We raised funding from Lollapalooza Capital and Impa Ventures to rebuild tools and infrastructure for people and agents

Today, we're announcing a multi-million-dollar funding round for Polygala, Inc., the company behind BoxLite. The round is backed by Lollapalooza Capital and Impa Ventures.

Thank you to everyone who has tried BoxLite, contributed code, reported a problem, or shared what they want their agents to do. And thank you to our team and investors for helping us build BoxLite.

Why we started

Before BoxLite, I spent years building open-source databases: the storage engine under TiKV, then data lake and cloud storage support in StarRocks. Both assumed a human at the controls, someone to read the docs, tune the settings, and step in when things broke.

Then the progress of large language models convinced me that we were entering the era of agents. I could see a future in which people would ask software to carry out work on their behalf, with agents writing code, using tools, and working through tasks with several steps.

By early 2025, that direction was becoming tangible. Anthropic introduced Claude Code in February, and OpenAI released new tools for building agents in March.

Agents would need somewhere dependable to run software, save files, and return to a task. We started building BoxLite to give them that place.

I wrote the first lines of BoxLite in May 2025 and open-sourced it in December 2025. We began with a runtime that gives agents isolated environments in which to run software.

Our mission is to rebuild tools and infrastructure for people and agents. We want anyone to be able to put the power of the cloud to work through an agent.

The experience we're building toward is simple: people ask, agents do the work, and cloud complexity stays behind the scenes.

The foundation we're building on

BoxLite is open source, from the runtime on your laptop to a cloud stack you can run in your own AWS or GCP account. It's also available as BoxLite Cloud, our managed service.

The runtime gives each agent a box: a micro-VM with its own kernel that keeps its files between turns. It runs as a library on your laptop, and the same engine powers BoxLite Cloud.

Today, you can build an app with your coding agent, and BoxLite Cloud runs it and gives you a public URL to share.

Our open-source community has helped us move that work forward. In less than a year as open source, BoxLite has earned about 2,500 GitHub stars. It has been downloaded more than 500,000 times across PyPI, npm, crates.io, and GitHub, and 40 people have contributed code to the project.

BoxLite has also been integrated into open-source agent projects from Databricks, Alibaba, and ByteDance:

• Omnigent, from Databricks' AI team, offers BoxLite as a sandbox option for running agents.
• AgentScope Runtime, from Alibaba, supports BoxLite as a local sandbox backend for tool execution.
• DeerFlow, from ByteDance, includes a BoxLite sandbox provider for agent workloads.

These integrations connect BoxLite to the agents people use. Keeping the runtime open source lets the people building those agents inspect it, improve it, and make it part of their own applications.

Why we raised

We raised this round to grow the team around two priorities.

First, open source. We'll keep investing in both halves of the open-source project: the local runtime and the cloud stack.

Second, BoxLite Cloud. We'll grow our managed service and make local and cloud feel the same: what works on your laptop works the same way in the cloud.

Open source comes first because we believe in democratizing agent infrastructure. Anyone should be able to run it, inspect it, and own it, on a laptop or in the cloud.

Lollapalooza Capital is a fund founded by Meituan co-founder Huiwen Wang. Impa Ventures is a Singapore fund that invests in frontier technologies powering AI at scale.

What comes next

We created BoxLite Cloud to give people a managed BoxLite experience. We want to expand it so everyone can have a micro cloud of their own, operated through an agent, with the complexity of traditional cloud kept behind the scenes.

Imagine asking your agent to build a website for your small business, organize a collection of documents, or prepare a report every week. You describe what you want. The agent gets the resources it needs, does the work, and brings back something you can use.

People should be able to see what their agent is doing, decide what it can access, and change its direction.

Our roadmap focuses on helping people:

• Start work and take it with them. Boxes will come with an agent ready to work, carrying their memory and files between a laptop and the cloud.
• Turn an idea into an app people can use. Sign-in, a database, storage, and a web address will come built in, so the agent can handle more of the setup.
• Take on bigger tasks. Resources will scale with the work, including GPUs for running their own models.
• Connect services while protecting their credentials. Agents will be able to use credentials without seeing the secrets themselves.

We plan to rewrite the entire BoxLite product in Rust. We believe Rust is the language of the AI era: as agents write more software, strong checks built into the language become more valuable. Rust's approach to types, ownership, and concurrency is part of how we aim to make the infrastructure underneath that work more dependable.

One part of this effort is already public: we're building our own virtual machine monitor. It gives us more control over how BoxLite runs agent workloads.

We're hiring software engineers and teammates in go-to-market, product, and developer relations. If you want to help build BoxLite, email founder@boxlite.ai.

You can explore the open-source project, try BoxLite Cloud, or join the community. Tell us what you'd like your agent to do for you.