Introducing LGND Code Assistant  

My Walk with LGND Code Assistant

A few days ago, my boss told me to vibe-code something of my choice on top of our embeddings API using our new code assistant, and write a blog post about it. What I actually did was spend ninety minutes on a mountain trail, talking to Claude through a custom voice setup, arguing a spec into existence. The code came later.

You'd reasonably ask why a LGND employee needs a code assistant to build on LGND's own API. The honest answer: the team has been shipping the embeddings API faster than even I can keep track of. That made me close enough to a fresh developer to be a useful test of the onboarding surface.

The code assistant is built on Dosu, which acts as an organizational memory layer for coding agents — continuously ingesting docs, code, and project activity into a searchable knowledge base it exposes through the MCP protocol. That way, when Claude queries our API, it gets answers grounded in the current state rather than guessed at or pulled from stale memory.

I started up the mountain with a vague idea: build a small toy that helps people develop intuition about embedding similarity. By the time I reached the top, Claude and I had argued through what the toy should be, what the API would actually need to do, and how the difficulty modes should work. A big chunk of that conversation was figuring out what the embeddings API actually exposes — my vague design idea didn't translate cleanly to the tools that exist, so we worked through what was there, how the pieces related to what I was trying to build, whether the whole thing was feasible, and how. The code assistant wasn't just there for the implementation phase. It was load-bearing in the design conversation, because Claude could read the actual docs as we talked. I came back down with a real spec.

I meant to dispatch the build to Claude Code from the trail, but I'd forgotten to leave my laptop on. So I waited until I got home, gave the spec to Claude Code, and came back later to an almost-working first pass. The bugs were the interesting part. The loop surfaced a few undocumented API behaviors, an internal API issue that needed fixing, and a small workflow finding about MCP tool discovery. Some got patched live; a couple are still on the backlog.

Hello ChipMatch!

What I built is called ChipMatch. You get a query thumbnail of a satellite chip and six candidates; pick the one with the highest cosine similarity to the query in Clay embedding space. Three difficulty modes, ten rounds, a leaderboard. It's harder than it looks — even for me. The chips are NAIP imagery from California, embedded with Clay v1.5. For the longer story on what the game reveals about how the embedding represents landscapes, the science blog has the deeper dive. 

Three rounds of ChipMatch are free with no signup needed. After that, a free LGND API key keeps you going on your own quota. Play the game here and grab a free API key here.