Anyone who's let an AI agent write code has seen the trick fail in the same quiet way. The output looks right. The syntax is clean, the function names make sense, the logic reads like something a competent engineer would ship. Then you run it, and it breaks, because the model confidently called a library method that doesn't exist in the version you're using. GitHits, a Helsinki startup, just raised 1.5 million euros to attack exactly that failure.

The pre-seed round was led by Vendep Capital, with Trind VC and a notable group of angels joining: Peter Sarlin, Zach Shelby, and Jerry Liu. The company spun out of Softlandia Venture Studio in late 2025, and it's launching the beta of its product today, with a commercial version planned for later this year.

The pitch is narrow on purpose. GitHits is building what it calls a version-aware index of all public open-source code, a search layer designed not for humans but for AI coding agents. The goal is to ground those agents in real implementations of real libraries, so they stop guessing how a dependency behaves and start checking. Less hallucination, fewer retry loops, working code on the first pass instead of the third.

To understand why this matters, you have to understand where coding agents actually fall down. "Coding agents are great at navigating your local codebase," said CTO and co-founder Olli-Pekka Heinisuo. "The problem is that modern software doesn't stop at the repository boundary. A large part of the system lives in frameworks, libraries, SDKs, and other open-source dependencies. Agents can't inspect those nearly as well, so AI has to guess."

That's the crux. Your own code is right there in front of the model. The vast machinery your code depends on, the thousands of open-source packages, each with its own versions, quirks, and breaking changes, is largely opaque to it. So the agent does what models do when they lack the facts. It fills the gap with something plausible. Plausible is not the same as correct, and in software the difference shows up the second you hit run.

GitHits gives agents tools to find working examples of open-source implementations and to inspect software components, including their dependencies and known vulnerabilities. To make that possible, it's building an AI-native, version-aware index of public code. Version-aware is the key phrase. The same library can behave completely differently across releases, and an answer that's right for version three is a bug in version five.

The team is the reason investors leaned in, and one credential stands out. Heinisuo previously created opencv-python, a software package with more than 100 million downloads. It was used by NASA in the Ingenuity helicopter that flew on Mars. When the person building your open-source code index has himself shipped open-source infrastructure that left the planet, the credibility problem mostly solves itself.

"Olli-Pekka is a quiet legend in open source and has lived inside this problem for years," said Timo Felin, Partner at Vendep Capital. "At this stage you invest in people, and this was an easy call." The idea itself was born from frustration. While working at AI consultancy Softlandia, Heinisuo kept giving colleagues the same manual tip for finding open-source information, and eventually realized the whole thing could be automated.

His colleague Jaakko Timonen, now CEO, got excited, and the two spun out a company with Softlandia's backing, assembling a team of four experienced co-founders. The venture-studio origin matters here. GitHits didn't start in a dorm room. It started inside a company that builds AI systems for a living and knew the pain firsthand, which tends to produce sharper early products than pure speculation does.

The conventional wisdom says you should never build a feature the big AI labs might absorb. GitHits is making the opposite wager, and there's a real argument behind it. Code isn't ordinary text. It has versions, dependency trees, and behavior that changes release to release, structure that a general-purpose search engine flattens and loses. The deeper that structure runs, the more room there is for a specialist to do something a generalist simply can't replicate by bolting on a code filter.

That's the bet against Exa, the well-funded generalist building search for all AI agents. GitHits isn't trying to out-raise it or out-scale it. It's trying to out-specialize it on the one surface where specialization pays: making sure an agent knows precisely how version 5 of a library behaves, not how version 3 did, not how the model imagines it might. In a world where a single wrong method call breaks the build, that precision is the product. Generalists struggle to match it because the long tail of versioned behavior is exactly what gets averaged away at scale.

There's a token-economics angle that could matter more than anyone expects. Every retry loop, every time an agent generates broken code and tries again, burns compute and money. If GitHits can ground agents well enough to cut those loops, it isn't just improving quality. It's lowering the running cost of every team that wires it in. Sell measurable savings to engineering leaders watching their AI bills climb and you've got a far easier conversation than selling a nice-to-have. Painkillers move budgets. Vitamins don't.

There's an obvious worry with any developer-tools startup in 2026: what happens when OpenAI, Anthropic, or Google decides to build your feature? GitHits has a clear answer. It isn't trying to be a coding agent. "GitHits doesn't compete with Codex, Claude Code or Cursor, but complements them," Heinisuo said. The product brings open-source code in as context for those agents, ending retry loops and cutting token consumption.

That positioning is shrewd. Instead of fighting the biggest, best-funded companies in the world head-on, GitHits wants to be the layer that makes their agents work better. The big labs build the brains. GitHits supplies a specific kind of memory, grounded knowledge of how the open-source world actually behaves. Picks-and-shovels, sold to the people selling the gold rush.

The competition GitHits does acknowledge sits one layer over: US-based Exa, which raised a 250 million dollar Series C in May at a 2.2 billion dollar valuation to build search for AI agents. Heinisuo draws a clean line. "Exa is building a general-purpose search for AI. GitHits focuses only on code." Whether narrow focus beats a well-funded generalist is one of the oldest questions in software, and it usually comes down to how deep the specialized problem really runs.

Everything about GitHits rides on one assumption: that searching code for AI agents is a hard enough, distinct enough problem to support a standalone company rather than a feature swallowed by a general search engine or a coding tool. Code has structure, versioning, and dependency graphs that ordinary text doesn't, which is the argument for specialization. It's the same instinct driving a wave of Nordic AI-infrastructure bets this year, founders going narrow and deep while the giants go wide.

The timing has a certain logic to it. As AI-generated code moves from novelty to default, the cost of agents that confidently produce broken output stops being an annoyance and becomes a real tax on engineering teams. If GitHits can measurably cut that, by reducing retries and grounding outputs in code that actually exists, it's selling a painkiller, not a vitamin. Teams pay for painkillers.

A 1.5 million euro pre-seed buys time to prove the index works and that developers will wire it into their agents. The beta launching today is the first real test. The vision, indexing all public open-source code so AI never has to guess again, is enormous. The near-term question is humbler and more important. Does the beta make agents demonstrably better? If it does, the rest is just scale.