sleep3r/mtproto.zig

Why this one? · Install · Update · Commands · Routing · Config · Dashboard · Build · Docker · Trust · Compatibility · FAQ

• You live somewhere Telegram is throttled or blocked and you just want it back.
• You're the one your family asks for help — and you want to protect your parents and friends with a link they tap once and never think about again.

It runs on your own server — your messages never pass through ours, and there's nothing to sign up for. Open source under MIT; the proxy deliberately never logs secrets or who connects.

A VPN announces itself — censors recognize the protocol and block it, and a whole-device VPN is slow and drains the battery. This looks like a plain HTTPS website, carries only Telegram, and the people you share it with install nothing: they tap one link and Telegram does the rest. Small enough for the cheapest VPS you can rent, it starts instantly, and there's nothing else to set up.

Most MTProto proxies are large, dependency-heavy, and use lots of memory. This one is different:

We chose Zig because it provides the raw performance and micro-footprint of C, but without the memory unsafety or build-system nightmares:

• No hot-path allocations: Connection-slot bookkeeping is pre-sized at startup; slot structs and their heavy buffers are created on first use and reused, so the relay path never allocates per byte. There is no garbage collector dropping frames under heavy load.
• Hermetic cross-compilation: Run zig build on macOS, and out comes a statically linked Linux binary. No Docker, no glibc version mismatches.
• Comptime: Costly operations like protocol definition mapping, endianness conversions, and bilingual string lookup for mtbuddy are resolved during compilation, giving instant startup times.

You don't need to understand any of the names below — the default install turns them all on for you. Under the hood, the proxy stacks more anti-censorship techniques than any other MTProto proxy, and keeps adapting as the blocks get smarter:

MiddleProxy is required for promotion tags and for media on non-Premium accounts. Without it, photos, videos, stories, and other media on non-Premium accounts should be treated as unavailable rather than flaky. Telegram calls are not supported by this proxy: Telegram only routes calls through SOCKS-style paths, and exposing SOCKS traffic cannot be masked by mtproto.zig as normal HTTPS.

All installation, updates, and management are done through mtbuddy — a native Zig CLI that ships alongside the proxy.

curl -fsSL https://raw.githubusercontent.com/sleep3r/mtproto.zig/main/deploy/bootstrap.sh | sudo bash

# Explicitly allow unsigned bootstrap mode (not recommended)
curl -fsSL https://raw.githubusercontent.com/sleep3r/mtproto.zig/main/deploy/bootstrap.sh | sudo bash -s -- --insecure
# or: MTPROTO_INSECURE=1

This downloads the latest mtbuddy binary, verifies minisign signature + SHA-256 checksum from the GitHub Release, and prints the next-step install command. Then install the proxy:

# Minimal — auto-generates a secret, enables all DPI bypass modules
sudo mtbuddy install --port 443 --domain rutube.ru --yes

# Bring your own secret and username
sudo mtbuddy install --port 443 --domain rutube.ru --secret <32-hex> --user alice --yes

# Disable all DPI modules (bare proxy only)
sudo mtbuddy install --port 443 --domain rutube.ru --no-dpi --yes

# Install using an existing config file (auto-maps port and domain)
sudo mtbuddy install --config /path/to/config.toml --yes

# Explicitly allow unsigned mode (not recommended)
sudo mtbuddy install --insecure --port 443 --domain rutube.ru --yes

At the end, mtbuddy prints a ready-to-use tg:// connection link.

Share it with someone you love. Send them this, with the link: "I set up a private door to Telegram for us. Tap this link, choose Connect, and Telegram will work again — nothing to install, nothing to pay, and it's only ours."

