A program structure interface designed for AI coding agents — exposing the structural semantics of codebases so agents can navigate, understand, and modify code with surgical precision.
Current AI coding agents operate primarily on raw text — they lack deep understanding of program structure. This leads to imprecise edits, broken references, and an inability to reason about complex codebases beyond the current file.
PSI bridges this gap by providing a structured interface that extracts and exposes function call graphs, data flow, type hierarchies, and symbol resolution — giving agents the same structural awareness that experienced developers have.
PSI provides a rich, queryable interface to the structural properties of any codebase.
Build complete, cross-file function call graphs. Understand call chains, entry points, and dependency relationships across your entire project.
Navigate code by meaning — go to definition, find all references, resolve imports, and traverse inheritance hierarchies programmatically.
A clean, tool-call compatible interface designed for LLM-based coding agents. Query structure, get context-aware code slices, and apply targeted edits.
Track how data moves through your program — from function parameters to return values, across modules and abstraction layers.
Extract and visualize class hierarchies, interface implementations, and generic type relationships for object-oriented and typed codebases.
Supports Python, JavaScript/TypeScript, Java, Go, and C/C++ with language-specific semantic analysis via tree-sitter and LSP integration.
PSI integrates seamlessly into your AI agent's tool ecosystem.
PSI parses and indexes the target codebase, building an in-memory structural graph of all symbols, relationships, and scopes.
The agent queries PSI through structured tool calls — "find callers of function X", "get class hierarchy of Y", "list all imports in module Z".
With structural context, the agent makes precise, semantics-aware code modifications — avoiding the "edit by text pattern" failures common in naive agents.