How I Stole PSI From Android Studio: Letting AI Navigate a Large C…
summary
Shows that text-based grepping by AI returns 2,880 noisy matches for a component when the correct count is below 500 and that LSP solutions spawn a headless IntelliJ using 2.6 GB RAM. Describes building a JetBrains plugin, Kotlin PSI MCP, that exposes IDE PSI over MCP so agents can call find usages, get Android Studio–exact results (≈452 refs), deterministic answers, lower token use and faster responses, plus extra tools: module dependency, call hierarchy, declaration metadata and VRT test-case generation.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:12Talk outline
- 2:16AI code navigation limits
- 7:17LSP as solution
- 11:35JetBrains PSI
- 15:35Kotlin PSI MCP
- 18:29Finding usages
- 22:48Advanced tools
- 26:43Derived workflows
- 28:26Development challenges
- 32:16Current limitations
- 33:18Choosing the right tool for AI code navigation
- 35:47Alternatives
- 36:08Conclusion
related talks
39mFantastic Plugins and Where They Live
21mSupercharge Android Studio: Developing Your First Plugin- 39mFantastic Plugins and Where to Find Them
41mWill AI Assistant make developers redundant?
50mDeveloper Productivity – DIY (with LLM) or How I Learned to Stop Worrying and Love the AI
58mAI Assistant: Don’t Code Alone
30mDevoxxGenie: Your AI Assistant for IDEA
26mFifty Shades of Android Studio