If you prefer to be walked through the setup:

sudo mtbuddy --interactive

1. Downloads the pre-built proxy binary from GitHub Releases (auto-detects CPU: x86_64_v3 → x86_64 → aarch64)
2. Generates a random secret (or uses --secret)
3. Creates a systemd service (mtproto-proxy)
4. Opens the port in ufw (if active)
5. Applies TCPMSS=88 iptables rules
6. Sets up Nginx masking + nfqws TCP desync (unless --no-dpi)
7. Prints tg:// link

⚠️ Pick --domain once. The tg:// links embed tls_domain, so changing it on a live deployment (including via mtbuddy setup masking --domain …) invalidates every link you've already shared. See ARCHITECTURE.md / COMPATIBILITY.md.

# Update to latest release (verifies minisign + checksum, checks CPU compat, auto-rollback on failure)
sudo mtbuddy update

# Pin to a specific version
sudo mtbuddy update --version v0.11.1

# Explicitly allow unsigned mode (not recommended)
sudo mtbuddy update --insecure

config validate, doctor, and print-effective report common TOML mistakes (strings without quotes, duplicate sections/keys, and unclosed quotes) with a line number and a correction hint. These commands do not modify the file. For example, put public_ip = "proxy.example.com" inside the existing [server] section; do not add a second [server]. This compatibility check supplements the runtime's semantic validation; it is not a full TOML implementation.

If the dashboard cannot parse the config, it shows the location and keeps updating system/process telemetry. Config-dependent panels and management actions are unavailable until the file is corrected; the next refresh loads the repaired file.

# Show proxy and module status
sudo mtbuddy status

# Validate and inspect config
sudo mtbuddy config validate
sudo mtbuddy config doctor
sudo mtbuddy config doctor --network
sudo mtbuddy config print-effective

# Print Telegram proxy links from config.toml (FakeTLS ee by default; +dd when fake_tls_only=false; sensitive output)
sudo mtbuddy links
sudo mtbuddy links --server proxy.example.com --config /opt/mtproto-proxy/config.toml

# Generate a fresh 32-hex user secret
mtbuddy secret

# Hot-reload config (SIGHUP, reloadable settings only)
sudo mtbuddy reload

# Setup DPI modules after the fact
sudo mtbuddy setup masking --domain rutube.ru
sudo mtbuddy setup nfqws
sudo mtbuddy setup recovery

# Optional kernel-level per-IP SYN rate-limit (anti-flood, OFF by default).
# Drops abusive first-SYN bursts in the kernel BEFORE accept(); separate systemd
# oneshot so CAP_NET_ADMIN is never granted to the proxy. Default preset is CGNAT-safer.
sudo mtbuddy setup syn-limit --preset soft   # soft 2/s·5 | medium 1/s·3 | hard 1/s·1
sudo mtbuddy setup syn-limit --status
sudo mtbuddy setup syn-limit --remove

# WEB proxy relay for Telegram Desktop 7.1+ (MTProto carried inside an ordinary HTTPS site).
# Needs a domain you own, with an A record pointing at this host.
sudo mtbuddy setup web --domain relay.example.com
sudo mtbuddy setup web --domain relay.example.com --mode behind   # you terminate TLS (CDN, spare IP)
sudo mtbuddy setup web --remove

# Install web monitoring dashboard
sudo mtbuddy setup dashboard
# Remove just the dashboard (leaves the proxy running)
sudo mtbuddy setup dashboard --remove

# VPN tunnel (for servers where Telegram DCs are blocked)
sudo mtbuddy setup tunnel /path/to/awg0.conf
sudo mtbuddy setup tunnel 'vpn://...'
sudo mtbuddy setup tunnel --iface awg1 /path/to/awg1.conf

# Egress from a VPN share-link — clean, hard-to-block upstream for the proxy→Telegram hop.
#   vless:// vmess:// trojan:// ss://  -> local sing-box TUN tunnel (type=tunnel, exactly like
#                                        AmneziaWG; VLESS-Reality camouflages the hop as real TLS).
#   wireguard://                       -> native kernel WG tunnel (same as `setup tunnel`).
#   multiple links                     -> a urltest failover pool.
sudo mtbuddy setup egress 'vless://...@host:443?security=reality&pbk=...&sni=...&flow=xtls-rprx-vision'
# vless / vmess / trojan / ss / hysteria2 — hysteria2 is QUIC, so measure UDP to the endpoint first
sudo mtbuddy setup egress 'wireguard://<privkey>@host:51820?publickey=...&address=10.0.0.2/32'

