MAX98357A for Radza Zero 3 – Giving Skelly a Voice with I2S and GPIO

When I started my AI Powered Skelly project built with a Radxa Zero 3 I quickly discover that the board does not ship with on‑board audio. This was a problem, because I really wanted to utilize the existing speaker within the 3ft Dancing Skeleton animatronic as a more natural source of audio output. To solve this I choose to use a MAX98357A mono amplifier to drive audio output to small speaker. This article walks you through everything you need to get the MAX98357A working on a Radxa Zero 3, from wiring the pins to enabling I2S software and finally creating the custom device‑tree overlay that makes the kernel aware of the hardware.

 MAX98357A

Why Choose MAX98357A for Radxa Zero 3?

FeatureMAX98357ATypical Alternatives
Power consumption< 100 mW (idle)> 200 mW for many DACs
Output typeMono, 3 W Class‑DStereo, often larger footprint
InterfaceI2S digital audio input onlyPCM/I2C/USB combos
Gain controlFixed (9 dB default) or external pinVariable via software registers
Shutdown pinOptional hardware lineOften not present

Because the amplifier needs only three I2S signals (BCLK, LRCK and DATA) plus power and ground, it can be wired directly to the standard I2S pins on the Radxa Zero 3 without any extra level shifters. The board’s 5 V rail supplies the amplifier’s VIN, while the GND pins provide a clean reference.

GPIO Pinout – Mapping MAX98357A to Radxa Zero 3

The Radxa Zero 3 exposes its I2S‑3 controller (named I2S3_M0) on the following header pins:

MAX98357A pinFunctionRadxa Zero 3 Pin #SoC GPIO name
VIN5 V power2 or 4 5v VDD_5V
GNDGround6, 9, 14, 20, 25, 30, 34, 39GND
BCLKI2S Bit Clock12I2S3_SCLK_M0
DIN (DATA)I2S Serial Data In40I2S3_SDO_M0
LRC (LRCK)I2S Left‑Right Clock35I2S3_LRCK_M0
GAIN (optional)Fixed gain selector – connect to GND for 6 dB, leave floating for default 9 dBany free GND (e.g., 20)GND
SD (Shutdown)Active‑low shutdown – pull low to muteany free GPIO (e.g., 16 with optional PWM)GPIO3_B1/PWM8_M0

Tip: The GAIN and SD pins are optional. For a simple “always on” configuration you can leave them floating. If you need hardware mute, connect the SD pin to a spare GPIO and drive it low when you want silence.

Wiring Diagram

Wiring MAX98357A for Radxa Zero 3
Radxa Zero 3                 MAX98357A
---------------------------  -----------------
Pin 4   (+5 V)               VIN  (Red)
Pins 6/9/... (GND)           GND  (Black)
Pin 12  I2S3_SCLK_M0 (BCLK)  BCLK (Blue)
Pin 40  I2S3_SDO_M0  (DATA)  DIN  (Orange)
Pin 35  I2S3_LRCK_M0 (LRCK)  LRC  (Green)
Pin 16  (optional)           GAIN (Purple Doted)
Pin 18 (optional)            SD   (Black Doted)

All connections are 5 V tolerant on the Radxa Zero 3, and the MAX98357A operates comfortably from 2.7 V up to 5.5 V, so no level shifting is required.

Preparing the Software – Enabling I2S in the Kernel

The Radxa Zero 3 runs a Debian‑based OS (or Ubuntu) with a Linux kernel that already contains an I2S controller driver (i2s-axg). However, the default device tree does not expose the I2S3_M0 peripheral for audio output. To make it usable you must:

