I deal with tens of repositories daily, in a company with thousands. Having a clear picture of real-time architecture relies on discipline and the goodwill of engineers to keep the (fragmented) documentation up to date. With coding agents, this problem grows at the speed of light.
LLMs are good at explaining parts of code, but are very bad at extracting precise and reliable architecture mapping of big codebases, not to say when dealing with multiple repos of a microservices fleet (read: distributed monoliths being dressed as microservices architecture).
I’ve built CodeRadius to try to solve this problem (and other adjacent ones) in a more deterministic way, leveraging static analysis combined with local LLMs and a graph database.
repo: https://github.com/coderadius-ai/coderadius
demo: https://coderadius.ai/acme-microservices-demo.html
What CodeRadius tries to do:
- maps dependencies paths across services through APIs, databases, and message brokers
- calculates the blast radius of a code change and detects breaking changes (BC) at a static level, before reaching production
- gives platform teams a governance tool to assess and enforce policies
- empowers your already existing catalogs (e.g. Backstage, Cortex) with the ground truth of the real software architecture
- provides package intelligence that helps tech leads assess the fragmentation of software versions across all their repos, and security teams to have a quick overview over the vulnerabilities of the used packages. It also detects internal packages and calculates the adoption of each release
- exposes all of this via MCP to coding agents
How do you handle cross-repo impact today? How do your platform teams govern the golden paths?