# IPv6 hopping
sudo mtbuddy ipv6-hop --check
sudo mtbuddy ipv6-hop --auto --prefix 2a01:abcd:ef00:: --threshold 5

# Update Cloudflare DNS A record
sudo mtbuddy update-dns 1.2.3.4 proxy.example.com

# Full help
mtbuddy --help
mtbuddy --lang ru --help

sudo systemctl status mtproto-proxy
sudo journalctl -u mtproto-proxy -f
sudo systemctl reload mtproto-proxy   # SIGHUP hot-reload (where possible)
sudo systemctl restart mtproto-proxy

The proxy supports multiple ways to route outgoing connections to Telegram DC servers.

If your VPS is in a region where Telegram DCs are blocked at the network level, you can route proxy traffic through a VPN tunnel pool with explicit socket policy routing. The proxy runs in the host namespace; only sockets marked by the proxy (SO_MARK=200) are routed through table 200. mtbuddy keeps that table pointed at the first healthy tunnel in the configured order.

Currently supported VPN types:

• AmneziaWG — DPI-resistant WireGuard fork (recommended for Russia/Iran)
• WireGuard — standard kernel WireGuard (via setup egress 'wireguard://…' or a WG .conf through the AmneziaWG-compatible backend)

Client → mtproto-proxy (host namespace)
                     │
                SO_MARK=200
                     │
        Linux policy routing table 200
                     │
          awg0 / awg1 / ... (pool)
                     │
             Telegram DC servers

sudo mtbuddy setup tunnel /path/to/awg0.conf
# or paste an Amnezia share link directly
sudo mtbuddy setup tunnel 'vpn://...'

# Add or replace a specific pool member
sudo mtbuddy setup tunnel --iface awg1 /path/to/awg1.conf

In the interactive mtbuddy menu, tunnel setup first asks for the VPN type (AmneziaWG or standard WireGuard), then shows the current tunnel pool. Choose Create new tunnel to append the next free awgN, or choose an existing interface to replace that pool member's config.

Configs mtbuddy did not create are never deleted. mtbuddy writes into /etc/amnezia/amneziawg/<iface>.conf — AmneziaWG's own directory, where your configs live too — so every config it writes carries a # Generated by mtbuddy banner, and only those are removed by mtbuddy uninstall or tunnel delete. Point --iface awg1 at an interface's existing config and mtbuddy adopts it in place: it normalises the file (drops DNS, adds Table = off) and adds the name to the pool, but leaves it, and the running interface, alone when you uninstall. Anything left behind is listed by path at the end of the uninstall so you can decide about it yourself.

mtbuddy keeps [general].use_middle_proxy unchanged and only configures transport ([upstream].type = "tunnel"). After setup, it installs mtproto-tunnel-pool.timer, validates policy routes (mark 200) to Telegram DC ranges, and prints operational commands. The pool controller probes Telegram through each tunnel and rewrites table 200 with ip route replace; automatic failover does not restart mtproto-proxy.

You can also explicitly configure the tunnel interface in config.toml:

[upstream]
type = "tunnel"

[upstream.tunnel]
interface = "awg0"
interfaces = ["awg0", "awg1"]
pinned_interface = ""   # optional; empty means priority auto-failback

Route DC connections through an external SOCKS5 proxy. Supports RFC 1928 auth.

[upstream]
type = "socks5"

[upstream.socks5]
host = "127.0.0.1"
port = 1080
username = "admin"    # optional, omit for no-auth
password = "secret"

Route DC connections through an HTTP CONNECT proxy. Supports Basic auth.

