GitHub - DNSCrypt/doh-server: Fast, mature, secure DoH and ODoH server proxy written in Rust. Previously known as doh-proxy and rust-doh.

A fast and secure DoH (DNS-over-HTTPS) and ODoH (Oblivious DoH) server.

doh-proxy is written in Rust, and has been battle-tested in production since February 2018. It doesn't do DNS resolution on its own, but can sit in front of any DNS resolver in order to augment it with DoH support.

Features

DNS-over-HTTPS (DoH) - Encrypts DNS queries using HTTPS
JSON API Support - Compatible with Google DNS-over-HTTPS JSON API format
Oblivious DoH (ODoH) - Provides additional privacy by hiding client IP addresses
EDNS Client Subnet - Forward client IP information to upstream resolvers for geo-optimized responses
High Performance - Built with Rust and Tokio for excellent performance
Flexible Deployment - Can run standalone with built-in TLS or behind a reverse proxy
Production Ready - Battle-tested in production environments since 2018
Multiple IP Support - Supports multiple external IP addresses for load balancing
Automatic Certificate Reloading - No downtime when updating TLS certificates
Configurable Caching - TTL management with configurable min/max values

Installation

Option 1: precompiled binaries for Linux

Precompiled tarballs and Debian packages for Linux/x86_64 can be downloaded here.

Option 2: from source code

This requires the rust compiler to be installed.

With built-in support for HTTPS (default):

cargo install doh-proxy

Without built-in support for HTTPS:

cargo install doh-proxy --no-default-features

Quick Start

Basic Usage

# Simple setup with a local DNS resolver
doh-proxy -H 'doh.example.com' -u 127.0.0.1:53

# With a specific public IP address
doh-proxy -H 'doh.example.com' -u 127.0.0.1:53 -g 203.0.113.1

# With built-in TLS support
doh-proxy -H 'doh.example.com' -u 127.0.0.1:53 -i /path/to/cert.pem -I /path/to/key.pem

Complete Usage Reference

USAGE:
    doh-proxy [FLAGS] [OPTIONS]

FLAGS:
    -O, --allow-odoh-post      Allow POST queries over ODoH even if they have been disabed for DoH
    -K, --disable-keepalive    Disable keepalive
    -P, --disable-post         Disable POST queries
    -h, --help                 Prints help information
    -V, --version              Prints version information

OPTIONS:
    -E, --err-ttl <err_ttl>                          TTL for errors, in seconds [default: 2]
    -H, --hostname <hostname>                        Host name (not IP address) DoH clients will use to connect
    -l, --listen-address <listen_address>            Address to listen to [default: 127.0.0.1:3000]
    -b, --local-bind-address <local_bind_address>    Address to connect from
    -c, --max-clients <max_clients>                  Maximum number of simultaneous client connections, TLS handshakes included [default: 512]
    -C, --max-concurrent <max_concurrent>            Maximum number of concurrent requests per client [default: 16]
    -X, --max-ttl <max_ttl>                          Maximum TTL, in seconds [default: 604800]
    -T, --min-ttl <min_ttl>                          Minimum TTL, in seconds [default: 10]
    -p, --path <path>                                URI path [default: /dns-query]
    -g, --public-address <public_address>            External IP address(es) DoH clients will connect to (can be specified multiple times)
    -j, --public-port <public_port>                  External port DoH clients will connect to, if not 443
    -u, --server-address <server_address>            Address to connect to [default: 9.9.9.9:53]
    -t, --timeout <timeout>                          Time budget of a request, in seconds (1 to 3600). Idle connections are closed after this delay, sooner when the server is busy [default: 10]
    -I, --tls-cert-key-path <tls_cert_key_path>
            Path to the PEM-encoded secret keys (only required for built-in TLS)

    -i, --tls-cert-path <tls_cert_path>
            Path to the PEM/PKCS#8-encoded certificates (only required for built-in TLS)

    --enable-ecs                              Enable EDNS Client Subnet
    --ecs-prefix-v4 <ecs_prefix_v4>         IPv4 prefix length for EDNS Client Subnet [default: 24]
    --ecs-prefix-v6 <ecs_prefix_v6>         IPv6 prefix length for EDNS Client Subnet [default: 56]

Example Configurations

Basic setup with custom DNS resolver:

doh-proxy -H 'doh.example.com' -u 8.8.8.8:53 -g 203.0.113.1

Multiple IP addresses for load balancing:

doh-proxy -H 'doh.example.com' -u 127.0.0.1:53 -g 203.0.113.1 -g 203.0.113.2 -g 2001:db8::1

This generates separate DNS stamps for each IP address, allowing clients to connect via any of them.

