TSCG - Token-Context Semantic Grammar Compiler for LLM Tool Schemas

Reference implementation of TSCG (Token-Context Semantic Grammar), a deterministic compiler for transforming JSON tool schemas into compressed, model-readable formats for use in agentic AI systems.

This bundle contains the source code of 4 npm packages:

@tscg/core (1.4.3): Core compiler with 8 operators (SDM, TAS, DRO, CFL, CFO, CAS, SAD, CCP). Sub-millisecond execution. Zero dependencies.

@tscg/mcp-proxy (1.4.3): Model Context Protocol-aware proxy with adaptive profile resolution and per-model cache reader.

@tscg/tool-optimizer (1.4.3): Framework integrations for LangChain, MCP, Vercel AI SDK.

@tscg/openclaw (1.4.3): OpenClaw plugin with adaptive sweep CLI for per-model empirical profile detection.

Validated across 20,000+ API calls on 15+ models (4B-32B parameters + frontier models GPT-4o, GPT-5.2/5.4/5.5, Claude Sonnet 4, Claude Opus 4.7).

Sub-millisecond execution. Zero dependencies. MIT licensed.

Companion findings library: github.com/SKZL-AI/tscg/tree/main/findings