[upstream]
type = "http"

[upstream.http]
host = "127.0.0.1"
port = 8080
username = "admin"    # optional, omit for no-auth
password = "secret"

Note: DC-bound relay traffic and MiddleProxy metadata refreshes (getProxyConfig / getProxySecret) use the configured upstream. Mask (camouflage) connections always go direct.

Dependency note: the "zero dependencies" claim holds for the default auto/direct egress. With socks5, http, or tunnel upstream modes, MiddleProxy metadata refresh shells out to curl, so curl must be installed on the host (the standard installer pulls it in).

Config lives at /opt/mtproto-proxy/config.toml. MTBuddy generates it on install; you can edit it manually and restart:

[general]
use_middle_proxy = true   # ME mode for promo-channel parity

[upstream]
type = "auto"            # auto | direct | tunnel | socks5 | http
# allow_direct_fallback = false   # fail-closed by default for socks5/http misconfig

[server]
port = 443
# public_ip = "proxy.example.com"   # Inbound IP/domain used in client links
# public_port = 443                 # Link port when behind HAProxy/Nginx
# middle_proxy_nat_ip = "203.0.113.10"   # Override outbound IPv4 seen by Telegram MiddleProxy (auto-detected through the egress otherwise)
max_connections = 512
# workers = 1            # SO_REUSEPORT epoll workers: 1 = single-threaded (default); 0 = one per CPU; N spreads load across cores
idle_timeout_sec = 120
# client_silence_close_sec = 0   # Close a relay whose server reply went unanswered by the client for N sec (breaks an iOS bad_salt wedge where "Updating" hangs ~90-120s) → instant clean reconnect. 0 = off; best-effort, can occasionally close a healthy conn — ~10-15 if you enable it
handshake_timeout_sec = 15
# dc_connect_timeout_sec = 10   # Per-endpoint TCP-connect deadline to a Telegram DC; fails a black-holed endpoint fast so failover advances to the next one (within handshake_timeout_sec). 0 = off; never affects healthy connects (<1s)
graceful_shutdown_timeout_sec = 15
log_level = "info"        # debug | info | warn | err
rate_limit_per_subnet = 0   # 0 = disabled (default; avoids carrier-NAT false positives). Set e.g. 30 for non-NAT hosts
handshake_flood_guard_enabled = false
handshake_flood_guard_threshold = 20
handshake_flood_guard_window_sec = 30
handshake_flood_guard_block_sec = 120
tag = ""                  # Optional: promotion tag from @MTProxybot

[censorship]
tls_domain = "rutube.ru"
mask = true
# mask_target = "host.docker.internal" # Optional: custom masking backend host (Docker/remote Nginx)
mask_port = 8443          # 8443 = local Nginx backend (what mtbuddy installs); 443 = front the real tls_domain (opt-in, single-round-x25519 domains only)
fast_mode = true          # Recommended: delegates S2C AES to the DC, saves CPU/RAM
drs = true                # Dynamic Record Sizing (mimics Chrome/Firefox)

[access.users]
alice = "00112233445566778899aabbccddeeff"
bob   = "ffeeddccbbaa99887766554433221100"

[access.direct_users]
alice = true   # bypass MiddleProxy for this user

Generate a secret: mtbuddy secret or openssl rand -hex 16

Print client links explicitly: sudo mtbuddy links. By default it prints FakeTLS (ee...domain) links only; it also prints secure padded (dd...) links when the dd transport is enabled (fake_tls_only = false). Runtime proxy logs intentionally hide secrets and proxy links.

The dd ("secure"/padded) transport is rejected by default ([censorship].fake_tls_only = true) — it is plain obfuscated MTProto with no TLS disguise, directly fingerprintable as MTProto by DPI. By default the proxy accepts only FakeTLS (ee), and mtbuddy links prints only ee links. To hand out dd links (lower-DPI / compatibility scenarios), set fake_tls_only = false. See THREAT_MODEL.md.

