Last year, the project that I led (Vesuvius Challenge) achieved a breakthrough in resurrecting an ancient library from the ashes of a volcano (and they’re still going). This success taught me an important lesson: an extremely fast-moving community can achieve incredible things. This seems obvious, but few do it! Since then I’ve been thinking about how to apply this lesson to other spaces, such as helping my friend Steve at Val Town prototype an AI agent.
Since then I’ve been thinking a lot about AI agents. They’re the closest I’ve seen to the dream of “programming for all”: all humans shaping our digital worlds (“moldable as clay”). This is something that has long interested me, and I think its time has come. Gone are the days of thinking of code directly—AI agents will produce code “by the kilo”. Enter: Kilo Code.
There are already a million AI agent companies out there, so why another one? Me and my team are an impatient bunch, and we want to quickly apply this idea of an extremely fast-moving community to AI agents. We’ve been living this extreme focus on speed already:
Two weeks ago, I dropped everything in my life to build the first version.
A week later we had assembled a team of 10 (full time).
We then forked the Roo Code VSCode extension (which itself is a fork of Cline).
Since then, we shipped a bunch of small improvements (more about that below).
By Friday, we’ll have fixed two of your requests.
Next week, we’ll do ~10 more improvements.
Most importantly, we want to build this with you. We would love your continuous feedback (on Github and Discord), and we have a free tier (with $15 in free tokens a month!) in order to thank you—and for the best feedback we’ll give out more free tokens. The goal is to create the most user-friendly AI coding agent, as quickly as possible.
Meet the team
The team came together within a week, which means that we’ve self-selected for folks who value speed, and are down for adventure. Here are a few of us:
I’m leading the team (JP Posma).
Justin Halsall leads developer relations. He has worked on an AI company for the last 5 years, is the open source maintainer of rrweb, and an Apache contributor.
Johan Otten is our lead developer. After building products for teaching kids to code, he’s now excited to better teach LLMs to code.
In addition, we’re well-funded, and hiring. (If you’re interested mail us at hi@kilocode.ai. We’re remote-first, but with offices in SF and Amsterdam. Expect to work really hard and have a ton of fun.)
Small fixes shipped
We’re serious about speed! In a few days we’ve plucked some low-hanging fruit:
No need to create an OpenRouter account (use our own proxy)
No need to pay and use a credit card (batteries included free tier using Claude 3.7 Sonnet)
No need to select a model (good defaults)
No need for telemetry acceptance (we send what we need to improve the product and are transparent about what we send and how we use it)
No need to set auto acceptance (vibe coding—but you can turn it off)
Support for DeepSeek in Fireworks (super fast)
Multiple onboarding improvements (after doing user testing)
Next week we’ll similarly focus on low-hanging fruit. Tell us what you want to see, and we’ll give out more free tokens for the best ideas!
Secret Master Plan
Besides quickly building with the community, we do need some rough direction of where to take things. We started with an AI agent (as opposed to a Copilot/Cursor-style UI), because we want to build the future we want to see, and which we think is coming. We want to build for the dream of billions of programmers; billions of artists; billions of scientists—using computing as moldable clay.
However, right now AI agents only work well for smaller projects, and for novice/non-programmers. But we think we can quickly make them better, and move upmarket to more advanced use cases (Clay Christensen “Innovator’s Dilemma”-style).
For this, we’ll be part of the larger community of AI coding companies. Hopefully we can bring meaningful new ideas to the space, as well as quickly implement the best ideas that we see from others. Since our extension is open source, we encourage everyone (e.g. our upstream Cline and Roo Code) to use our best code and ideas.
We have some ideas already, many of which are already being pioneered by other tools, and some which are more out there:
Instant app - Type an idea, press Enter (or Google-style “I’m feeling lucky!”), get a live app with a backend and database.
Up-to-date docs - Crawl docs automatically, so whenever you need to find the latest version of your favorite tool, everything is up to date.
Prompt/product-first workflows - Not even looking at any code; instead looking at the product in a browser, together with an agent.
Browser IDE - A full development environment in the browser. Sandboxed, secure, no setup.
Local/on-prem models - For extra speed, security, and privacy.
Live collaboration - Share a link and work with your team in real time, prompting together.
Parallel-agents - Run many agents in the same app in separate threads.
Code variants - Create multiple code implementations from one prompt. Pick the best.
Shared context - Reduce API calls by storing project knowledge across the team.
Open source sharing - Allow snippets of your code to be shared as open source, and get used by others who are trying to do the same task (generalized caching).
MCP marketplace - Making it easier to extend your agents.
Integrated CI - Agent writes and runs tests, fixes things until all pass.
Monitoring/production agents - Agent sees runtime stats and errors, and can fix or escalate.
Security agents - Pentesting by agent, fixing by another agent.
Sketching - draw the UIs or data architecture you want, or show with video recordings what to change.
Generate kilos of code for free in VS Code
Kilo Code is live in VS Code. Try it out and generate kilos of code for free. Sign up with your Google Account and get $15 worth in free Claude 3.7 usage without needing a credit card. And tell us what needs to improve, on Github and Discord. We’re building fast—your feedback shapes what comes next.
will there be a version for neovim?