Production setup with TLS and custom limits:

doh-proxy -H 'doh.example.com' \
          -u 127.0.0.1:53 \
          -l 0.0.0.0:443 \
          -i /etc/letsencrypt/live/doh.example.com/fullchain.pem \
          -I /etc/letsencrypt/live/doh.example.com/privkey.pem \
          -c 1000 \
          -C 32

Behind a reverse proxy (nginx/Caddy):

doh-proxy -H 'doh.example.com' -u 127.0.0.1:53 -l 127.0.0.1:3000

Deployment Architectures

Behind a Reverse Proxy (Recommended)

The recommended deployment is behind a TLS termination proxy such as nginx, Caddy, HAProxy, or a CDN. This allows:

Sharing port 443 with existing web services
Leveraging existing TLS certificate management
Using HTTP/2 and HTTP/3 features from the proxy
Better DDoS protection and rate limiting

Example with nginx:

server {
    listen 443 ssl http2;
    server_name doh.example.com;

    ssl_certificate /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;

    location /dns-query {
        proxy_pass http://127.0.0.1:3000;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
    }
}

Example with Caddy:

doh.example.com {
    reverse_proxy /dns-query localhost:3000
}

Standalone with Built-in TLS

For simpler deployments or when running on separate infrastructure:

doh-proxy -H 'doh.example.com' \
          -u 127.0.0.1:53 \
          -l 0.0.0.0:443 \
          -i /path/to/fullchain.pem \
          -I /path/to/privkey.pem

Certificate Requirements:

Certificates and keys must be in PEM/PKCS#8 format
Can be stored in the same file or separately
Automatically reloaded when changed (no restart needed)

If using ECDSA certificates that start with -----BEGIN EC PRIVATE KEY-----, convert to PKCS#8:

openssl pkcs8 -topk8 -nocrypt -in example.key -out example.pkcs8.pem

Using Let's Encrypt with acme.sh:

# Install acme.sh
curl https://get.acme.sh | sh

# Get certificates
acme.sh --issue -d doh.example.com --webroot /var/www/html

# Run doh-proxy with Let's Encrypt certificates
doh-proxy -H 'doh.example.com' \
          -u 127.0.0.1:53 \
          -i ~/.acme.sh/doh.example.com/fullchain.cer \
          -I ~/.acme.sh/doh.example.com/doh.example.com.key

Note: Once HTTPS is enabled, HTTP connections will not be accepted. A sample self-signed certificate localhost.pem is available for testing.

Integration Examples

With Encrypted DNS Server

Encrypted DNS Server can handle both DNSCrypt and DoH on the same port:

# In encrypted-dns-server.toml
[tls]
upstream_addr = "127.0.0.1:3000"

This provides:

Support for both DNSCrypt and DoH protocols
Built-in DNS caching
Server-side filtering
Connection reuse and DDoS protection

With nginx

location /dns-query {
    proxy_pass http://127.0.0.1:3000;
    proxy_set_header Host $host;
    proxy_set_header X-Real-IP $remote_addr;
    proxy_http_version 1.1;
    proxy_set_header Connection "";
}

The example above opens a fresh backend connection for each request. To enable connection reuse, use an upstream block with keepalive:

upstream doh_backend {
    server 127.0.0.1:3000;
    keepalive 16;
}

location /dns-query {
    proxy_pass http://doh_backend;
    proxy_http_version 1.1;
    proxy_set_header Connection "";
    proxy_next_upstream error timeout non_idempotent;
}

doh-proxy closes connections that stay idle, like most servers. If nginx reuses a connection just as it closes, that request attempt can fail. nginx retries GET requests by default, and proxy_next_upstream ... non_idempotent allows it to retry POST queries as well. DoH and ODoH queries can safely be repeated. Retries are only possible before nginx starts sending the response to the client.

Setting keepalive_timeout 5s; in the upstream block can reduce these retries at normal load with the default --timeout 10, but is optional. Under connection pressure, doh-proxy may close idle connections sooner, so a shorter nginx timeout does not replace retries.

With HAProxy

backend doh_backend
    mode http
    server doh1 127.0.0.1:3000 check

JSON API

The server supports Google's DNS-over-HTTPS JSON API format, making it compatible with applications that use this format.

