v0.6.1 Released•Sub-Millisecond AI Gateway & Multi-Account Router

The Sub-Millisecond AI Gateway for Autonomous Coding Agents

Aggregate 330+ providers, eliminate rate limits with recursive failover cascades, and run in < 5 MB RAM with Rust SIMD acceleration.

npm install -g goroute
0.86 ms
Routing Overhead
4.7 MB RSS
Daemon Memory
Recursive Combos
Zero Rate Limits
330+ Providers
Auto-Configured
goroute daemon :8128 (online)
Incoming POST /v1/messages | Model: "frontier" | Client: Claude Code v2.1
[router] resolving virtual combo 'frontier' → cascade target: 'claude-models'
⚠ upstream antigravity/claude-sonnet-4-6 returned HTTP 429 Quota Exhausted
✓ failover triggered in 0.86ms → routed to openai-compatible/deepseek-chat
[stream] bidirectional SSE tunnel open | TTFT: 142ms | Tokens: 3,842 out
[status] 200 OK | Zero disruption to agent session | Cooldown applied to exhausted key (120s)
Showdown Benchmarks

Stop Burning Gigabytes of RAM on Background Gateways

Traditional AI gateways run heavy Node.js runtimes or Python virtual environments that idle at 400 MB+ and spike into gigabytes under load. GoRoute was built from the ground up in Go + Rust to make local proxying invisible.

OmniRouteNode.js / Next

Requires Node 20+, 800+ npm dependencies, heavy Next.js frontend engine running 24/7.

Idle Memory~380 MB RSS
Routing Latency18.4 ms
Cold Startup3,200 ms
Dependencies800+ node_modules
Heavy background battery & RAM drain
98.8% Less RAM
GoRouteGo + Rust SIMD

Single static binary. Zero CGO runtime. Self-contained SQLite WAL engine and embedded UI.

Idle Memory4.7 MB RSS
Routing Latency0.86 ms (21x faster)
Cold Startup11 ms
DependenciesZero (Single executable)
Continuous 24/7 daemon with zero system impact
LiteLLMPython Runtime

Interpreted Python WSGI server requiring virtual environments, pip wheels, and GIL overhead.

Idle Memory~420 MB RSS
Routing Latency24.1 ms
Cold Startup2,800 ms
DependenciesPython 3.10+ / Pip
GIL lock bottlenecks concurrent agent SSE streams

Concurrent Agent Concurrency Scaler

Adjust the active agent slider to compare system RAM consumption under parallel workload:

10 Concurrent Agents
1 agent (idle)25 agents (team)50 agents (CI/CD)100 agents (enterprise)
GoRoute (Go + Rust)5.0 MB
OmniRoute (Node.js)408 MB
LiteLLM (Python)454 MB
Internal Pipeline

Zero-Copy Architecture Built for High-Throughput Concurrency

Every layer in GoRoute is optimized for microsecond dispatch. Requests enter via the Chi router, pass through Rust SIMD fast-paths, evaluate recursive composite trees, and tunnel directly to upstream SSE sockets.

1. Client Agents & Developer Tools

Ingress

Connects Claude Code, Cursor, CLI coding agents, and any tool speaking standard OpenAI or Anthropic protocols with zero SDK modifications.

2. Rust SIMD Dispatch & Protocol Normalizer

Engine

Zero-copy HTTP engine dynamically linked to Rust cdylib via Purego. Translates requests between OpenAI, Anthropic, and Gemini schemas in sub-milliseconds.

3. Recursive Composite Cascades

Router

Hierarchical combo routing engine. Combos can reference other combos with automatic cycle detection, depth bounds, and health circuit breakers.

4. Multi-Account Pool & Circuit Breakers

Persistence

Autonomous OAuth token refresh loops and SQLite WAL persistence. Rotates accounts seamlessly and isolates unhealthy endpoints with backoff cooldowns.

5. Upstream AI Providers & Inference Networks

Egress

Direct, authenticated streaming tunnels to 330+ supported AI providers, local LLMs, and frontier foundation models with zero payload mutation.

Layer Deep Dive
Live Diagnostic

3. Recursive Composite Cascades

Pipeline Stage 3 of 5

Hierarchical combo routing engine. Combos can reference other combos with automatic cycle detection, depth bounds, and health circuit breakers.

Engine Guarantees:
  • •Recursive combo nesting: e.g. 'frontier' &rarr; 'claude-models' + 'deepseek' + 'gpt'
  • •Instant failover under 1ms if an upstream provider yields 429, 503, or rate exhaustion
  • •Exponential-weighted moving average (EWMA) latency selection across pooled accounts
Core Capabilities

Engineered Without Compromise for High-Stakes Agentic Workflows

From single-developer setups to multi-agent teams running dozens of autonomous Claude Code tasks simultaneously, GoRoute provides enterprise-grade reliability in an ultra-compact binary.

Zero Rate Limits

Recursive Composite Combos

Group accounts and models into hierarchical failover cascades. Combos can reference other combos with cycle detection and depth bounds. If Claude Sonnet 4.6 yields HTTP 429, GoRoute immediately pivots to DeepSeek or Gemini within 0.86ms without dropping the client stream.

•virtual_combo: "frontier"Composite Strategy
→ combo: "claude-models" (Target 1: Google Antigravity OAuth Pool)Active
→ combo: "deepseek" (Target 2: OpenAI-Compatible API)Standby
→ combo: "gpt" (Target 3: OpenAI Official)Standby
Cycle-safe resolution with max recursion depth gates

< 5 MB Daemon

Compiled directly to pure native machine code. Runs quietly as a background systemd user daemon or launchd agent in under 5 MB of physical RAM.

