n8n-nodes-gcf

n8n community node for GCF (Graph Compact Format) -- bidirectional conversion between GCF and JSON with 71% fewer tokens and zero config.

GCF is a structured data wire format optimized for LLM tool responses. It achieves 90.7% comprehension across 10 models while cutting token costs by 71% compared to JSON.

Installation

In your n8n instance:

npm install n8n-nodes-gcf

Or install via the n8n UI: Settings > Community Nodes > Install > n8n-nodes-gcf

Operations

Encode (JSON to GCF)

Converts JSON input to GCF format. Automatically detects whether to use the graph profile (for tool payloads with tool + symbols fields) or the generic profile (everything else).

• Input Data: Drag and drop JSON data, use {{ $json }}, or reference a field name
• Output Field: Field name for the GCF output (default: data)
• Include Token Metrics: Adds estimated token counts for JSON vs GCF with reduction percentage

Decode (GCF to JSON)

Parses GCF text back into JSON. Automatically detects graph vs generic profile from the header.

• Input Data: A GCF-formatted string (field reference or literal text)
• Output Field: Field name for the JSON output (default: data)

AI Agent Integration

This node has usableAsTool: true, so n8n AI agents can use it directly to compress tool responses before sending them to the LLM, saving tokens on every call.

Example Workflow

1. HTTP Request node fetches API data (JSON)
2. GCF node encodes it to GCF (Encode operation)
3. AI Agent node receives the compressed payload, saving 71% on input tokens
4. GCF node decodes agent output back to JSON (Decode operation)

Token Savings

Enable "Include Token Metrics" to see per-item comparisons:

{
  "data": "GCF profile=generic\n...",
  "tokenMetrics": {
    "jsonTokens": 1200,
    "gcfTokens": 350,
    "tokensSaved": 850,
    "reductionPercent": 70.83
  }
}

Links

• GCF Specification
• Interactive Playground
• Benchmark Results
• GitHub
• GCF TypeScript Library

Why GCF over TOON?

	GCF	TOON
Token savings vs JSON	71%	40%
GCF vs TOON (15 datasets)	25.5% fewer	baseline
LLM comprehension (500 records)	100% on every frontier model	Fails on GPT-5.5
LLM generation validity	5/5 on every frontier model	Rejected by 7/9 models
Source format support	JSON, YAML, TOML, CSV, MessagePack	JSON only
Runtime dependencies	Zero	1
Lossless round-trips verified	33 billion+	None published
Session deduplication	Yes (92% savings by 5th call)	No
Graph/relationship encoding	Yes (local IDs, typed edges)	No

Full comparison with benchmarks

How GCF Works

GCF replaces JSON's verbose syntax with a compact, deterministic encoding:

• No braces, brackets, or quotes for keys
• Tabular arrays encoded as pipe-separated rows
• Nested structures via indentation
• Zero config: encodeGeneric(data) and decodeGeneric(text) handle everything

The encoding is fully deterministic and round-trip safe.

License

MIT