The code intelligence platform that turns ten million lines of legacy into something you can actually read on a Tuesday morning.
Photographed in browser tabs across San Francisco, Oakland, and roughly 190 living rooms - Sourcegraph hasn't had a real office since the meeting room became optional.
Open a browser tab inside Uber, Lyft, Cloudflare, or PayPal and there's a fair chance you'll find a Sourcegraph window already loaded. It is, depending on who you ask, the company's grep, the company's compass, or the company's last line of defense against the chaos of a ten-year-old monorepo.
None of those descriptions are wrong. Sourcegraph indexes every line of code an organization owns - across GitHub, GitLab, Bitbucket, that one server nobody is allowed to turn off - and lets engineers search, navigate, and refactor it like a single document. On top of that index sits Cody, an AI assistant that, unlike most of its peers, has actually been introduced to the code it's asked about.
It is, in other words, a search engine. Just one that decided the world's most valuable text was not on the web.
The dirty secret of software engineering is that almost no one writes code from scratch. Engineers read. They read other people's code, their own code from six months ago, and the occasional comment that says // TODO: explain this later from someone who has since left the company.
By 2013, the tools for reading code had not meaningfully improved since the 1990s. grep -r. ctags. A right-click menu in your IDE that worked, statistically, about half the time. If your company had more than one repository - and by 2013 most of them did - the tools simply gave up.
Quinn Slack and Beyang Liu, two Stanford grads who had done internships at places like Google and Palantir, noticed the same thing every developer noticed. They just refused to accept it as load-bearing.
The wager was structural. If a codebase was treated not as a pile of files but as a graph - functions calling functions, files importing files, repos depending on repos - then almost every developer chore became a graph query. Find usages? Walk the edges. Refactor? Rewrite the node and patch its neighbors. Onboard a new hire? Hand them the map.
The wager looked obvious in retrospect. At the time it looked like building Google for a market that had four customers.
What made the bet survivable was the team's willingness to be patient. Sourcegraph was founded in 2013. It did not announce a Series A until 2018. In between, the founders did what founders are not supposed to do: they used the product themselves, all day, every day, and got irritated by it until it was good.
Caption: Two Stanford grads, one whiteboard, and a stubborn refusal to accept that grep -r was the final form of search.
Quinn Slack and Beyang Liu start Sourcegraph in San Francisco.
Series A. Universal Code Search ships to its first big enterprise customers.
Series B and C close back-to-back. Batch Changes launches.
Series D at a $2.625B valuation. The company is now officially a unicorn.
Cody debuts - an AI assistant with full-repo context, not just file context.
Core repo goes closed-source. Engineering focus shifts hard to Cody and enterprise.
Agentic workflows and codebase-aware autocomplete become the headline feature.
Sourcegraph's product surface looks broad until you notice every piece is doing the same job: shrinking a codebase down to something a human can hold in their head.
Universal search across every repository, branch, and host - regex, structural, and semantic.
An AI assistant with deep codebase context. Chat, autocomplete, and inline edits, grounded in your repo.
Author a change once. Push it across ten thousand repos. Watch the PRs roll in.
Dashboards that track migrations, deprecations, and engineering health over time.
Precise go-to-definition and references that don't pretend the language ended at the file boundary.
Caption: Five products, photographed in their natural habitat - a developer's third monitor, the one with all the tabs.
If you're going to claim to be the search engine for code, the question becomes uncomfortable quickly: who actually uses it, and how much have they paid? The answers are, in order: a lot of people, and a lot of money.
Source: company disclosures, Crunchbase. Bars are scaled to the largest round.
Customers include Uber, Lyft, Cloudflare, PayPal, Plaid, GE, Indeed, and - according to the company - more than half of the world's ten largest banks. None of these are companies that pick a code-search tool for fun.
It's a strange mission statement for a company that sells to enterprises. But it's the one Sourcegraph has stuck with. The argument runs like this: if reading code is the hard part of building software, and reading code can be made dramatically easier, then the population of people who can meaningfully participate in software grows.
That population already includes engineers at the world's largest banks and rideshare companies. With Cody, the company is betting it will expand to anyone willing to type a question in natural language - and willing to verify the answer before shipping it.
Sourcegraph remains fully remote, with a famously public handbook that's been copied as a template by dozens of other startups. The company's transparency about its strategy - including the 2024 decision to close-source the core product - is unusual for a venture-backed company at this stage. It's also entirely on-brand for a team that believes the best way to be trusted is to be legible.
Caption: A remote-first company writing its strategy in public, possibly because writing it in private didn't make it any better.
Every year, large language models get more impressive at writing code. Every year, the world's codebases get larger, weirder, and more interconnected. The gap between what a model can generate and what a model can be trusted with - inside a real codebase, with real customers - is filled by context.
Context is the boring part of AI. Indexing repos. Resolving symbols. Knowing which version of which function the user actually means. Sourcegraph spent a decade building exactly this kind of unglamorous infrastructure before anyone called it context engineering. Now it's the substrate underneath their AI bet.
If the next ten years of software look anything like the last two, the question facing every engineering org is the same: how does an AI agent ship a change inside a codebase no single human understands? Sourcegraph's answer is the only one they've ever given. Make the codebase smaller. Not in lines. In legibility.
Closing the loop. Open the browser tab again. The one inside Uber, or Lyft, or Cloudflare. The Sourcegraph window is still there. But now there's a chat panel on the right, and an engineer is asking the codebase a question in plain English, and getting an answer that references the actual code. The query took half a second. The reading it would have replaced would have taken half a day. That's the change. That's the bet, paying off in installments.
Official site, social, code, and a few worthwhile rabbit holes.