50-92% fewer tokens than JSON. 100% comprehension on every frontier model. Delta encoding compounds across a session. 43 billion+ lossless round-trips across 5 formats (JSON, YAML, TOML, CSV, MessagePack). Reverse-engineered from attention-level research. Spec v3.4.1.
One command. 50-92% fewer tokens. Zero code changes.
SDK pip install gcf-proxyCopynpm i -g @blackwell-systems/gcf-proxyCopygo install github.com/blackwell-systems/gcf-proxy@latestCopy
JSON, YAML, TOML, CSV, MessagePack: any format in, GCF in the context window, any format out. One format that speaks every format. No other wire format operates across format boundaries.decode(encode(value)) == value for every structured value. Verified across 43,000,000,000+ lossless round-trips in 5 formats and 6 languages.
50-69% on a single call (generic to graph profile). Up to 92% with session deduplication (graph) and delta encoding (both profiles) across repeated tool calls.
At 1000 orders, JSON doesn't even fit in a 200K context window. GCF fits in 47K. And 29% fewer tokens than TOON across 16 real-world datasets (15/16 wins).
LLMs produce valid GCF with a 3-line primer.
33% smaller output than TOON. 5/5 generation validity on every frontier model across 3 providers. Zero training.
The only format that never fails. Tested across Claude, GPT-5.5, and Gemini with zero format instructions. On structurally complex code graphs, GCF scores 91.2% where JSON drops to 54.1% and TOON to 68.8%. 2,500+ evaluations, 4 providers.
43B+ (yes, really) round-trips across JSON, YAML, TOML, CSV, and MessagePack. Zero failures. 204 conformance fixtures. Cross-language 6x6 encode/decode matrix verified. Read the spec.
GCF's grammar was reverse-engineered from attention-level research, then validated by three published papers: controlled experiments across two architectures and two scales.
Clean delimiters produce up to 738x lower structured-data perplexity, at zero cost to natural language. No other wire format ships a falsifiable theory of why it works. The science.
Arrays of uniform records. Same data, fewer tokens, zero loss.
{
"orders": [
{"id": 1001, "customer": "Acme Corp",
"total": 49.99, "status": "shipped",
"items": 1},
{"id": 1002, "customer": "Globex Inc",
"total": 150.49, "status": "pending",
"items": 2},
{"id": 1003, "customer": "Initech LLC",
"total": 250.99, "status": "processing",
"items": 3}, {"id": 1004, ...},
{"id": 1005, ...},
{"id": 1006, ...},
{"id": 1007, ...},
{"id": 1008, ...},
{"id": 1009, ...},
{"id": 1010, ...}
]
}GCF profile=generic
## orders [10]{id,customer,total,status,items}
1001|Acme Corp|49.99|shipped|1
1002|Globex Inc|150.49|pending|2
1003|Initech LLC|250.99|processing|3
1004|Umbrella Co|351.49|delivered|4
1005|Stark Ind|451.99|shipped|5
1006|Wayne Ent|552.49|pending|6
1007|Oscorp|652.99|shipped|7
1008|LexCorp|753.49|processing|8
1009|Cyberdyne|853.99|delivered|9
1010|Soylent|954.49|shipped|1061% fewer tokens. Scales to 71%+ at production sizes.
10 orders, 5 fields. Full payload, tiktoken cl100k.
from gcf import encode_generic, decode_generic
gcf_string = encode_generic(data)
original = decode_generic(gcf_string)The generic profile is a strict subset of the graph profile. Call encode(), the LLM reads it natively, call decode() at the end.
Graph-shaped data is the fastest-growing data shape in AI: knowledge systems, ontologies, GraphRAG, code intelligence, agent memory. No other token-efficient format treats graphs as first-class. GCF is the only format with native graph syntax: local IDs, typed edges, distance grouping, and session deduplication that compounds to 92% savings across multi-turn sessions. Delta encoding, content-addressed identity, and streaming are not graph-only: the generic profile has them too, so tabular data gets the same incremental and multi-turn savings.
Any structured value in, same value out. Verified lossless across 43 billion+ round-trips with JSON, YAML, TOML, CSV, and MessagePack. 71% fewer tokens than JSON.