Idle Memory:4.7 MB RSS
CGO Runtime:0 (Pure Go)
Dependencies:Zero external
Leaves 99% RAM free for your IDE

Rust SIMD Engine

Optional cdylib linking via Purego. Accelerates high-frequency token counting, JSON AST normalization, and string scanning with vector SIMD instructions.

crates/goroute-fast
AVX2 / NEON vectorized resolver
Purego bridge (No CGO needed)
0.86ms average dispatch latency under 500 concurrent workers

Universal Protocol

Transparently converts between Anthropic Messages (`/v1/messages`), OpenAI Chat (`/v1/chat/completions`), and Google Gemini formats with full thinking token support.

  • •Anthropic thinking budget passthrough
  • •Tool & function calling translation
  • •count_tokens endpoint emulation
Seamless Claude Code & Cursor interop

1-Click Vault Migration

Migrating from OmniRoute? One click in the Web Dashboard or running `goroute sync omniroute` reads, decrypts, and imports all accounts and combos in 0.2s.

$ goroute sync omniroute
→ Decrypting SQLite credentials...
→ 5 Antigravity + 3 Kiro accounts imported
Zero manual account re-authentication

Full-Fledged Web Portal Built-In

Open `http://localhost:8128/` anytime to inspect active quotas, toggle provider cooldowns, visually reorder combo cascade priorities, and test completions in the real-time playground.

25+ Auto-Configured TemplatesAgent Fleet 1-Click SetupInteractive Playground & Metrics
Portal Ingresshttp://127.0.0.1:8128
Compiled Into Binary//go:embed web/dist
Memory Overhead0 MB extra
Universal Ecosystem

Connect Any Model. Any Lab. 330+ Providers Out-of-the-Box

From autonomous Google Cloud Code OAuth pools to custom OpenAI-compatible proxies and self-hosted Ollama instances, GoRoute handles authentication and payload normalization universally.

Anthropic
AnthropicFrontier Lab
OpenAI
OpenAIFrontier Lab
DeepSeek
DeepSeekReasoning Models
Groq
GroqLPU Inference
Mistral AI
Mistral AIEuropean Frontier
Qwen
QwenAlibaba Cloud
Cerebras
CerebrasWafer-Scale CS-3
Ollama
OllamaLocal Offline
Cohere
CohereEnterprise NLP
Cloudflare Workers AI
Cloudflare Workers AIEdge GPU
Perplexity
PerplexitySearch & Sonar
Meta LLaMA
Meta LLaMAOpen Source
Anthropic
AnthropicFrontier Lab
OpenAI
OpenAIFrontier Lab
DeepSeek
DeepSeekReasoning Models
Groq
GroqLPU Inference
Mistral AI
Mistral AIEuropean Frontier
Qwen
QwenAlibaba Cloud
Cerebras
CerebrasWafer-Scale CS-3
Ollama
OllamaLocal Offline
Cohere
CohereEnterprise NLP
Cloudflare Workers AI
Cloudflare Workers AIEdge GPU
Perplexity
PerplexitySearch & Sonar
Meta LLaMA
Meta LLaMAOpen Source
AWS Bedrock
AWS BedrockCloud Infrastructure
Azure OpenAI
Azure OpenAICloud Infrastructure
Brave Search API
Brave Search APIWeb Retrieval
Vercel AI SDK
Vercel AI SDKServerless
Chutes AI
Chutes AIDecentralized Compute
Cartesia
CartesiaVoice & Audio
Clarifai
ClarifaiComputer Vision
Baidu Qianfan
Baidu QianfanAsia Frontier
Arcee AI
Arcee AISmall Language Models
AssemblyAI
AssemblyAISpeech-to-Text
DeepSeek Reasoner
DeepSeek ReasonerThinking Tokens
Anthropic Claude
Anthropic ClaudeMessages v1
AWS Bedrock
AWS BedrockCloud Infrastructure
Azure OpenAI
Azure OpenAICloud Infrastructure
Brave Search API
Brave Search APIWeb Retrieval
Vercel AI SDK
Vercel AI SDKServerless
Chutes AI
Chutes AIDecentralized Compute
Cartesia
CartesiaVoice & Audio
Clarifai
ClarifaiComputer Vision
Baidu Qianfan
Baidu QianfanAsia Frontier
Arcee AI
Arcee AISmall Language Models
AssemblyAI
AssemblyAISpeech-to-Text
DeepSeek Reasoner
DeepSeek ReasonerThinking Tokens
Anthropic Claude
Anthropic ClaudeMessages v1
SSRF Link-Local Protected • Zero Telemetry Leakage • Strictly Localhost Daemon
Installation & Setup

Up and Running in Under 60 Seconds

No Docker daemon required. No Python dependencies. Install a single static executable and immediately route your favorite coding agents through resilient multi-account cascades.

curl -fsSL https://raw.githubusercontent.com/ziuus/goroute/main/install.sh | bash
1

Launch Daemon

Start the local background gateway. Run once or enable automatic system boot launch.

goroute autostart enable
• Web portal live on http://127.0.0.1:8128
2

Connect Accounts

Authenticate Google Cloud Code, import AWS Builder ID, or add API keys via the Web Studio.

goroute auth login antigravity
• 25+ auto-configured URL presets included
3

Link Coding Agent

Auto-configure Claude Code, Cursor, or your entire agent fleet with a single CLI command.

goroute setup claude
• Ready to code with model "frontier"