Kodesage Raises 6.6M to Fix Legacy Code AI Cannot Reach

The hottest tools in software, from GitHub Copilot to Cursor to Devin, are brilliant at writing new code. They are nearly useless against the code that actually runs the world: decades-old banking, insurance, and government systems that no single living engineer fully understands. A Budapest startup called Kodesage just raised $6.6 million to attack exactly the part of software AI has been quietly avoiding.

What Actually Happened

Kodesage, founded in 2024, closed a $6.6 million seed round, roughly 5.67 million euros, to build an AI platform that helps large organizations understand, document, maintain, and modernize complex legacy software. The round was led by VentureFriends, with participation from Portfolion, which had backed the company at pre-seed. The capital is earmarked for go-to-market expansion across the United States and Europe, plus engineering and product hiring as the company moves from prototype to deployed enterprise tool.

The cap table is unusually pointed for a seed. Angel investors include Christian Szegedy, a co-founder of xAI and one of the most cited researchers in deep learning, and Mario Gotze, the footballer who scored the 2014 World Cup winning goal. Kodesage was started by Gergely Dombi, Miklos Szurdi, and Gyorgy Szilagyi, a founding team betting that the next frontier for applied AI is not generating fresh code but decoding the tangled systems enterprises already depend on and cannot afford to break.

The product itself connects an organization's codebase with its scattered context: documentation, tickets, databases, and wikis. From that, it delivers natural language question-and-answer over the code, automated documentation, and visual architecture diagrams, all through secure on-premise or virtual private cloud deployments. That last detail is the whole game. Kodesage is built so that sensitive source code never leaves the customer's own environment, which is the precondition for selling to the banks, insurers, and agencies that own the messiest and most valuable legacy estates.

Why This Matters More Than People Think

The AI coding boom has a blind spot shaped exactly like the enterprise. Tools like Copilot and Cursor shine on greenfield projects and modern stacks, where code is clean, documented, and safe to send to a cloud model. The trillions of lines that run payroll, settle trades, and process insurance claims are the opposite: undocumented, interdependent, and legally barred from leaving the building. Kodesage is wagering that the real prize was never code generation at all. It is comprehension, the unglamorous work of figuring out what forty years of accumulated software actually does.

The market math explains the urgency. Legacy modernization is estimated at $29.39 billion in 2026 and projected to reach $66.21 billion by 2031, a compound annual growth rate near 17.64 percent, according to Mordor Intelligence. That spend exists because the people who wrote these systems are retiring, taking irreplaceable knowledge with them, while the systems themselves grow more brittle and more expensive to touch every year. A tool that turns an opaque monolith into something a current engineer can query in plain language attacks a cost center every large enterprise already funds.

The geography is a tell about where this market is heading. A serious legacy-AI company emerging from Budapest, not San Francisco, reflects that Europe sits on a deep base of regulated, on-premise enterprises in banking and government, precisely the customers cloud-first American tools struggle to serve. European founders live closer to the data-sovereignty rules that define the opportunity, and a product engineered for those constraints travels well into US banks and agencies facing the same audit regimes. The non-obvious advantage is empathy for the buyer compliance reality, which is hard to retrofit onto a cloud-native starting point built for startups that never had these limits.

There is a strategic reason this matters for the entire agentic-AI thesis. Autonomous coding agents are only as good as their understanding of the system they operate on, and on legacy estates that understanding is the binding constraint. An agent let loose on a forty-year-old core banking platform without a map will cause damage, not progress. By building the comprehension and context layer first, Kodesage is positioning itself underneath the coming wave of enterprise coding agents, as the system of record for what the code means rather than just another tool that writes more of it.

The Competitive Landscape

Kodesage is small and the field is crowding fast. Sourcegraph has spent years on code search and understanding and is pushing into AI-assisted comprehension. Moderne, built on the OpenRewrite project, automates large-scale code transformation. Mechanical Orchard, founded by Pivotal veteran Rob Mee, raised heavily to rewrite legacy systems with AI. And the giants loom: IBM markets watsonx Code Assistant for translating COBOL on the mainframe, while Microsoft, Amazon Q, and Cognition all want a share of enterprise code work. A $6.6 million seed is a slingshot against several armories.

The differentiator Kodesage is leaning on is deployment posture plus breadth of context. Many rivals either focus on transformation rather than understanding, or assume a cloud-friendly customer. Kodesage's pitch is comprehension across the full mess, code plus tickets plus wikis plus databases, delivered air-gapped or in a private cloud so that a regulated bank's compliance team can actually approve it. In enterprise software, the vendor that clears security review first often wins the account, and on-premise readiness is a moat that pure cloud tools cannot quickly cross.

Scale of context is the second axis of competition, and it is subtler than raw model quality. Understanding a legacy system is not just reading source files; it is reconciling code with the tickets that explain why a strange workaround exists, the wiki page a departed engineer once wrote, and the database schema the logic secretly depends on. A tool that ingests only code produces confident but shallow answers. Kodesage is betting that stitching these fragmented sources into one queryable model is the harder, stickier problem, and that whoever assembles the most complete picture of a system becomes the default place engineers go to ask what anything actually does.

