@blackwell-systems/gcf-proxy 0.4.0 → 0.6.0

package/README.md

@@ -1,8 +1,8 @@
 # gcf-proxy
 
-MCP proxy that re-encodes JSON tool responses as [GCF](https://gcformat.com/) — the most token-efficient wire format for LLMs. Drop-in, zero code changes. A TOON/JSON alternative that saves 63-79% of tokens.
+Bidirectional MCP proxy that translates between [GCF](https://gcformat.com/) and JSON. Drop-in, zero code changes to your server or client.
 
-**79% fewer tokens than JSON. 34% fewer than TOON. 90.5% average comprehension accuracy across 10 models and 3 providers. 1,300+ LLM evaluations. Zero training.**
+**79% fewer input tokens. 63% fewer output tokens. 90.7% comprehension accuracy across 10 models and 3 providers. 1,300+ LLM evaluations. Zero training.**
 
 Docs: [gcformat.com](https://gcformat.com/) · [Proxy Guide](https://gcformat.com/guide/proxy.html) · [Playground](https://gcformat.com/playground.html) · [GCF vs TOON](https://gcformat.com/guide/vs-toon.html)
 
@@ -24,64 +24,30 @@ npm install -g @blackwell-systems/gcf-proxy
 {"mcpServers": {"yours": {"command": "gcf-proxy", "args": ["your-mcp-server"]}}}
 ```
 
-Your server keeps outputting JSON. The LLM receives GCF. Nothing else changes.
-
 Works with Claude Code, Claude Desktop, VS Code, Cursor, and any MCP client.
 
 ## What it does
 
-```
-LLM  ←──  GCF  ←──  gcf-proxy  ←──  JSON  ←──  Your Server
-```
-
-1. Spawns your MCP server as a subprocess
-2. Proxies stdin/stdout between client and server
-3. Detects JSON payloads in tool responses
-4. Re-encodes as GCF (graph profile for code intelligence, generic profile for everything else)
-5. Non-convertible responses (text, HTML, errors) pass through unchanged
-
-## Savings
-
-Tested on a real MCP tool response (10 symbols, 8 edges):
+Translates in both directions:
 
 ```
-JSON   2,506 bytes  ~626 tokens
-GCF      916 bytes  ~229 tokens
-
-Savings: 63% fewer tokens
+Responses:  Your Server (JSON) -> gcf-proxy encodes -> LLM reads GCF   (79% input savings)
+Requests:   LLM writes GCF    -> gcf-proxy decodes -> Your Server (JSON) (63% output savings)
 ```
 
-On a real agent-lsp blast_radius response (7 symbols, 47 callers):
-
-```
-JSON   6,515 bytes  ~1,628 tokens
-GCF    4,866 bytes  ~1,216 tokens
-
-Savings: 25% fewer tokens (generic encoding, no graph profile)
-```
-
-## When to use
-
-- You can't modify the server (third-party binary, another team's code)
-- You want to test GCF savings without writing any code
-- You want zero-effort adoption on any existing MCP server
-
-For session deduplication (92.7% savings) and delta encoding (81.2% savings), use the [GCF libraries](https://gcformat.com/ecosystem/implementations.html) directly.
+- **Responses**: JSON tool results from the server are encoded as GCF
+- **Requests**: GCF strings in tool call arguments are decoded to JSON (4-byte prefix check, zero overhead)
+- Non-convertible content passes through unchanged in both directions
+- Neither the server nor the client needs to know about GCF
 
 ## Benchmarks
 
 | Format | Accuracy | Tokens | vs JSON |
 |--------|----------|--------|---------|
-| **GCF** | **90.5%** avg (10 models) | **11,090** | **79% fewer** |
+| **GCF** | **90.7%** avg (10 models) | **11,090** | **79% fewer** |
 | TOON | 68.5% avg | 16,378 | 69% fewer |
 | JSON | 53.6% avg | 53,341 | baseline |
 
-GCF wins all 6 datasets on TOON's own benchmark. 42% smaller on semi-uniform data, 34% on mixed-structure.
-
-Reproduce comprehension eval: `git clone https://github.com/blackwell-systems/gcf-go && cd gcf-go/eval && GOWORK=off go test -run TestComprehension -v -timeout 0`
-
-Reproduce token benchmark: `git clone https://github.com/blackwell-systems/toon && cd toon && git checkout gcf-comparison && cd benchmarks && pnpm install && pnpm benchmark:tokens`
-
 ## Also available on
 
 - PyPI: `pip install gcf-proxy`
@@ -92,5 +58,4 @@ Reproduce token benchmark: `git clone https://github.com/blackwell-systems/toon
 - [Full Setup Guide](https://gcformat.com/guide/proxy.html)
 - [GCF Specification](https://gcformat.com/reference/spec.html)
 - [GCF vs TOON](https://gcformat.com/guide/vs-toon.html)
-- [Playground](https://gcformat.com/playground.html)
 - [GitHub](https://github.com/blackwell-systems/gcf-proxy)

package/package.json

@@ -1,17 +1,17 @@
 {
   "name": "@blackwell-systems/gcf-proxy",
-  "version": "0.4.0",
-  "description": "MCP proxy that re-encodes JSON tool responses as GCF. Zero code changes.",
+  "version": "0.6.0",
+  "description": "Bidirectional MCP proxy that translates between JSON and GCF. 79% fewer input tokens, 63% fewer output tokens. Zero code changes.",
   "bin": {
     "gcf-proxy": "bin/gcf-proxy"
   },
   "optionalDependencies": {
-    "@blackwell-systems/gcf-proxy-darwin-arm64": "0.4.0",
-    "@blackwell-systems/gcf-proxy-darwin-x64": "0.4.0",
-    "@blackwell-systems/gcf-proxy-linux-arm64": "0.4.0",
-    "@blackwell-systems/gcf-proxy-linux-x64": "0.4.0",
-    "@blackwell-systems/gcf-proxy-win32-x64": "0.4.0",
-    "@blackwell-systems/gcf-proxy-win32-arm64": "0.4.0"
+    "@blackwell-systems/gcf-proxy-darwin-arm64": "0.6.0",
+    "@blackwell-systems/gcf-proxy-darwin-x64": "0.6.0",
+    "@blackwell-systems/gcf-proxy-linux-arm64": "0.6.0",
+    "@blackwell-systems/gcf-proxy-linux-x64": "0.6.0",
+    "@blackwell-systems/gcf-proxy-win32-x64": "0.6.0",
+    "@blackwell-systems/gcf-proxy-win32-arm64": "0.6.0"
   },
   "keywords": [
     "gcf",
@@ -20,8 +20,11 @@
     "llm",
     "tokens",
     "json",
+    "json-alternative",
     "wire-format",
-    "toon-alternative"
+    "toon-alternative",
+    "ai",
+    "agent"
   ],
   "homepage": "https://gcformat.com/",
   "repository": {