  1. Install the overlay tooling – Radxa provides rsetup and a set of scripts to compile and install custom overlays.
   sudo apt-get update
   sudo apt-get install rsetup device-tree-compiler
  1. Disable any conflicting overlay – The board ships with an i2s3-m0 overlay that is intended for microphone input. You need to remove or disable it before loading your speaker overlay.
   sudo rsetup overlay disable i2s3-m0
  1. Compile the custom overlay (shown in the next section) and install it:
   sudo rsetup overlay add radxa-zero3-max98357a.dts
  1. Reboot to let the kernel load the new device‑tree node.
   sudo reboot
  1. Verify that the I2S PCM device appears under /dev.
   aplay -l
   # You should see something like:
   # card 0: Rockchip [rockchip i2s], device 0: I2S3 PCM [...]
  1. Test audio playback – Use aplay with a raw PCM file or any WAV that matches the default format (16‑bit, 44.1 kHz, mono).
   aplay -D plughw:0,0 /usr/share/sounds/alsa/Front_Center.wav

If you hear sound from your speaker, the hardware and software stack are correctly configured.

The Custom Device‑Tree Overlay – radxa-zero3-max98357a.dts

Below is the complete overlay source that tells the kernel to expose I2S3_M0 as a PCM output and optionally controls the GAIN and SD pins as GPIOs.

/dts-v1/;
/plugin/;

#include <dt-bindings/gpio/gpio.h>
#include <dt-bindings/pinctrl/rockchip.h>

/ {
    metadata {
        title = "Enable MAX97357A on I2S3-M0";
        compatible = "radxa,zero3";
        category = "audio";
        description = "Enable MAX97357A on I2S3-M0";
        exclusive = "GPIO3_A0", "GPIO3_A3", "GPIO3_A4", "GPIO3_A5", "GPIO3_A6", "i2s3_2ch";
    };
};

&{/} {
    max98357a_codec: max98357a {
        #sound-dai-cells = <0>;
        compatible = "maxim,max98357a";
        sdmode-gpios = <&gpio3 RK_PA0 GPIO_ACTIVE_HIGH>;
        sdmode-delay = <5>;
        status = "okay";
    };

    sound_ext_card: sound-ext-card {
        #address-cells = <1>;
        #size-cells = <0>;
        status = "okay";
        compatible = "simple-audio-card";
        simple-audio-card,format = "i2s";
        simple-audio-card,mclk-fs = <256>;
        simple-audio-card,name = "snd_max98357a_dac";
        simple-audio-card,dai-link@0 {
            reg = <0>;
            format = "i2s";
            cpu {
                sound-dai = <&i2s3_2ch>;
            };
            codec {
                sound-dai = <&max98357a_codec>;
            };
        };
    };
};

&i2s3_2ch {
    pinctrl-0 = <&i2s3m0_lrck &i2s3m0_sclk &i2s3m0_sdi &i2s3m0_sdo>;
    status = "okay";
};

How the Overlay Is Applied

  1. Compilersetup overlay add radxa-zero3-max98357a.dts runs dtc -@ -I dts -O dtb internally and places the resulting .dtbo file under /boot/overlays.
  2. Activate – The same command updates /boot/extlinux/extlinux.conf to load the overlay at boot.
  3. Boot – During kernel initialization, the overlay patches the base device tree, making the simple-audio-card node visible to ALSA.

Controlling Shutdown from Userspace

Because we exposed the optional pins as standard Linux GPIOs, you can manipulate them with the sysfs interface or via the newer libgpiod library.

Using sysfs (quick test)

# Export GPIO11 (shutdown) and keep high (amp on)
echo 16 > /sys/class/gpio/export
echo out > /sys/class/gpio/gpio16/direction
echo 1 > /sys/class/gpio/gpio16/value  # 1 = not shutdow

Using libgpiod (modern approach)

#include <gpiod.h>

int main(void) {
    struct gpiod_chip *chip;
    struct gpiod_line *sd;

    chip = gpiod_chip_open_by_name("gpiochip3");
    sd   = gpiod_chip_get_line(chip, 1);    // GPIO16

    gpiod_line_request_output(sd,   "max98357a_sd",   1);

    /* ... later you can toggle them */
    gpiod_line_set_value(sd,   0); // shutdown amp
}

Container‑Friendly Deployment

If you are running audio workloads inside Docker on the Radxa Zero 3, keep the following in mind:

RequirementHow to satisfy
Access to /dev/sndAdd --device /dev/snd:/dev/snd (Docker) or mount the device in the pod spec.
GPIO controlMount /sys/class/gpio read‑write or use the gpiod socket (/run/gpiod).
Real‑time scheduling (optional)Use --cap-add SYS_NICE and set ulimit -r 99.
Overlay persistenceThe overlay is applied at boot, so containers do not need to modify it. Just ensure the host kernel has loaded the PCM device before container start.