The historical parallel is the Y2K remediation industry of the late 1990s, when a generation of consultancies built fortunes combing through COBOL to fix two-digit dates. That wave proved two things that apply directly today: enterprises will pay enormous sums to de-risk legacy systems, and the work is ultimately about comprehension under deadline, not elegance. The difference in 2026 is that AI can read and explain code at a speed no army of contractors could match, which is why a three-person Budapest team can credibly target a problem that once required thousands of billable consultants.

Hidden Insight: The Most Valuable AI Coding Market Is the One Models Are Not Allowed to See

Here is the angle almost everyone misses. The AI coding narrative is dominated by how much new software models can produce, but the largest and most defensible enterprise market is defined by a constraint, not a capability. The code that matters most, in finance, defense, healthcare, and government, legally cannot be shipped to a frontier model's cloud. That single fact carves out a protected market where the flashy consumer tools simply cannot compete, and where the winner is whoever can run capable models inside the customer's walls. Kodesage is building for that constraint from day one.

This inverts the usual moat logic of AI. In consumer and startup tooling, distribution and model quality win, and a better foundation model can erase a competitor overnight. In regulated legacy, the moat is the boring, expensive work of secure deployment, compliance certification, and integration with systems that predate the internet. Those are exactly the advantages that compound slowly and resist disruption, because no amount of model progress shortcuts a bank's procurement and security review. The startup that does the unglamorous integration work early earns durable lock-in.

There is also a deeper shift in what AI is actually being sold to do inside enterprises. The first wave promised to write code faster. The more valuable wave, the one Kodesage targets, promises to recover lost institutional knowledge before it walks out the door with retiring engineers. Comprehension, documentation, and a queryable map of a system are insurance against operational risk, which is a board-level concern, not a developer-productivity nice-to-have. Selling risk reduction to executives is a fundamentally stronger motion than selling autocomplete to programmers.

The timing also rides a demographic cliff that makes the pitch urgent rather than speculative. The engineers who built the mainframe systems still running global finance are retiring in large numbers, and universities stopped teaching their languages decades ago. Every year of delay raises both the price of understanding these systems and the danger of a failure no one left can diagnose. That converts Kodesage from a productivity upgrade into something closer to a continuity plan, and continuity plans get funded in downturns precisely when discretionary developer tools get cut. A product framed as protection against an unfixable outage sells in the exact budget cycles that starve nice-to-have software.

The bear case, however, is real and worth stating plainly. Critics argue that on-premise AI is a punishing business: hard to deploy, hard to support, and slow to scale compared to a clean cloud product, which can crush margins for a company with only $6.6 million in the bank. Skeptics point out that LLMs still hallucinate when reasoning about gnarly, interdependent code, and a confidently wrong explanation of a core banking system is worse than no explanation. The risk is also that the moat erodes from above: as on-premise open models like Llama and Nemotron mature, IBM, Microsoft, or Sourcegraph could bundle comparable comprehension into deals they already own, leaving a tiny seed-stage startup to evangelize a market the incumbents then harvest.

Pull the threads together and Kodesage is a focused bet that the AI software market is misweighted. The attention and capital have rushed toward generating code, while the larger, stickier, better-defended opportunity is helping enterprises understand the code they already cannot live without. It is a less exciting story than autonomous agents writing apps from a prompt, but it maps onto budgets that already exist and risks executives already lose sleep over. The open question is whether a three-person team can out-execute deep-pocketed incumbents long enough to turn that insight into a category before the giants notice the same gap.

What to Watch Next

Over the next 30 days, watch for named design partners or reference customers, ideally in banking, insurance, or the public sector. In enterprise software, a single credible logo in a regulated vertical does more than any feature announcement, because it proves the security and compliance story holds up under a real buyer's scrutiny. Silence on customers would suggest the round is funding a promising prototype rather than validated demand.

Over the next 90 days, track Kodesage's US go-to-market build-out and any published accuracy data. The hard question for any code-comprehension tool is how often it is right about systems no human fully understands, so look for benchmarks, case studies, or third-party validation of how faithfully it explains real legacy code. Watch the hiring too: enterprise sales leaders and solutions engineers signal a company preparing to sell into procurement, while a purely research-heavy team signals it is still proving the science.

By the 180-day mark, the decisive signals are competitive and commercial. Watch whether Sourcegraph, IBM, or Microsoft responds with on-premise comprehension features, which would validate the market while squeezing the category. Watch retention and expansion inside early accounts, because land-and-expand is the only path that justifies an on-premise model's cost. And watch for an early Series A, since investors funding modernization will move quickly if Kodesage shows that comprehension, not generation, is where enterprise budgets actually unlock.

One quieter metric will reveal more than any headline: how many of Kodesage early deployments move from a single team to an enterprise-wide rollout. On-premise enterprise software rarely fails at the pilot; it fails at the expansion, where support burden and integration edge cases multiply. If a bank that started with one legacy system rolls Kodesage across its full estate within two quarters, the model is working. If pilots stall as isolated experiments, the on-premise thesis is harder than the pitch deck admits, and the funding runway gets tight fast.

The most valuable code on earth is the code no AI is allowed to see. Whoever can read it inside the bank's own walls owns a market the consumer tools can never touch.

---