Gortex Exposes Tree-Sitter Knowledge Graphs via Model Context Protocol to Cut Agent Tokens

Tree-Sitter AST Graphs over Model Context Protocol

When coding agents execute repository-wide edits, initial exploration frequently burns through massive context allocations. Gortex replaces linear string scanning with an on-disk SQLite knowledge graph generated through tree-sitter. The binary indexes functions, classes, call chains, HTTP endpoints, and cross-service dependencies, exposing them directly through a long-lived Model Context Protocol (MCP) daemon.

Parsing is categorized into three language tiers:

• Core Tier (16+ languages): High-precision AST parsing via hand-tuned queries for Go, Rust, TypeScript, Python, Java, Kotlin, C, C++, C#, PHP, Swift, Zig, Ruby, Elixir, OCaml, and Haskell.
• Heuristic Tier (~60 languages): Regex-based symbol extraction.
• Generic Tier (~165 languages): Structural signature fallback.

Multi-Repository Cross-Contract Mapping

For microservice architectures, Gortex extracts and unifies distributed contracts. It detects routes from web frameworks (FastAPI, Express, gin, Spring), alongside gRPC, GraphQL, and broker topics (Kafka, RabbitMQ). It normalizes endpoints into standardized IDs such as http::GET::/api/users/{id}, revealing missing consumers, orphaned routes, and schema drift across distinct git repositories.

Benchmarked Performance and Token Reductions

Project benchmarks report indexing the 70,333-file Linux kernel in approximately 3 minutes (300 files/sec, 5.07 GB peak heap on Apple Silicon) and the 10,762-file VS Code repository in roughly 1 minute (143 files/sec, 580 MB peak heap).

Because agents retrieve surgical AST declarations rather than raw files, sample ledgers report up to a 93.3% token reduction on complex queries. Precomputed depth-3 reach queries provide instant blast-radius feedback during code modifications.