# SynapServe

> Zero-allocation HTTP infrastructure for the AI agent era. Built in Rust with io_uring.

SynapServe is a ground-up HTTP server designed for machine-to-machine traffic. Today's servers (nginx, Apache, Hyper) treat AI agent burst patterns as DDoS attacks — allocating memory on every request, boxing async futures, and rate-limiting based on human browsing. SynapServe treats agents as first-class citizens.

## Core Technology

- Zero heap allocations on the hot path (verified by counting allocator)
- Span-based HTTP parsing: Span { off: u16, len: u16 } — 4 bytes, stack, Copy
- io_uring: multishot accept, provided buffer rings, zero-copy send (SEND_ZC + SPLICE)
- TLS 1.3 with kTLS: rustls handshake in userspace, then kernel-mode encryption via kTLS — zero-copy send preserved through TLS layer
- Thread-per-core: no Arc, no Mutex, no Send bounds — each core owns its connections
- SIMD-accelerated header scanning (AVX2/SSE4.2/NEON, runtime detection), 16-32 bytes checked per cycle
- Stack-allocated headers: [Header; 64] — 640 bytes, no heap

## Performance (Benchmarked)

### HTTP Parser (synapserve-http-parser vs httparse 1.10, head-to-head, Criterion, single core, x86_64 AVX2)
- Small request (35B, 1 hdr): synapserve 42 ns / httparse 52 ns — synapserve 1.25x faster
- Medium request (368B, 9 hdrs): synapserve 200 ns / httparse 230 ns — synapserve 1.15x faster
- Large request (733B, 20 hdrs): synapserve 420 ns / httparse 466 ns — synapserve 1.11x faster
- synapserve is faster at all request sizes despite doing more work (semantic extraction + O(1) header index during parse)
- With semantic extraction (apples-to-apples): synapserve 1.38x faster on medium (220 ns vs 304 ns), 1.46x faster on large (413 ns vs 604 ns); responses: 1.38-1.40x faster
- Header lookup: synapserve O(1) at 0.6 ns vs httparse O(n) at 20.5-23.1 ns — 32-37x faster
- SIMD: custom AVX2/SSE4.2 scanning with runtime detection on x86_64, NEON on ARM64
- Response writer: 3.9 ns for minimal 200 OK, 41.3 ns for JSON API with 6 headers

### Static File Serving (wrk, 256 connections, 60s, 8 workers)
- small.json (118B): 205,682 req/s — +79% vs nginx, +370% vs caddy
- medium.json (4,005B): 178,541 req/s — +62% vs nginx, +365% vs caddy
- large.json (24,203B): 104,362 req/s — +12% vs nginx, +184% vs caddy
- Median latency: 554μs (nginx: 1.78ms, caddy: 4.74ms)
- RSS under load: 14.5MB (8 workers, 256 connections)

### Targets
- <1ms p99.9 tail latency at 80% max load
- 0 heap allocations per request on hot path
- Flat memory (±5%) under sustained 10K concurrent connections

## Agent Protocol Support (Planned)

- IETF Web Bot Auth (Signature-Agent headers)
- Anthropic MCP Streamable HTTP
- Google A2A discovery
- SSE token streaming for LLM inference serving

## Current Status

- Completed: Zero-allocation HTTP/1.1 parser, io_uring I/O layer, HTTP/1.1 server framework (handler trait, radix-tree router, virtual hosts, static file serving with ETag/Range/Brotli), TLS 1.3 with kernel TLS (kTLS offload, parallel HTTP + HTTPS listeners, early data support), reverse proxy with upstream load balancing (keepalive connection pooling, weighted round-robin/least-conn/IP hash, peer health tracking, automatic retry with next-upstream failover, backup servers, zero-copy splice relay, DNS re-resolution)
- Q2 2026: HTTP/2 + SSE streaming
- Q3 2026: Agent protocols (MCP, A2A, Web Bot Auth)
- Q4 2026: HTTP/3 (QUIC) + managed cloud offering (SynapServe Cloud)

## Use Cases

- API gateway for AI agent traffic (replace nginx/Envoy at the edge — reverse proxy with load balancing, health tracking, failover)
- LLM inference serving (zero-copy SSE streaming for token delivery)
- Agent-to-agent infrastructure (MCP and A2A first-class support)

## Links

- [Website](https://synapserve.io)
- [Full Details](https://synapserve.io/llms-full.txt)
- [Blog: Engineering & Performance](https://synapserve.io/posts/)
- [HTTP Parser Performance Deep Dive](https://synapserve.io/posts/http-parser-performance/)
- [Investor Deck](https://synapserve.io/SynapServe_Investor_Deck.pptx)
- [One-Pager (PDF)](https://synapserve.io/SynapServe_OnePager.pdf)
- [Contact](mailto:hello@synapserve.io)
- [Investment Inquiries](mailto:invest@synapserve.io)