A minimal Dockerfile for an audio player might look like:

FROM debian:bookworm-slim

RUN apt-get update && \
    apt-get install -y alsa-utils libgpiod2 && \
    rm -rf /var/lib/apt/lists/*

COPY myplayer.sh /usr/local/bin/
ENTRYPOINT ["/usr/local/bin/myplayer.sh"]

Run it with:

docker run --rm -it \
  --device /dev/snd:/dev/snd \
  -v /sys/class/gpio:/sys/class/gpio:rw \
  my-audio-player

Now your container can play back WAV files through the MAX98357A and mute/unmute using GPIO.

Troubleshooting Checklist

SymptomLikely CauseFix
No sound, aplay returns “Invalid argument”I2S node not enabledVerify overlay loaded (dmesg | grep i2s3) and run aplay -l.
Crackling or clicksClock mismatch or bad power supplyEnsure VIN is stable 5 V, add a decoupling capacitor (10 µF) near the amplifier.
Speaker only plays one channelLRC pin mis‑wired or wrong pinmuxDouble‑check that Pin 35 is connected to LRC and i2s3_m0_pins includes it.
Amplifier stays silent even after GPIO highSD pin tied lowRemove the shutdown connection or set GPIO16 value to 1.
Gain not changingGAIN pin left floating while expecting -6 dBConnect GAIN to a GPIO and drive it low, or permanently tie to GND for fixed gain.

Use journalctl -k to see kernel messages about the audio card; any “probe failed” lines usually indicate a mis‑matched overlay.

Security Considerations

Even though this guide focuses on hardware integration, running audio services on an edge device still raises security questions:

  1. GPIO Exposure – Unrestricted access to /sys/class/gpio allows any local user (or compromised container) to toggle the amplifier’s pins, potentially causing denial‑of‑service or odd behavior. Restrict GPIO permissions using Linux capabilities or mount namespaces.
  2. Audio Injection – If your device accepts network streams, validate source authenticity (TLS, signed manifests) before feeding data to ALSA.
  3. Kernel Attack Surface – Custom overlays modify the kernel’s view of hardware. Ensure only trusted administrators can write overlay files; use immutable boot partitions where possible.

Conclusion

The MAX98357A for Radxa Zero 3 offers a lightweight, low‑cost way to add mono audio output to any project that runs on this SBC. By wiring three I2S pins and enabling the I2S controller through a custom device‑tree overlay I was able to give the AI Skelly the power to speak.

Up next we look at how to programmatic trigger an event, in order to capture an image from a simple attached webcam. Then how to process the image via a standard openAI style API call to a local AI service like ollama serve or LMStuido server. In order to ask targeted questions about the capture image, such as the Halloween costumes in frame. The we can use the prompt response to generate a unique voice line response with services like piper. All while keeping everything completely offline for privacy and security.

Bringing Skelly to Life: A Radxa Zero 3W Powered AI Halloween Skeleton

Halloween is my favorite time of year, and I love creating interactive experiences for trick-or-treaters. Last year, Alex Volkov’s project on Weights & Biases – building an AI-powered skeleton using a Raspberry Pi – sparked an idea. This year, I wanted to take that concept further, aiming for a smaller footprint and increased processing power by leveraging the Radxa Zero 3W single board computer. This blog post details my journey of transforming a standard Home Depot 3ft Halloween Classics Animated LED Dancing Skeleton into a locally AI-driven greeter. We’ll cover everything from dismantling the original animatronic, wiring up the Radxa Zero 3W, and setting the stage for integrating local vision models to recognize costumes.

