available for faster node bring-up when capacity needs to expand. AWS load balancing distributes traffic dynamically across nodes.

Each browser is a microVM, fully isolated at the hardware level. That’s something that enterprise customers value a lot.

— Pasha Dudka, Head of Engineering, Tinyfish

Tinyfish’s browser images are relatively large with almost 1 GB compared to the few megabytes common in other deployments. Tinyfish was the customer that drove Unikraft to dramatically improve efficiency for large root filesystems, optimizing load time, snapshot size, and overall memory consumption when many instances share the same base image.

Beyond performance, what mattered equally for Tinyfish was getting a production-ready solution without having to build it. Unikraft’s platform already offered request routing, built-in reverse proxies, and scaling strategies. This is infrastructure they would have otherwise needed to implement themselves.

There were so many great things offered by Unikraft that we didn’t have to build. MicroVMs, fast startup, isolation…we simply package everything as an OCI image and it just spins up — amazing!

— Pasha Dudka, Head of Engineering, Tinyfish

The Outcome

The migration to Unikraft gave Tinyfish enterprise-grade browser infrastructure without having to build the underlying platform:

• Hardware-level isolation out of the box: Each browser session is strongly isolated via its own microVM — the security posture enterprises demand, delivered without custom engineering.

• ~10ms cold starts at scale: Tens of thousands of sessions available at millisecond scale, critical when hundreds of thousands of runs happen in short time windows.

• Zero infrastructure to build: Routing, proxies, scaling, and load balancing came with the platform, freeing Tinyfish to focus entirely on their AI agent product.

• Expanding use cases: Tinyfish is exploring additional Unikraft workloads including development sandboxes and AI coding assistants — workloads that follow the same spin-up, execute, terminate pattern.

Learn how Unikraft can help you build faster infrastructure: Book a call with our CEO Felipe

(He is an engineer, no sales fluff here)