Both abuse guards are off by default so large carrier-NAT, VPN-egress, or shared-office networks (many legitimate clients behind one source IP/subnet) aren't false-positived and blocked together: the per-subnet new-connection rate limit (rate_limit_per_subnet = 0) and the exact-IP handshake flood guard (handshake_flood_guard_enabled = false). Access is already gated by the per-user secret, the global handshake-inflight budget, and max_connections. On a single-tenant / non-NAT host under real abuse, turn them on: set rate_limit_per_subnet (e.g. 30) and handshake_flood_guard_enabled = true (tune handshake_flood_guard_threshold / window / block).

Both of the above run after accept() (they're in-proxy). For an additional kernel-level layer that drops abusive first-SYN bursts before they cost a socket/accept(), there's an optional per-source-IP SYN rate-limiter: sudo mtbuddy setup syn-limit --preset soft. It's an iptables hashlimit rule installed as a separate mtproto-syn-limit.service oneshot, so CAP_NET_ADMIN is never granted to the proxy. Also off by default and subject to the same carrier-NAT caveat (the soft 2/s·burst-5 preset is the CGNAT-safer default); --remove to undo, status + drop counter in mtbuddy status. Re-run it after changing the proxy port.

Telegram Desktop 7.1 added a fourth proxy type, WEB. It is an ordinary MTProxy whose carrier is a browser: the client opens no MTProto socket at all. A hidden native WebView loads https://<your-domain>/?bridge=<capability>, and the page shuttles multiplexed frames over a same-origin WebSocket to a relay, which dials a stock MTProxy — this one — for every logical stream.

Telegram Desktop
  → hidden WebView (WKWebView / WebView2 / WebKitGTK)
  → https://relay.example.com/           ← a real browser TLS handshake to a real site
  → mtproto-web-relay
  → mtproto-proxy
  → Telegram

What a censor sees is a genuine browser fingerprint completing a genuine TLS handshake with a CA-chained certificate, then speaking HTTP — because it is one. There is no FakeTLS ServerHello to mismatch and no MTProto handshake on the wire.

Set it up:

sudo mtbuddy setup web --domain relay.example.com
sudo mtbuddy links          # now also prints tg://webproxy links

Existing WEB links keep working after updating. The bridge now issues a short-lived carrier token instead of repeating the link capability in WebSocket URLs. Open pages may reconnect once when the relay restarts.

For ordinary visitors, serve your own static site by setting [web].public_dir to a readable directory containing index.html and its assets, then restart the proxy/relay. The former generated placeholder has been removed: without public_dir, public requests return a plain 404 while authenticated WEB links still work. Files are read once at startup (up to 256 files, 2 MiB each, 16 MiB total); dotfiles and symlinks are not served. Use exact file paths for asset links. The bridge is a separate minimal document and does not execute scripts from your public site.

The WEB health timer probes [web].listen (host is also accepted). A completed local HTTP response, including 404, keeps the relay running; only three failed local transfers trigger a relay restart. The public HTTPS check accepts 2xx, 3xx, and 404, while other statuses and TLS/network failures report an error without restarting the relay. These are liveness checks, not authenticated end-to-end tests. When upgrading from a version with the old health script, run sudo mtbuddy update --force once after the upgrade: the first update finishes using the previous updater's embedded scripts, even after replacing its binary.

When the TLS terminator is on another host, list its IP in [web].trusted_http_sources; this is separate from relay_sources, which controls relay-to-MTProto connections. Configure the terminator to overwrite the client-IP header, and disable request URI/header logging, including error logs.

You need a domain you control with an A record pointing at this host; mtbuddy issues a Let's Encrypt certificate over HTTP-01 (so port 80 is opened) and installs a renewal hook.

Bring your own certificate if you already have one — acme.sh, a wildcard, a Cloudflare Origin certificate with Cloudflare in front. mtbuddy then issues nothing, installs no renewal hook, and points the vhost at your files; the paths are persisted in [web] so mtbuddy update keeps them:

sudo mtbuddy setup web --domain relay.example.com \
  --cert /etc/ssl/relay/fullchain.pem --key /etc/ssl/relay/privkey.pem

The certificate still has to be one the client's WebView will accept — that means a public CA, or a private one only where whatever sits in front of you is what presents it to the client. A self-signed certificate served straight to clients is refused by the WebView and the bridge never loads.