The Inspiration & Goals

Volkov’s project was brilliant: using online AI services like Google AI Studio, ElevenLabs (for voice), Cartesia, and ChatGPT to create a responsive skeleton that could greet trick-or-treaters. However, relying on cloud services introduces latency, requires a stable internet connection, and could raise privacy concerns – not ideal for the often chaotic Halloween night. My goal was to replicate the interactive experience but move all processing local, using a more compact and powerful board than the Raspberry Pi 4. The Radxa Zero 3W seemed like the perfect fit. It packs significant punch in a tiny form factor, offering Wi-Fi connectivity, Bluetooth, and ample GPIO pins for controlling the animatronic components.

Disassembly & Component Identification: Getting to Know Skelly

The first step was understanding how the original skeleton worked. This involved carefully dismantling the 3ft dancing skeleton. Start by removing the back of the skull and chest plate; this provides access to the control board, battery pack, motors, and speaker. Be gentle – these animatronics aren’t built for extensive tinkering!

Inside the skull, you’ll find a DC motor controlling the mouth movement (yellow positive, white ground wires) and LEDs illuminating the eyes (red positive, black ground wires).

Photo of the skeleton's internal components after removing the skull backplate. Highlight the mouth motor and eye LEDs

Under the chest plate you will see a hardware speaker with two blue wires and another DC motor powering the body/arm movements (positive red, ground black). All these wires converge on a small control board.

Photo of the skeleton's internal components after removing the chest backplate. Highlight the control board, battery pack,  body motor, and speaker

It’s crucial to document everything as you go. I took numerous photos and created a wiring diagram to ensure I could reassemble everything correctly (or at least understand where things went if something went wrong!).

Close-up documenting the Home Depot 3ft Halloween Classics Animated LED Dancing Skeleton control board wiring

The original manufacturer uses transistors and capacitors to compensate for the fluctuating battery voltage – typically between 1.4V and 1.66V with three AA batteries in parallel, reaching around 5V. This is a good reminder that relying solely on the battery pack’s power output isn’t ideal; we’ll address this later.

Wiring Up the Radxa Zero 3W: The Heart of the Operation

The plan was to intercept the signals going to each component – mouth motor, eye LEDs, body motor, and speaker – and control them via the Radxa Zero 3W’s GPIO pins. This required carefully unsoldering these wires from the original control board.

Once unsolder, I connected each wire to a set of jumper pins, allowing me to easily breadboard and test connections before committing to permanent soldering. This also provides flexibility for future modifications.

Note: I included a 220 Ohm inline to help prevent the eye LEDs from burning out. Its not required, but its recommended to avoid burning out the LEDs during tinkering.

Close-up photo showing the wires from the skeleton's components connected to jumper cables.

Here’s the GPIO pinout I utilized on the Radxa Zero 3W (gpiochip3):

  • PIN_7 (gpiochip3 20) – Mouth motor open/close
  • PIN_11 (gpiochip3 1) – Eye LEDs illumination
  • PIN_15 (gpiochip3 8) – Body motor for dancing movement

The Radxa Zero 3W’s official documentation outlines the 40-pin GPIO interface. Since we’re using pins 12, 40 and 35 for our mono amp (more on that later), this leaves a good selection of readily available pins on gpiochip3 to control relays.

  • Pin 7: GPIO3_C4 (also PWM14_M0)
  • Pin 11: GPIO3_A1
  • Pin 15: GPIO3_B0
  • Pin 16: GPIO3_B1 (also PWM8_M0)

Note: PWM (Pulse-Width Modulation) can be used on Pin 7 and Pin 16 by enabling the correct device tree overlay using rsetup and/or u-boot. This allows for finer control over LEDs (dimming) and DC motor speeds, but wasn’t necessary for this initial implementation.

