Legacy app refactoring with AI: software modernisation through reverse engineering

Many mid-sized companies run core processes on applications in Delphi, Visual Basic 6, Access, PowerBuilder, .NET Framework or Java 8 – some for more than 20 years. The people who built them are retiring, documentation is patchy, and operating systems and runtimes are falling out of support.

Until now, modernisation often failed on the effort of understanding: before a single line can be rewritten, it must be clear what the old system actually does. That is exactly the part AI agents speed up most – they read code at scale, explain it and generate tests.

The split of work between agent and team is the most important success factor.

Reverse Engineer Anything (REA) is an open-source project under the MIT licence. Through an MCP server it connects coding agents such as Claude Code or Codex to established reverse engineering tools like Ghidra, IDA and Hopper, plus browser and runtime analysis. An agent can then examine native programs, JavaScript and Electron apps or .NET assemblies without knowing the source code.

The repository was created in April 2026 and within a few weeks in October 2026 became one of the most-watched projects on GitHub, with more than 400,000 stars. Perplexity added REA to its Computer agent as the /reverse-engineer skill. A widely cited example from the project: a sound function from the game DX-Ball was reconstructed so that it passes all 3,205 original test cases.

Important context: REA does not recover the original source code. The output is a reconstruction of the logic with evidence and a list of open questions. For modernisation that is exactly what is valuable – but it does not replace a business review.

For companies the value lies less in cloning other people's apps than in recovering knowledge about their own. Typical cases: the source code is lost or differs from the production version; custom software came from a vendor that no longer exists; interfaces exist only as compiled libraries; file formats are documented nowhere.

Three perspectives complement each other: static analysis of code or binaries, dynamic analysis of the running program including database queries and network traffic, and a black-box view of inputs and outputs from real cases. The result is a specification of business rules in which every statement points to evidence.

Legacy refactoring and replacing SaaS tools share the same core: derive a specification from existing behaviour, rebuild with coding agents, prove equivalence with tests. The difference is the source of knowledge.

How a legacy application gradually becomes a modern solution without interrupting operations.

Assumption: an internal warehouse management application with around 150,000 lines of Delphi code and stored procedures; one of the two developers who know the system will retire soon. We compare a classic rewrite with an AI-assisted one. The figures are assumptions for an order of magnitude, not measurements.

For your own software with your own usage rights, reverse engineering is unproblematic. For licensed software, Section 69d(3) of the German Copyright Act allows authorised users to observe, study and test how a program functions. Decompilation is only permitted to a limited extent under Section 69e, for example to achieve interoperability with an independently created program. The German Trade Secrets Act (Section 3 GeschGehG) allows examining lawfully acquired products unless this has been effectively excluded by contract.

Check licence and maintenance contracts before analysing third-party components. This is not legal advice.

MAKONIS assesses your legacy applications in a modernisation assessment, sets up isolated analysis environments with sovereignly operated models and modernises with AI software engineering – context engineering, review gates and tests included.

The target is usually a cloud-native application in your Azure landing zone; on request we plan the cloud/Azure migration at a fixed price. Start with a free 30-minute call with our engineers.