convez.ai
MCP ServerIntegração & APIsIntermediário
blackwell-systems

BetterThanJSON - GCF para Servidores MCP

blackwell-systems/betterthanjson

Converta JSON em GCF: 79% menos tokens, 90.7% acurácia de compreensão

Ver no GitHub
InicialPluginScore · de 0 a 100

MCP server que converte respostas JSON em GCF (Graph Codex Format), um formato otimizado para LLMs que reduz tokens em 79% e melhora compreensão para 90.7% em média. Funciona como proxy transparente — seu servidor segue enviando JSON, o modelo recebe GCF.

Instalação

$npx convez-ai blackwell-systems/betterthanjson

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
14%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Otimizar custos de API com redução drástica de tokens de entrada
  • Melhorar precisão de LLMs ao processar grandes volumes de dados estruturados
  • Integrar com MCP servers existentes sem modificar código

Veredito

Ferramenta sólida com benchmarks robustos (1.300+ avaliações, 10 modelos testados) que resolve um problema real: JSON é ineficiente para LLMs. A implementação como proxy MCP é elegante e reduz fricção de adoção. Ideal para aplicações com alto volume de dados estruturados onde custo e precisão importam.

README

A tradução para o português deste README ainda está na fila.

betterthanjson.com

GCF is better than JSON for AI agents. Full benchmark data proving it.

Why JSON fails at scale

JSON was designed in 2001 for browser-to-server communication. When an LLM receives 500 records as JSON, it gets 53,341 tokens of repeated field names ("qualified_name":, "kind":, "score": on every record). At this scale, GPT-5.5 returns empty strings. Claude Opus spends 143 lines manually enumerating symbols and still gets the wrong answer. JSON averages 53.6% comprehension accuracy across 10 models.

What GCF does differently

GCF declares field names once in a header. Rows are positional values. Section headers (## targets, ## related) replace per-record metadata. The same 500-record payload uses 11,090 tokens instead of 53,341.

  • 79% fewer input tokens than JSON
  • 63% fewer output tokens than JSON
  • 90.7% comprehension accuracy across 10 models and 3 providers (Anthropic, OpenAI, Google)
  • Four models hit 100%: Claude Sonnet, Gemini 2.5 Pro, Gemini 3.1 Pro, Gemini 3.5 Flash
  • 5/5 generation validity on every frontier model with zero prior training
  • Wins all 6 datasets on TOON's own benchmark using their datasets and tokenizer

The benchmark

1,300+ LLM evaluations. 23 comprehension runs. 28 generation runs. 10 models. 3 providers. Deterministic ground truth, no LLM judge, reproducible from one command.

Comprehension (can the model read it?)

ModelGCFTOONJSON
Claude Opus 4.696.2%84.6%73.1%
Claude Sonnet 4.6100%73.1%53.8%
Claude Haiku 4.596.2%69.2%57.7%
GPT-5.584.1%67.7%45.8%
GPT-5.478.0%56.0%44.1%
GPT-5.4-mini71.8%64.1%54.2%
Gemini 2.5 Pro100%76.9%58.3%
Gemini 3.1 Pro100%76.9%46.2%
Gemini 3.5 Flash100%61.5%46.2%
Gemini 2.5 Flash80.6%54.6%57.0%

GCF > TOON > JSON on every model from every provider. No exceptions.

Generation (can the model write it?)

ModelGCFTOONJSON
Claude Opus 4.65/50/55/5
Claude Sonnet 4.65/52-3/55/5
GPT-5.54-5/51-2/55/5
GPT-5.45/50/55/5
Gemini 2.5 Pro5/51/55/5
Gemini 3.1 Pro5/50/55/5

TOON's official decoder rejects LLM-generated output on 7 of 9 models. GCF achieves 5/5 on every frontier model with zero training.

Token efficiency

FormatTokens (500 symbols)vs JSON
GCF11,09079% fewer
TOON16,37869% fewer
JSON53,341baseline

How to try it

pip install gcf-proxy

Wrap any MCP server. Zero code changes. Your server keeps outputting JSON, the model receives GCF.

Links

License

MIT - Dayna Blackwell

Ver o original no GitHub

Bom para

Relacionados