Two topologies, both supported:

How it interacts with the rest of the proxy:

• Same users, same secrets. A WEB link carries the same 16-byte secret as the FakeTLS link, encoded as dd… rather than ee… — Telegram Desktop reports an ee (FakeTLS) secret as unsupported for a WEB proxy, because the relay is a raw byte pipe that adds no TLS-emulation record. Everyone in [access.users] gets both links; per-user connection caps and expirations apply unchanged. [access.user_max_ips] needs [web].mask_backend to see real browser addresses — without it every WEB client reaches the proxy as 127.0.0.1, and those connections are exempt from the quota rather than crowding into one slot.
• fake_tls_only stays on. The public internet still never gets a dd responder to fingerprint. Only the relay's own source address (loopback by default, plus anything in [web].relay_sources) is allowed through that gate, and the decision is made from the address accept() reported — never from a client-supplied PROXY header.
• WEB-only ([web].only) takes that one step further: the proxy stops answering direct MTProto altogether. A FakeTLS or dd handshake from anyone but the relay is handed to the masking backend on the same code path a wrong secret takes, so a probe cannot tell the difference — see below.
• Real client IPs are preserved. The relay prefixes each backend connection with a PROXY protocol header carrying the browser's address, so per-IP accounting, the flood guard and Telegram itself see the actual client rather than 127.0.0.1.
• Probe resistance. The bridge capability is HMAC-SHA256(user secret), so a visitor who cannot present one derived from a configured secret never sees the bridge. Ordinary requests use the operator-owned public site (or a plain 404 when unset). Carrier credentials expire after two minutes and cannot resume an adopted session.
• Capacity. Each WEB client costs one masked connection plus one connection per logical MTProto stream against [server].max_connections. Budget roughly 3–5× a direct client, and raise max_connections accordingly.

Limitations: desktop only (7.1+; Android has the same bridge contract but no released UI yet), no voice calls, and the carrier is more RTT-sensitive than a direct connection. It is a fallback for when the direct link is blocked, not a replacement for it.

Some networks blocklist an IP the moment a Telegram client connects to it directly — the proxy works, and then the whole host stops answering on every TCP port for days. If that is what you are seeing, stop offering the direct door:

sudo mtbuddy setup web --domain relay.example.com --only
sudo mtbuddy links          # now prints tg://webproxy links and nothing else

setup web --only verifies HTTPS, the authenticated bridge and a real MTProto reply before disabling direct links. Python 3 is required for this deployment check; a missing interpreter or failed check leaves the previous WEB-only setting unchanged.

With [web].only the data plane answers MTProto for the relay and no one else. Everyone else — including a real Telegram client holding a valid ee link — is forwarded to the masking backend, byte for byte the same way a wrong secret is, so the proxy is indistinguishable from the website it fronts. mtbuddy links prints only the WEB links, because they are the only ones that still connect.

What it costs:

• Every ee/dd link you have already handed out stops working. There is no partial rollout; turn it off with sudo mtbuddy setup web --no-only.
• Desktop 7.1+ only — phones cannot use a WEB link at all.
• [web].max_sessions becomes your user ceiling, since every client now arrives through the relay.
• It needs [censorship].mask = true. Without a cover site the refusal is a bare close, which is a cleaner signal to a prober than answering MTProto was — mtbuddy refuses to enable WEB-only without the local masking backend for exactly that reason.

mtproto_web_only_masked_total counts what the gate turned away.

A lightweight web dashboard (~30 MB RAM) shows live connections, CPU/memory, network throughput, proxy stats, tunnel pool health/failover state, user management, and streaming logs.

The dashboard is embedded directly into the mtbuddy binary — no extra files needed.

