Two profiles, no third one. The shapes below are canonical xray-core client outbounds; the Rust shell materializes profiles into exactly this JSON.
See src/lib/profile.ts for the Zod source of truth
and DEVELOPMENT.md §4 for the validation rules.
cdn-ws and reality are the only deployment profiles where the
wire traffic — to a passive observer or an active prober — is
indistinguishable from a legitimate browser-to-CDN session, end to end:
- TLS 1.3 handshake with a real CDN's certificate (Cloudflare for
cdn-ws, the borrowed SNI's CDN forreality). - Chrome-fingerprinted ClientHello via uTLS (
fp=chromeis the default), bit-indistinguishable from a real Chrome browser. - No proxy-shaped framing inside TLS — Vision (REALITY) and standard WebSocket Upgrade (cdn-ws) are both well-known, well-distributed patterns, not custom layers a fingerprinting model can latch onto.
- Active probes hit a real CDN's backend. Probing the IP without
client credentials lands on Cloudflare (cdn-ws) or the borrowed CDN's
edge (reality), and gets a response indistinguishable from genuine CDN
traffic — see
docs/SECURITY.mdfor the threat model.
Every other protocol or transport combination listed in What we do not support has at least one of: fingerprintable framing, observable handshake patterns, cert-chain information leaks, or vulnerability to active probing. Operators have been blocking those for years; we do not implement them by design.
VLESS + WebSocket + TLS, fronted by a CDN (typically Cloudflare).
{
"protocol": "vless",
"settings": {
"vnext": [{
"address": "<cdn-edge-ip-or-host>",
"port": 443,
"users": [{ "id": "<uuid>", "encryption": "none" }]
}]
},
"streamSettings": {
"network": "ws",
"security": "tls",
"tlsSettings": {
"serverName": "<sni>",
"alpn": ["h2", "http/1.1"],
"fingerprint": "chrome",
"allowInsecure": false
},
"wsSettings": {
"path": "/?ed=2560",
"headers": { "Host": "<host>" }
}
}
}Validation rules:
addressis often a clean Cloudflare edge IP, not the apex domain.hostandserverNamecarry the real domain.fingerprintdefaults tochromeand shapes the ClientHello even with WS.allowInsecure: trueis rejected at parse time. The materializer always emitsfalse.- The early-data hint (
?ed=2560) inpathis preserved verbatim.
VLESS + Vision flow + REALITY over raw TCP.
{
"protocol": "vless",
"settings": {
"vnext": [{
"address": "<server-ip>",
"port": 443,
"users": [{
"id": "<uuid>",
"flow": "xtls-rprx-vision",
"encryption": "none"
}]
}]
},
"streamSettings": {
"network": "tcp",
"security": "reality",
"realitySettings": {
"serverName": "<dest-sni>",
"fingerprint": "chrome",
"publicKey": "<server-public-key>",
"shortId": "<short-id>",
"spiderX": ""
}
}
}Validation rules:
flowMUST equalxtls-rprx-vision(literal). Any other value is rejected.networkMUST equaltcp. REALITY combined withws,grpc,xhttp, orhttpupgradeis rejected at parse time.fingerprintwhitelist:chrome,firefox,safari,ios,android,edge,random. Unknown values are a hard failure, not a fallback.serverNameis the borrowed SNI of the destination the server steals TLS from (e.g.www.microsoft.com), not the proxy server's own domain.publicKeyis base64url, 43 chars (32-byte Curve25519, no padding).shortIdis hex, 0–16 chars, even length.
The classifier surfaces these as skipped with a structured reason;
nothing legacy lands in the pool. The reason each is rejected is
concrete, not aesthetic — every entry in this table has a known
weakness against the threat model in SECURITY.md.
| Rejected | Why |
|---|---|
| VMess | The custom AEAD header is fingerprint-able and the protocol's "no-header MD5" mode has known active-probe attacks. Deprecated upstream; ships with a non-zero attack surface. |
| Shadowsocks (SS), ShadowsocksR (SSR) | Pre-TLS era. Stream cipher modes leak length distributions; even SS-2022's AEAD-2022 streams have observable connection-setup patterns and have been flagged by deep-packet-inspection systems. |
| Trojan, Trojan-Go | Relies on a fixed password-hash prefix immediately after the TLS handshake; an active prober can replay an arbitrary GET, observe whether the server rejects vs. proxies, and confirm a Trojan endpoint. REALITY's steal-routine doesn't have this oracle. |
| VLESS + TLS direct (no REALITY, no CDN) | The TLS cert is for a domain you control. The cert chain reveals the operator's domain, the SNI is uniquely yours, and post-handshake traffic patterns are fingerprintable as VLESS. REALITY fixes all three by impersonating a CDN. |
| REALITY + WebSocket / gRPC / XHTTP / HTTPUpgrade | REALITY's whole value is that post-handshake bytes look like a normal browser→CDN session. Layering an HTTP-style framing on top adds a structure (HTTP/2 frames, WebSocket Upgrade, chunked padding) that fingerprinting models can pick up. xray-core lets you build these combinations; the audit cost is too high for the camouflage value lost. |
| mKCP / QUIC inbound / raw UDP transports | UDP-based, with timing and packet-size patterns that are fingerprintable. No major CDN exposes a comparable camouflage target. |
| HTTP / SOCKS as outbound transports | No encryption, no camouflage. Useful only as the local inbound listener for the user's apps, never as the upstream link to the proxy server. |
If any of those land in your subscription, they're surfaced verbatim in the "X servers skipped: …" line under each subscription, so you know exactly why the count is what it is.