Usage

Send GET requests to /dns-query with Accept: application/dns-json header:

# Query A records
curl -H "Accept: application/dns-json" \
  "http://localhost:3000/dns-query?name=example.com&type=1"

# Query with multiple parameters
curl -H "Accept: application/dns-json" \
  "http://localhost:3000/dns-query?name=example.com&type=28&cd=1&do=1"

Supported Parameters

name - Domain name to query (required)
type - DNS record type (default: 1 for A records)
cd - Disable DNSSEC validation (0 or 1)
do - Request DNSSEC data (0 or 1)
edns_client_subnet - Client subnet for EDNS

Response Format

{
  "Status": 0,
  "TC": false,
  "RD": true,
  "RA": true,
  "AD": false,
  "CD": false,
  "Question": [{
    "name": "example.com",
    "type": 1
  }],
  "Answer": [{
    "name": "example.com",
    "type": 1,
    "TTL": 300,
    "data": "93.184.216.34"
  }]
}

EDNS Client Subnet (ECS)

Overview

EDNS Client Subnet (ECS) is a DNS extension that allows the DoH proxy to forward client IP information to upstream DNS resolvers. This enables geo-optimized DNS responses by allowing authoritative nameservers to return results based on the client's actual location rather than the proxy's location.

Configuration Options

Enable ECS support with the following command-line options:

--enable-ecs - Enable EDNS Client Subnet functionality
--ecs-prefix-v4 <length> - IPv4 prefix length (default: 24)
--ecs-prefix-v6 <length> - IPv6 prefix length (default: 56)

How It Works

Client IP Extraction: The proxy extracts the client's IP address from:
- X-Forwarded-For header (takes the first IP if multiple are present)
- X-Real-IP header
- Direct connection IP (if no headers are present)
IP Truncation: For privacy, client IPs are truncated to the configured prefix length:
- IPv4: Default /24 (e.g., 192.168.1.100 → 192.168.1.0/24)
- IPv6: Default /56 (e.g., 2001:db8::1 → 2001:db8::/56)
DNS Query Enhancement: The truncated client subnet is added to outgoing DNS queries using the EDNS0 Client Subnet option (RFC 7871).
Geo-Optimized Responses: Upstream resolvers and authoritative nameservers can use this information to return geographically appropriate results.

Examples

Basic ECS setup:

doh-proxy -H 'doh.example.com' -u 8.8.8.8:53 --enable-ecs

Custom prefix lengths for more privacy:

doh-proxy -H 'doh.example.com' -u 8.8.8.8:53 \
          --enable-ecs \
          --ecs-prefix-v4 16 \
          --ecs-prefix-v6 48

Behind nginx with ECS enabled:

server {
    listen 443 ssl http2;
    server_name doh.example.com;

    location /dns-query {
        proxy_pass http://127.0.0.1:3000;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
    }
}

Testing ECS functionality:

# With X-Forwarded-For header
curl -H "X-Forwarded-For: 1.2.3.4" \
     -H "Accept: application/dns-json" \
     "https://doh.example.com/dns-query?name=example.com&type=1"

# The DNS query sent to upstream will include ECS: 1.2.0.0/24

Privacy Considerations

IP Truncation: Client IPs are never sent in full. The default settings provide a good balance between geolocation accuracy and privacy.
Opt-in Only: ECS is disabled by default and must be explicitly enabled.
Header Trust: Only trust X-Forwarded-For and X-Real-IP headers from known reverse proxies.
Logging: The proxy does not log client IPs when ECS is enabled, but upstream resolvers might.

For maximum privacy, avoid enabling ECS or use larger prefix values (smaller subnets) like /8 for IPv4 or /32 for IPv6.

Oblivious DoH (ODoH)

Oblivious DoH is similar to Anonymized DNSCrypt, but for DoH. It requires relays, but also upstream DoH servers that support the protocol.

This proxy supports ODoH termination (not relaying) out of the box.

However, ephemeral keys are currently only stored in memory. In a load-balanced configuration, sticky sessions must be used.

Currently available ODoH relays only use POST queries. So, POST queries have been disabled for regular DoH queries, accepting them is required to be compatible with ODoH relays.

This can be achieved with the --allow-odoh-post command-line switch.