# Install the dashboard on the server
sudo mtbuddy setup dashboard

# Open via SSH tunnel (binds to 127.0.0.1:61208 by default)
ssh -L 61208:localhost:61208 root@<server_ip>
# → http://localhost:61208

The dashboard requires HTTP Basic auth (username: any; password auto-generated at /opt/mtproto-proxy/monitor/dashboard.token — cat it on the server). It is a root-privileged control plane, so keep it on the loopback/SSH-tunnel path and never expose plain HTTP to the internet — front it with HTTPS + a reverse proxy if you must.

Local dashboard history can be disabled independently of Prometheus/Grafana. Set traffic_history_enabled = false in the existing [monitor] section and leave [metrics].enabled = true. At the next sample (normally within 30 seconds), the dashboard stops scraping metrics and writing history; it keeps the database and shows that history is disabled. Set the option back to true to resume. The first sample after startup or resume establishes a baseline, so traffic from a collection pause is not added later.

mtproto-proxy can expose an embedded Prometheus-compatible metrics endpoint on a dedicated port.

For a complete Docker-based monitoring stack with mtproto-zig, Prometheus, Grafana, and an importable dashboard, see hack/docker/README.md.

[metrics]
enabled = true
host = "127.0.0.1"
port = 9400

The endpoint is plaintext HTTP and serves:

GET /metrics

Typical Docker usage:

docker run --rm \
  -p 443:443 \
  -p 9400:9400 \
  -v "$PWD/config.toml:/etc/mtproto-proxy/config.toml:ro" \
  mtproto-zig

It exposes proxy counters plus process metrics such as RSS, virtual memory, CPU time, and open file descriptors.

Requires Zig 0.16.0.

git clone https://github.com/sleep3r/mtproto.zig.git
cd mtproto.zig

make build      # cross-compile ReleaseFast binaries for Linux x86_64_v3+aes
make test       # run Zig tests
make e2e        # run E2E/integration harness
make fmt        # format Zig sources
make deploy     # build + deploy to SERVER (see Makefile)
make dashboard  # SSH tunnel for web dashboard (localhost:61208)

# optional performance checks
zig build bench
zig build soak

Release builders can override the default pinned minisign key if needed:

zig build -Dminisign_pubkey=RW... -Doptimize=ReleaseFast -Dtarget=x86_64-linux

Cross-compile for Linux from macOS:

zig build -Doptimize=ReleaseFast -Dtarget=x86_64-linux -Dcpu=x86_64_v3+aes
scp zig-out/bin/mtproto-proxy root@<SERVER>:/opt/mtproto-proxy/

Docker support is provided for testing, packaging experiments, and simple deployments where only the proxy binary is needed. The project is primarily designed for a native Linux host managed by mtbuddy: DPI modules, tunnel pool failover, policy routing, Nginx masking, nfqws, and recovery timers are host-level integrations and are not fully represented by the container.

docker pull ghcr.io/sleep3r/mtproto.zig:latest

docker run --rm \
  -p 443:443 \
  -v "$PWD/config.toml:/etc/mtproto-proxy/config.toml:ro" \
  ghcr.io/sleep3r/mtproto.zig:latest

MiddleProxy media/promo traffic is sensitive to the outbound source IP:port used in its encrypted handshake. The egress IPv4 is auto-detected through the configured upstream — direct probes the host, socks5/http probes through the proxy (so it learns the SOCKS exit IP), and tunnel probes through the tunnel interface — so SOCKS/tunnel egress + ad-tag work without manual setup. For Docker deployments that need MiddleProxy, prefer host networking (--network host) or a native mtbuddy install. [server].public_ip is only the inbound address shown to clients; set [server].middle_proxy_nat_ip only to override when auto-detection can't see the real egress. Bridge or remote NAT that rewrites source ports can still break MiddleProxy handshakes.

Build locally:

docker build -t mtproto-zig .
# multi-arch
docker buildx build --platform linux/amd64,linux/arm64 -t your-registry/mtproto-zig:latest --push .

