TZP v1.0 · Open Standard

AI-to-AI Communication
Reinvented

TokenZip Protocol (TZP) is the universal semantic shared memory standard for heterogeneous AI agents. Replace 10,000-token payloads with a 15-character pointer. Try the interactive demo below.

0%
Bandwidth Reduction
0%
Latency Reduction
0%
Cost Savings
GitHub
The Problem
The Fatal Bottleneck of AI-to-AI Communication
Today's multi-agent systems rely on passing full natural-language context between agents — resulting in explosive latency and runaway API costs at scale.
Traditional

Pass-by-Value

Agent A ──── [10,000 tokens full context: "Q1 2026 Global AI Market Report...The market showed strong upward momentum, driven by: 1. Central bank monetary policy shifts...2. AI-driven industrial transformation...3. Easing geopolitical tensions..."] ────▶ Agent B
Transfer Size~40 KB
Latency~2,000 ms
Cost per Call$0.030
Tokens Used10,000
TZP

Pass-by-Reference

Agent A ──── "See [TZP: tx_us_8f9A2bXr7]" ────▶ Agent B ✅ 15-char pointer ✅ O(1) transfer complexity
Transfer Size~7.6 KB
Latency~50 ms
Cost per Call$0.001
Tokens Used0
Protocol Lifecycle
The Four-Phase TZP Workflow
From capability negotiation to zero-overhead addressing, a complete TZP exchange completes in milliseconds.
0

Capability Exchange

Agent A confirms Agent B supports TZP-Core via an out-of-band handshake. 5-second timeout auto-falls back to full-text transfer.

{ "tzp_capability_response": { "supported": true, "features": ["TZP-Core"] } }
I

Semantic Quantization

Long text is chunked semantically, embedded via all-MiniLM-L6-v2 into 384-d vectors, then percentile-based Int8 quantized — 81% compression.

10,000 words → 20 chunks × 384d Float32 → 20 × 384 Int8 ≈ 7.6 KB
II

Edge Caching

The semantic payload is POSTed to the global edge network. The server generates a globally unique TrexID pointer with configurable TTL and geo-routing.

POST /v1/payloads → 201 { "trex_id": "tx_us_8f9A2bXr7" }
III

Zero-Overhead Addressing

Agent A embeds only the TrexID marker in its prompt. Agent B's Interceptor auto-detects, fetches, dequantizes, and injects the full context.

"Analyze the report. Context: [TZP: tx_us_8f9A2bXr7]"
Live Playground
Interactive API Demo
The controls below call the running TrexAPI backend in real time. Experience the full Push / Pull workflow live.

Push POST /v1/payloads

Results will appear here...

Pull GET /v1/payloads/:id

Enter a TrexID and click Pull to retrieve the payload...
Performance
Benchmark Results
Based on the reference implementation. Averaged over 1,000 round-trips with 10,000-token English documents.
Avg. Transfer Size
Traditional
TZP
Avg. End-to-End Latency
Traditional
TZP
API Cost per 1K Calls
Traditional
TZP
Semantic Fidelity (Cosine Sim.)
Baseline
TZP
Ecosystem
Open Ecosystem & Roadmap
TZP is an open protocol. Anyone can build a compliant SDK or operate an edge node.
🐍

Python SDK

Full semantic quantization, payload push/pull, E2EE encryption. Integrates with HuggingFace Transformers.

In Development
🟦

TypeScript SDK

Lightweight browser & Node.js dual-runtime support. ONNX Runtime inference, zero native dependencies.

In Development
🐹

Go SDK

Optimized for high-concurrency workloads. Kubernetes-native deployment and gRPC gateway support.

In Development
🦀

Rust SDK

Maximum performance and memory safety. Targeting embedded edge nodes and WASM runtimes.

Planned
🌐

TrexAPI Edge Network

Globally distributed across 12 regions. Built on Cloudflare Workers for sub-50ms latency worldwide.

In Development
🤝

MCP / A2A Interop

Seamless integration with Model Context Protocol and Agent-to-Agent standards as a transport-layer optimization.

In Development
🏅

TZP-Core

Quantization + Push/Pull

🛰️

TZP-Network

+ Global Edge Cache Network

🏢

TZP-Enterprise

+ E2EE / SOC 2 / 99.9% SLA