Powering Skelly: The Radxa Zero 3W to the Rescue

Initially, I considered powering the components directly from the battery pack. However, as mentioned earlier, the inconsistent voltage proved problematic. The solution? Leverage the Radxa Zero 3W’s 5V GPIO power rails! This provides a stable and reliable power source for all components.

To manage the current requirements of the motors and LEDs, I incorporated relays inline with each component’s wiring. Relays act as electrically controlled switches, allowing the Radxa Zero 3W to control the flow of power from its 5V output to the skeleton’s components.

showing how the Radxa Zero 3W controls the skeleton's components via relays

Relay Implementation: The Switching Mechanism

Each relay requires a control pin on the GPIO, which when activated, allows power to flow through it. The wiring is as follows:

  1. Connect the Radxa Zero 3W’s 5V output to both sides of each relay.
  2. Ground each relay and component to the Radxa Zero 3W’s ground pins.
  3. Connect the control pin on the GPIO to the relay’s control input.
  4. The output power of each relay connects to the corresponding component’s jumper (mouth motor, body motor, eye LEDs).

When the GPIO pin is set HIGH (active state), the relay closes, allowing power to flow from the Radxa Zero 3W to the component. When the pin is LOW, the relay opens, cutting off the power supply. This effectively gives us programmatic control over each animatronic function.

Breadboarding & Testing: Bringing it All Together

I started by breadboarding everything – connecting the Radxa Zero 3W, relays, jumpers, and a temporary power source to verify functionality. This is where patience is key! Double-check all connections before applying power. A multimeter is your best friend during this phase. Once I confirmed that each component responded correctly to the GPIO signals, I removed the breadboard and connected everything directly to the jumper wires for a more permanent connection.

Mounting & Final Assembly: Skelly Gets an Upgrade

With the wiring complete, it was time to mount the Radxa Zero 3W inside the skeleton’s chest cavity. I repurposed the original control board’s mounting point and used some 2.5mm standoffs to secure the Radxa Zero 3W in place. This ensured a snug fit without interfering with any existing components.

 Photo showing the Radxa Zero 3W and relays mounted inside the skeleton’s chest cavity

I then stuffed all the “guts” back into the body, wrote a simple web control page, using Flask, to test the functionality remotely, and ran through final testing. Success! Skelly was now responding to commands from my computer, ready for the next phase: AI integration.

The Next Phase: Local Vision Models & Costume Recognition

With Skelly reasonably buttoned up and his basic movements working, I’m moving on to the most exciting part of the project – leveraging a connected camera and locally hosted/trained vision models to determine trick-or-treaters’ costumes. This will involve using services like LMStuido on an AMD AI workstation to leverage models like Gemma 3 for vision to text. I plan to document this process in a future blog post, so stay tuned! But up next a deep dive on how to leverage device tree overlays and I2S via GPIO, to power an existing hardware speaker with a MAX98357A mono amp.

Resources & Further Exploration

This project has been a fantastic learning experience, combining hardware tinkering with software development and AI integration. The Radxa Zero 3W proved to be an excellent platform for this application, offering the power and flexibility needed to bring Skelly to life. I hope this blog post inspires you to create your own interactive Halloween experiences!

Enforcing Device Trust with Certificate‑Based Client Authentication – A Practical Guide for Modern Identity Providers

Introduction

In today’s zero‑trust world, granting access solely based on a user’s password or even an OTP is no longer sufficient. Threat actors are increasingly targeting compromised credentials, and the security community has responded with identity provider empowered device trust, a model that ensures only verified devices can obtain tokens from an identity provider (IdP).

When you combine device trust with TLS client‑certificate authentication, you create a powerful barrier: before any session token is issued for a sensitive application or service, the IdP validates a device‑bound certificate presented by the client. This approach works equally well for on‑premises workloads, cloud native services, and containerized applications.

In this post we’ll walk through how to configure popular IdPs like Keycloak, Authentik, Okta, OneLogin, and Auth0 to require Certificate ClientAuth using device trust certificates. We’ll also discuss best‑practice considerations for certificate issuance, rotation, and revocation.