Published linux/amd64 images are built with -Dcpu=x86_64+aes — the x86-64 baseline plus AES-NI for fast crypto. This requires a CPU with AES-NI (virtually all VPS hardware since ~2011); on a QEMU/KVM guest exposing a CPU model without aes, build your own image with -Dcpu=x86_64 to avoid Illegal instruction.

For production censorship-bypass deployments, prefer the native mtbuddy install flow. OS-level mitigations (iptables TCPMSS, nfqws, tunnel policy routing, masking/recovery units) are not applied inside the container; only the proxy binary runs there.

• SECURITY.md - vulnerability reporting policy and response process
• THREAT_MODEL.md - security goals, non-goals, adversary model, residual risks
• CONTRIBUTING.md - dev workflow (fmt/test/e2e/bench) and PR expectations
• CHANGELOG.md - release history
• LICENSE - MIT license terms

Repository governance:

• .github/CODEOWNERS
• issue templates under .github/ISSUE_TEMPLATE

For a full model see THREAT_MODEL.md. Quick operational summary:

• Known limitations
  • This is a transport-hardening proxy, not an anonymity network.
  • Bypass quality can degrade as DPI strategies evolve.
  • Dashboard/metrics are plaintext by default; do not expose publicly without auth/TLS.
  • Telegram calls do not work through this proxy. Calls require Telegram's SOCKS-style call path, which is outside the MTProto/TLS-masking model and cannot be disguised cleanly as normal HTTPS here.
  • Without MiddleProxy ([general].use_middle_proxy = true), media on non-Premium accounts will not load. MiddleProxy is required for photos, videos, stories, and promotion tags.
• Region-specific caveats
  • ISP behavior differs by country/region; configs are not universally portable.
  • IPv6 and AAAA handling vary heavily across providers and can impact iOS/Desktop connection latency.
  • Tunnel routing depends on host policy routing and allowed VPN protocols in that region.
• Telegram client compatibility
  • Official Telegram Android/iOS/Desktop: expected to work on current releases.
  • Third-party clients: best effort only.
• Kernel/OS compatibility matrix
  • Linux x86_64: supported (primary target)
  • Linux aarch64: supported
  • Docker on Linux: supported with caveats (OS-level DPI modules are host-side)
  • macOS/Windows runtime: not supported (Linux runtime target only)
• What can break after Telegram/DC changes
  • MiddleProxy metadata and endpoint behavior
  • handshake expectations in newer Telegram clients
  • DC/media routing edge cases (for example DC203 behavior)

1. AAAA record exists but IPv6 doesn't work on the server. DNS has an AAAA → iOS tries IPv6 first → timeout → slow fallback to IPv4. Fix: remove AAAA until IPv6 routing is fully configured.

dig +short proxy.example.com AAAA
ip -6 route

2. Home Wi-Fi blocks the server's IPv4. Mobile networks usually work (they use IPv6). Home routers often block the destination IPv4. Fix: enable IPv6 Prefix Delegation (IA_PD) on your router.

3. VPN is dropping MTProto traffic. Commercial VPNs often DPI and drop proxy traffic. Fix: switch VPN protocol, or use a self-hosted AmneziaWG.

4. Co-located WireGuard/Docker on the same server. Docker's bridge drops packets from VPN subnet. Fix: iptables -I DOCKER-USER -s 172.29.172.0/24 -p tcp --dport 443 -j ACCEPT

5. DC203 media resets on non-premium clients. Check logs: journalctl -u mtproto-proxy | grep -E "dc=203|Middle". The proxy auto-refreshes DC203 metadata from Telegram on startup. If core.telegram.org is unreachable, it uses bundled fallback addresses. With [upstream].type = "socks5" or "http", metadata refreshes use that upstream; run sudo mtbuddy config doctor --network to verify the proxy endpoint and Telegram metadata fetch path.

MIT © 2026 Aleksandr Kalashnikov