Feature request: Add support for AmneziaWG 2.0 protocol

[argio11]

Hello!

I would like to request support for the AmneziaWG 2.0 protocol in sing-box.

AmneziaWG 2.0 is an enhanced version of WireGuard designed to bypass DPI (Deep Packet Inspection) and other network restrictions. It achieves this by obfuscating traffic so that it resembles other protocols, making it significantly more resistant to blocking by ISPs.

There is already an implementation available:
https://github.com/hoaxisr/amnezia-box

This is a fork of sing-box where support for AmneziaWG 2.0 has been added.

Request

If this protocol is of interest to you, could you please review the implementation in that repository? It may be possible to reuse or adapt parts of the code to integrate AmneziaWG 2.0 support into the official sing-box project.

Why this is useful

• Helps bypass DPI and censorship
• Extends WireGuard functionality with obfuscation
• Useful for users in restricted network environments
• Already has a working reference implementation

Thank you for your work and for considering this feature!

[Leadaxe]

Context, history, and a counterpoint on the standing rejection.

I'm a developer of an Android client built on sing-box, so this protocol's exclusion affects users I ship to. Aggregating prior threads first, then offering a counterargument.

Prior history

This protocol has been requested and rejected multiple times:

• Request for AmneziaWG Protocol Integration #1928 (Jul 2024) — closed in 1 day; labels: invalid, wontfix, spam. Maintainer's note: "Commercial and unprofessional. They do not have a correct understanding of China's Internet censorship."
• [Feature Request] Obfuscation for wireguard #2557 (Feb 2025) — "Obfuscation for wireguard"; canonical technical rationale here.
• PR Added support for AmneziaWG #2670 (Mar 2025) by @reflog — full working implementation depending on AmneziaWG Support wireguard-go#3. Closed in 2 days without merge, referencing [Feature Request] Obfuscation for wireguard #2557.
• [Features Request]: Add support for AmneziaWG, limit “1 uuid = 1 active connection”, users_list. #3159 (Jul 2025) — closed in 1 day; label: spam.

Working forks shipping it today: amnezia-vpn/amneziawg-go, hoaxisr/amnezia-box (a sing-box fork specifically tracking upstream for this one feature).

The standing position

From #2557:

"This is not obfuscation, it's just adding invalid data to the packet header. Not only is it not a standardized protocol, it's just a trick used by low-tech commercial companies. Since there are already projects like swgp-go that add real encryption to WireGuard, I won't consider this stuff."

I respect this stance and personally favor a purist core — sing-box's value comes precisely from its discipline about what gets in. The technical critique is correct: AmneziaWG is header-padding + magic-byte spoofing, not cryptographic obfuscation; swgp-go is architecturally cleaner.

Why I think this is worth re-examining

1. Every peer multi-protocol client has shipped it.

• mihomo — v1.18.9
• Hiddify
• Karing

When three competing projects converge on the same decision, it's a market signal, not a fad. As a downstream client developer, "we don't have AmneziaWG" is now a recurring user complaint with no good answer except "use a fork."

2. sing-box already accommodates non-standard DPI-evasion tricks at the TLS layer:

• Reality — TLS handshake forgery, not an RFC, designed against GFW
• ShadowTLS v3 — TLS mimicry
• VLESS xtls-vision — splice tricks against TLS-in-TLS detection

The "not a real protocol, just a trick" critique applies equally to these. AmneziaWG does the same thing at the UDP/WireGuard layer — and the threat it counters (RU TSPU's WireGuard-handshake fingerprinting) is a deployed, active block, not hypothetical.

3. swgp-go is good, but solves a different problem. swgp-go requires a paired server-side wrapper and is not interoperable with the existing AmneziaWG server ecosystem (already deployed by Amnezia, Hiddify panels, community providers in censored regions). Adding AmneziaWG would let sing-box clients connect to existing infrastructure; recommending swgp-go as the substitute asks operators to deploy parallel, non-interoperable infra — which most won't.

4. Cat-and-mouse is the cost of operating in this space. Yes, AmneziaWG parameters need rotation as DPI adapts. So do Reality fingerprints. So does every TLS mimicry technique. That's the cost of the censorship-evasion territory in general — not something AmneziaWG uniquely introduces. Opting in is part of being a tool for users behind state-level DPI; opting out leaves them to peer projects.

In closing

This is your project and your call, and I understand a "no" has been on the record since 2024. But that decision was made when the market hadn't yet converged. In 2026, three peer clients have shipped this, server-side infrastructure is deployed across multiple panels, and @reflog's reference PR (#2670) is still sitting there as a starting point. I'd like to respectfully ask you to reconsider.

[nekohasekai]

If you're cut from the same cloth as Hiddify and Karing, then take your AI slop and get lost.