[
{"name":"validateToken",
"kind":"func","refs":18},
{"name":"refreshSession",
"kind":"func","refs":6},
{"name":"getConnection",
"kind":"func","refs":34},
{"name":"runMigration",
"kind":"func","refs":3}
]## results [4]{name,kind,refs}
validateToken|func|18
refreshSession|func|6
getConnection|func|34
runMigration|func|3Knowledge graphs, code intelligence, ontologies, relationship networks. Graph-shaped data is the fastest-growing data shape in AI. No other format treats it as a first-class citizen.
{
"symbols": [
{"id":1,"kind":"func",
"name":"handleReq"},
{"id":2,"kind":"func",
"name":"validate"},
{"id":3,"kind":"iface",
"name":"AuthCfg"}
],
"edges": [
{"src":1,"tgt":2,
"type":"calls"},
{"src":2,"tgt":3,
"type":"implements"}
]
}## symbols [3]
@0 fn handleReq 0.95
@1 fn validate 0.87
@2 iface AuthCfg 0.60
## edges [2]
@1<@0 calls
@2<@1 implementsData changed since the last call? Send only what changed. A keyed delta works for any tabular set (generic) or symbol and edge topology (graph); the graph profile can also dedup known symbols to bare references. 88-95% fewer tokens on small changes, up to 92% across a session, with content-addressed roots so the consumer can verify it applied cleanly.
[
{"id":1002,"total":29.99,
"status":"shipped"},
{"id":1003,"total":129.50,
"status":"shipped"},
{"id":1004,"total":75.00,
"status":"pending"}
]GCF profile=generic delta=true base_root=sha256:aaa... new_root=sha256:bbb... key=id
## added [1]{@id,total,status}
1004|75.00|pending
## changed [1]{@id,total,status}
1002|29.99|shipped
## removed [1]{@id}
1001Two profiles. Session-aware. No ambiguity.
## symbols [4]{kind,qname}
## edges [3]
## targets
##! summary## section start[N] element count{fields} inline schema[?] deferred count (streaming)##! summary trailer## users [3]{name,role,active}
Alice|admin|true
Bob|dev|true
Carol|dev|false| pipe-separated valuesname=Alice
age=30
active=true
missing=-
empty=""key=value for primitives- null, ~ absent, "" empty string| or newline## orders [2]{id,total,"customer>name","customer>tier"}
1001|249.99|Alice|premium
1002|89.50|Bob|standard"customer>name" flattens nested field into column> separates path levelsGCF profile=generic delta=true base_root=sha256:aaa new_root=sha256:bbb key=id
## added [1]{@id,total,status}
1004|75.00|pending
## changed [1]{@id,total,status}
1002|29.99|shipped
## removed [1]{@id}
1001key=id names the identity column## added / ## changed / ## removed sections## changed replaces the whole row by identity## removed carries identity values only@0 fn handleReq 0.95
@1 iface AuthCfg 0.60 ast
@2
@3 fn revoke 0.91@N local ID, kind, qname, score, provenance@N = session ref (already transmitted)@2<@1 calls
@3<@1 calls
@4<@2 implements@target<@source type## targets [3]
@0 fn handleReq 0.95 lsp
@1 fn validate 0.87 lsp
@2 fn connect 0.91 lsp
## related [2]
@3 fn helper 0.60 ast
@4 iface Config 0.55 asttargets, related, extended by relevanceGCF profile=graph tool=blast_radius symbols=6 session=true
## targets
@0 # previously transmitted
@1 # previously transmitted
@5 fn pkg.NewFunc 0.85 lspsession=true signals bare refs present@N # previously transmitted = sent in prior callGCF profile=graph tool=topology delta=true tokens=30 savings=85%
## removed
fn pkg.OldHandler
## added
@0 fn pkg.NewHandler 0.85 lsp 0
## edges_added
pkg.Router -> pkg.NewHandler callsdelta=true signals diff from prior payload## results [?]{name,kind}
validate|func
connect|func
handle|func
##! summary counts=3[?] count unknown upfront##! trailer finalizes count