Operational recommendations

DoH can be easily detected and blocked using SNI inspection. As a mitigation, DoH endpoints should preferably share the same virtual host as existing, popular websites, rather than being on dedicated virtual hosts.
When using DoH, DNS stamps should include a resolver IP address in order to remove a dependency on non-encrypted, non-authenticated, easy-to-block resolvers.
Unlike DNSCrypt where users must explicitly trust a DNS server's public key, the security of DoH relies on traditional public Certificate Authorities. Additional root certificates (required by governments, security software, enterprise gateways) installed on a client immediately make DoH vulnerable to MITM. In order to prevent this, DNS stamps should include the hash of the parent certificate.
TLS certificates are tied to host names. But domains expire, get reassigned and switch hands all the time. If a domain originally used for a DoH service gets a new, possibly malicious owner, clients still configured to use the service will blindly keep trusting it if the CA is the same. As a mitigation, the CA should sign an intermediate certificate (the only one present in the stamp), itself used to sign the name used by the DoH server. While commercial CAs offer this, Let's Encrypt currently doesn't.
Make sure that the front-end supports at least HTTP/2 and TLS 1.3.
Internal DoH servers still require TLS certificates. So, if you are planning to deploy an internal server, you need to set up an internal CA, or add self-signed certificates to every single client.

Timeouts and connection limits

--timeout (10 seconds by default, between 1 and 3600) is the time budget of a single request. It starts once the request headers have been received, and covers reading the request body, waiting for the upstream resolver, and retrying over TCP when a UDP response was truncated. A body that doesn't arrive in time gets a 408 response, and a resolver that doesn't answer in time gets a 502.

The same value controls how long connections are kept, so there is nothing else to tune:

A connection with no request in progress is closed once it has been idle for the timeout. This includes connections that never send anything, and incomplete HTTP/1 request headers.
When the number of connections gets close to --max-clients, idle connections are closed sooner: after half the timeout from 70% of the limit, and after a tenth of it (never less than a second) from 90%. The delays go back up once the number of connections falls below 80%, and then below 60% of the limit. Requests in progress are never cut short because of this.
TLS handshakes have to complete within the same delays, counted from the moment the connection was accepted, but they always get at least two seconds. Once the handshake is done, the idle delay starts over, so the time spent on the handshake is not taken away from the first request.
Every connection is retired once it is 6 times the timeout old, or 60 seconds if that is longer. HTTP/1 connections finish the response in progress and close with Connection: close. HTTP/2 connections send a GOAWAY frame and complete the streams they had already accepted. Clients simply open a new connection.
A retiring connection has twice the timeout to finish its work. After that, it is closed even if a response is still waiting to be delivered, for example to a client that stopped reading.

With the default timeout, this gives:

+------------------------------------------------+------------+
| Event                                          | Default    |
+------------------------------------------------+------------+
| Request budget                                 | 10 s       |
| Idle connection closed after                   | 10 s       |
|   from 70% / 90% of the connection limit       | 5 s / 1 s  |
| TLS handshake limit                            | 10 s       |
|   from 70% / 90% of the connection limit       | 5 s / 2 s  |
| Connection retired at the age of               | 60 s       |
| Retiring connection closed at the latest after | 20 s       |
+------------------------------------------------+------------+

--max-clients counts connections from the moment they are accepted, TLS handshakes included. Connections above the limit are closed right away, before any TLS work.

A proxy that reuses connections to doh-proxy should retry requests that fail because an idle connection was closed at the same moment.

Truncated UDP responses are retried over TCP. The number of simultaneous TCP connections to the upstream resolver is limited separately, to one for every 8 allowed clients (and at least one). A query that needs one waits for a free connection, within its own time budget.

These limits bound how long the server holds on to its resources, but they are a safety net, not a guarantee. Under heavy overload, or with an upstream resolver that stops answering, some queries will still fail.

DNS Stamps and Certificate Hashes

Use the online DNS stamp calculator to compute the stamp for your server.

Add it to the [static] section of dnscrypt-proxy and check that everything works as expected.

Then, start dnscrypt-proxy with the -show-certs command-line flag to print the hashes for your certificate chain.

Here is an example output:

[NOTICE] Advertised cert: [CN=dohtrial.att.net,O=AT&T Services\, Inc.,L=Dallas,ST=Texas,C=US] [f679e8451940f06141854dc94e1eb79fa5e04463c15b88f3b392da793c16c353]
[NOTICE] Advertised cert: [CN=DigiCert Global CA G2,O=DigiCert Inc,C=US] [f61e576877da9650294cccb5f96c75fcb71bda1bbc4646367c4ebeda89d7318f]

The first printed certificate is the certificate of the server itself. The next line is the one that signed that certificate. As you keep going down, you are getting closer to the certificate authority.

