TERSE Tool Catalog (TTC) — Extension Specification v1.0

TERSE Tool Catalog (TTC) is an extension of the TERSE Format specification (Carvalho, 2025) that defines a compact, semantically annotated representation for Model Context Protocol (MCP) tool catalogs in Large Language Model (LLM) contexts.

A typical MCP tool definition consumes 100–270 tokens of JSON Schema. A system with 50 installed tools therefore places 5,000–13,500 tokens of catalog overhead into the context window before any user instruction, memory, or reasoning occurs. Beyond cost, MCP JSON Schema was designed as a machine-readable execution contract, not a semantic contract for tool selection: there is no explicit trigger condition, no error contract, and no retrieval taxonomy.

TTC v1.0 addresses both problems simultaneously through three changes: (1) it removes the syntactic and documentary overhead of JSON Schema scaffolding; (2) it adds three first-class semantic fields — WHEN (semantic trigger), ERR (failure contract), and TAGS (retrieval taxonomy) — that MCP JSON Schema does not carry; (3) it defines a normative WHEN vocabulary that enforces inter-server consistency through controlled intent and action verbs.

A measured benchmark over 10 representative real MCP tool definitions (cl100k_base BPE tokenization, consistent with GPT-4 / Claude tokenization) shows a 66.6% reduction in tokens (1,948 → 650) while explicitly adding the three new semantic fields. Extrapolated to a 100-tool catalog, this corresponds to ~12,980 tokens saved per request. A separate TF-IDF cosine-similarity simulation over a 12-tool, 36-query benchmark indicates that TTC's controlled-vocabulary WHEN field improves tool-selection accuracy by 8.3 percentage points and average selection confidence by 0.041 over MCP free-form descriptions.

The result is not pure compression: it is a context-budget reallocation toward higher-value reasoning signals. The specification includes the TOOL block syntax, MCP server catalog grouping, ABNF grammar extension, normative WHEN vocabulary, a Python reference converter for MCP JSON Schema → TTC, and an explicit roadmap of future extensions (EXAMPLE, COST, AUTH, CHAIN, ALIAS, DEPRECATED).

This specification is complementary to prompt caching (which addresses financial cost only), and is designed to work natively with dynamic tool retrieval (RAG over tools) — patterns that prompt caching cannot serve.

Reference Python converter: https://github.com/RudsonCarvalho/terse-format