sabrix-bench ⚡

Ultra-fast, zero-bloat developer CLI and benchmark harness for Model Context Protocol (MCP) traffic inspection and agent loop latency profiling.

🎯 The Problem

Developers building autonomous AI agents locally lack visibility into raw Model Context Protocol (MCP) JSON-RPC tool traffic and have no lightweight way to measure per-turn serialization and proxy latency overhead.

Legacy approaches introduce massive performance and security taxes:

• Legacy Python / Node Proxies: Add 30ms – 50ms of serialization and runtime tax per tool call.
• SaaS AI Firewalls: Incur 100ms – 250ms of WAN network latency, TLS handshakes, and third-party cloud data egress per turn.
• Over a 30-turn agent loop, legacy firewalls add 3.6+ seconds of dead wait time and leak raw database queries and internal system commands outside your perimeter.

sabrix-bench gives you real-time visibility into your local MCP tool calls and benchmarks your agent loops in microseconds ($< 2\ \mu\text{s}$) with zero network egress.

🚀 Quickstart & Installation

Install via Cargo (Crates.io)

cargo install sabrix-bench

Or Install via Git / Source

cargo install --git https://github.com/Pro-Kla/sabrix-bench

Or Build & Install from Local Source

# Clone the repository and install the binary
git clone https://github.com/Pro-Kla/sabrix-bench.git
cd sabrix-bench
cargo install --path .

Or run directly with Cargo:

cargo run -- --help

🛠️ CLI Usage & Subcommands

1. sabrix-bench trace — Real-Time MCP Security Inspector

Inspect JSON-RPC 2.0 requests (tools/call, resources/read, tools/list) and flag dangerous tool mutations (destructive shell commands, SQL injections, leaked credentials).

Run Built-in Demo Scenarios:

sabrix-bench trace --demo

Inspect from Inline JSON Payload:

sabrix-bench trace -p '{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "database_query",
    "arguments": { "sql": "DROP TABLE users; --" }
  }
}'

Pipe from Stdin:

cat mcp_message.json | sabrix-bench trace

2. sabrix-bench bench — Multi-Turn Agent Loop Benchmark

Simulate multi-turn autonomous agent loops to measure local parsing overhead, memory serialization, and latency percentiles ($p50$, $p95$, $p99$).

# Run a 30-turn benchmark with real-time progress
sabrix-bench bench --turns 30

# Benchmark larger context payloads (10x scale)
sabrix-bench bench --turns 50 --scale 10

# Output machine-readable JSON for CI/CD pipelines
sabrix-bench bench --turns 20 --json

3. sabrix-bench compare — Live Multi-Turn Latency Comparison

Run an empirical, turn-by-turn benchmark measuring safe-Rust in-process execution on your CPU against the WAN network overhead of remote SaaS AI firewalls (~120 ms).

# Run dynamic 30-turn live latency comparison
sabrix-bench compare --turns 30

# Output machine-readable comparison JSON for CI/CD pipelines
sabrix-bench compare --turns 20 --json

# View full static architectural comparison matrix
sabrix-bench compare --matrix

📊 Benchmark & Latency Tax Breakdown

🛡️ Built-in Risk Checks

sabrix-bench evaluates deterministic security rules locally in sub-microsecond time:

• MCP-SEC-001: Destructive Filesystem Operations (rm -rf, mkfs, dd, chmod 777)
• MCP-SEC-002: Remote Code Execution / Reverse Shells (curl | bash, nc -e, /dev/tcp/)
• MCP-SEC-003: Destructive SQL Mutations & Injections (DROP TABLE, TRUNCATE, DELETE FROM, WHERE 1=1)
• MCP-SEC-004-007: Leaked API Keys & Secrets (OpenAI sk-, GitHub ghp_, AWS AKIA, PEM Private Keys)
• MCP-SEC-008: Sensitive Path Egress (/etc/passwd, ~/.ssh/id_rsa, ~/.aws/credentials, .env)
• MCP-SEC-009: Unconstrained Arbitrary Execution Tool Invocations

🌟 Deploying to Production?

Enforce zero-egress In-VPC MCP security with millisecond-grade deterministic policy control:

👉 Deploy Sabrix In-VPC Gateway

📜 License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.