Unless you are using intermediate certificates, your safest option is probably to include the last printed hash certificate in your DNS stamp.

Go back to the online DNS stamp calculator, and copy&paste the hash (in this example: f61e576877da9650294cccb5f96c75fcb71bda1bbc4646367c4ebeda89d7318f).

If you are using Let's Encrypt, the last line is likely to be:

Advertised cert: [CN=Let's Encrypt Authority R3,O=Let's Encrypt,C=US] [444ebd67bb83f8807b3921e938ac9178b882bd50aadb11231f044cf5f08df7ce]

There you have it. Your certificate hash is 444ebd67bb83f8807b3921e938ac9178b882bd50aadb11231f044cf5f08df7ce.

This Go code snippet can also compute the hash of certificates given a .der file.

Why Certificate Hashes in DoH Stamps Matter

Background

For DNSCrypt resolvers, the stamp includes the resolver’s public key, ensuring authenticity by design.

For DoH (DNS-over-HTTPS) resolvers, however, authenticity normally depends only on the WebPKI (the set of trusted Certificate Authorities in the system). By default, any trusted CA can issue a valid TLS certificate for your DoH server’s domain.

This is where certificate hashes in stamps come in.

Why They’re Important

Defends against MITM via extra CAs Many systems have extra root CAs installed by enterprises, antivirus tools, or governments. Without pinning, these CAs can issue certificates for your DoH domain and intercept queries. Including certificate hashes ensures the client only accepts the intended TLS chain.
Prevents silent domain takeover If your DoH hostname changes ownership, the new owner can obtain a valid certificate and impersonate your resolver. Pinning the parent certificate’s hash (ideally a dedicated intermediate CA) ensures the client rejects impostors.
Restores DNSCrypt-level assurance DNSCrypt stamps always bind to a known key. With certificate hashes, DoH stamps gain a similar property: encryption and authenticity.

How Certificate Hashes Work in Stamps

A DoH stamp can include one or more SHA-256 hashes of certificates in the resolver’s TLS chain.
A client must see at least one matching certificate hash when connecting.
Operators can list multiple hashes to support certificate rotations smoothly.

Common certificate hashes

Let's Encrypt E1:
- cc1060d39c8329b62b6fbc7d0d6df9309869b981e7e6392d5cd8fa408f4d80e6
Let's Encrypt R3:
- 444ebd67bb83f8807b3921e938ac9178b882bd50aadb11231f044cf5f08df7ce
Let's Encrypt R10:
- e644ba6963e335fe765cb9976b12b10eb54294b42477764ccb3a3acca3acb2fc
ZeroSSL:
- 9a3a34f727deb9bca51003d9ce9c39f8f27dd9c5242901c2bab1a44e635a0219

Troubleshooting

Common Issues

Port already in use:

# Check what's using port 3000
lsof -i :3000
# Or use a different port
doh-proxy -l 127.0.0.1:3001 ...

Certificate errors:

Ensure certificate file contains the full chain
Convert ECDSA keys to PKCS#8 format
Check file permissions (readable by the doh-proxy user)

DNS resolution failures:

Verify the upstream DNS server is reachable
Check firewall rules for port 53 (UDP/TCP)
Test with: dig @127.0.0.1 -p 53 example.com

Performance Tuning

For high-traffic deployments:

doh-proxy -H 'doh.example.com' \
          -u 127.0.0.1:53 \
          -c 10000 \     # Max clients
          -C 100 \       # Max concurrent streams per client
          -t 30          # Timeout in seconds

Clients

Compatible DoH clients include:

dnscrypt-proxy - Supports both DNSCrypt and DoH
cloudflared - Cloudflare's DoH proxy
Firefox, Chrome, Edge (native DoH support)
doh-client - Rust DoH client
Android 9+ and iOS 14+ (native DoH support)

Public Deployments

doh-proxy powers several public DNS services including:

doh.crypto.sx - Public DNS resolver
Many other services listed in public encrypted DNS servers

Contributing

Contributions are welcome! Please feel free to submit pull requests or open issues on GitHub.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Name		Name	Last commit message	Last commit date
.github		.github
contrib/openrc		contrib/openrc
src		src
.gitignore		.gitignore
Cargo.toml		Cargo.toml
LICENSE		LICENSE
README.md		README.md
localhost.pem		localhost.pem
logo.png		logo.png

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Table of Contents