Why Certificate‑Based Device Trust?

BenefitExplanation
Strong mutual TLS (mTLS)The client proves possession of a private key bound to a trusted device, eliminating credential stuffing attacks.
Device bindingCertificates are issued per device (or per container/compute) and can be tied to hardware TPMs or secure enclaves, making them hard to steal.
Zero‑trust enforcementEven if a user’s password is compromised, an attacker cannot obtain a token without the correct device certificate.
Fine‑grained policy controlIdPs can apply separate authentication policies based on certificate attributes (e.g., OU=Mobile, OU=Workstation).
Auditable provenanceEvery successful login includes the certificate fingerprint, simplifying forensic investigations.

Device Trust Certificate Lifecycle

  1. Enrollment – During first‑boot or via automated provisioning, the device presents its hardware-bound public certificate to an enrollment endpoint (often part of the IdP or Third-party service like ACME Attestation service).
  2. Provisioning – A trusted CA (internal PKI or managed service) issues a short‑lived X.509 certificate to each device. The private key is bound by the vTPM, HSM, or container runtime.
  3. Authentication – When a user attempts to access a protected app, the client initiates TLS with client‑certificate request. The presented certificate is validated against Device Trust certificate chain.
  4. Renewal / Rotation – Before expiry (typically 30–90 days), the device automatically requests a new cert via the provisioner. Old certificates are revoked or marked as expired in the CRL database.

Identify Provider Empowered Device Trust Authentication Flow

identity provider empowered with device trust Session flow diagram

The diagram illustrates that no token is ever issued unless the device presents a valid certificate that matches the IdP’s trusted CA list. This is the essence of an identity provider empowered device trust.

Preparing Your PKI

All IdPs discussed support the use of external CA integration. The steps are similar across platforms:

  1. Create a dedicated CA hierarchy – Root CA (offline) → Intermediate “Device‑Trust” CA (online).
  2. Define certificate profile – Include extensions such as subjectAltName for device ID, extendedKeyUsage = clientAuth, and optionally certificatePolicies to tag the trust domain.
  3. Deploy an automated provisioning service – For example, cert‑manager in Kubernetes, ACME with device attestation, or a simple SCEP service.

Note: Alternatively you can use smallstep step-ca as outlined in my practical cloud-native guide.

Sample OpenSSL config for the Device‑Trust CA:

[ req ]
distinguished_name = req_distinguished_name
prompt = no

[ req_distinguished_name ]
C  = US
ST = WA
L  = Seattle
O  = AcmeCorp
OU = DeviceTrustCA
CN = device-trust.acme.local

[ v3_intermediate_ca ]
basicConstraints = critical, CA:true, pathlen:0
keyUsage = critical, digitalSignature, cRLSign, keyCertSign
subjectKeyIdentifier = hash
authorityKeyIdentifier = keyid:always,issuer

Use this to generate the intermediate CA and then sign device CSRs with -extensions client_auth:

openssl req -new -nodes -newkey rsa:2048 \
  -keyout device.key -out device.csr \
  -subj "/C=US/ST=WA/L=Seattle/O=AcmeCorp/OU=Workstation/CN=device01"

openssl ca -config openssl.cnf -extensions client_auth \
  -days 30 -notext -md sha256 \
  -in device.csr -out device.crt

Configuring Identity Providers

Below we present the minimal configuration needed to enable Certificate ClientAuth for each IdP. The examples assume you already have a running instance of the IdP and that your PKI’s intermediate certificate is available as device-trust-ca.pem.

Keycloak

Keycloak provides TLS client‑certificate authentication via the X509 Authentication flow.

# keycloak-realm.yaml – add a new authentication flow called "DeviceTrust"
realm: myrealm
authenticationFlows:
  - alias: DeviceTrust
    providerId: basic-flow
    topLevel: true
    builtIn: false
    authenticationExecutions:
      - authenticator: x509-browser-authenticator
        requirement: REQUIRED
        priority: 10
        config:
          # Truststore containing device‑trust CA
          trustStoreFile: /opt/keycloak/conf/device-trust-ca.p12
          trustStorePassword: changeit
          # Map certificate subject DN to user attribute "deviceId"
          principalAttribute: cn
      - authenticator: auth-cookie
        requirement: REQUIRED
        priority: 20
  1. Upload the CA – Convert device-trust-ca.pem into a PKCS‑12 keystore (keytool -importcert).
  2. Create the flow via the Admin Console or import the YAML above.
  3. Set the flow as default for the desired client (application) under Authentication → Flows.

Now, any request to /auth/realms/myrealm/protocol/openid-connect/auth will trigger a TLS handshake that expects a device‑trust certificate.

For More information about Keycloak clientAuth review the x509 documentation.

Authentik

Authentik’s Certificate provider can be combined with the Device stage.

# authentik.yaml – DeviceTrustProvider definition
providers:
  - name: device-trust-mtls
    kind: cert
    config:
      ca_file: /etc/authentik/certs/device-trust-ca.pem
      allowed_usages:
        - clientAuth
      map_subject_to_user_attribute: "device_id"
stages:
  - name: DeviceTrustStage
    kind: authentication
    flow: default-authentication-flow
    providers:
      - device-trust-mtls
  1. Place device-trust-ca.pem under /etc/authentik/certs.
  2. Restart Authentik; the stage will now reject any TLS handshake lacking a valid client cert.

For more information about configuring Authentik authentication review the stag flow documentation.

Okta

Okta’s Certificate Authentication is configured via API Access Management.

{
  "type": "CERTIFICATE",
  "name": "DeviceTrustPolicy",
  "settings": {
    "trustedCertificates": [
      { "x5c": ["MIID..."] }   // Base64‑encoded device‑trust CA cert
    ],
    "subjectMatchPattern": "CN=*.device.acme.com"
  }
}

Steps:

  1. In the Okta Admin Console, navigate to Security → Authenticators → Certificate and add a new authenticator with the above JSON.
  2. Attach this authenticator to an Authentication Policy that protects your sensitive app (e.g., “Sensitive‑App‑Policy”).

Okta will now enforce mTLS for any request hitting the OIDC /authorize endpoint of that app.

For more information on configuring certificate-based authentication for Okta review the official docs.

OneLogin

OneLogin supports X.509 certificate authentication through its MFA configuration.

<!-- one-login-mfa-config.xml -->
<CertificateAuthenticator>
    <TrustedCA>$CA_Chain_URL</TrustedCA>
    <SubjectRegex>CN=([a-z0-9\-]+)</SubjectRegex>
    <MapToUserAttribute>device_id</MapToUserAttribute>
</CertificateAuthenticator>

Upload the XML via Settings → Security → Multifactor Authentication → Certificate. Then enable this factor for the Security Policy that guards your high‑value applications.

For more information about how Onelogin handles third-party certificate authentication to validate a trusted device, review the official documentation.

Auth0

Auth0 uses Custom Database Connections with a pre‑login hook to verify client certificates.

// auth0-pre-login.js – Deploy as an Action (Pre‑Login)
exports.onExecutePostLogin = async (event, api) => {
  const certHeader = event.request.headers['x-client-cert'];
  if (!certHeader) {
    return api.access.deny('client_certificate_missing');
  }

  // Decode PEM and verify against trusted CA
  const forge = require('node-forge');
  const pki = forge.pki;
  const caPem = `-----BEGIN CERTIFICATE-----
MIID...
-----END CERTIFICATE-----`;
  const caCert = pki.certificateFromPem(caPem);
  const clientCert = pki.certificateFromPem(certHeader);

  // Basic chain validation
  const verified = pki.verifyCertificateChain(pki.createCaStore([caCert]), [clientCert]);
  if (!verified) {
    return api.access.deny('invalid_device_certificate');
  }

  // Optional: map CN to user metadata
  const deviceId = clientCert.subject.getField('CN').value;
  event.user.app_metadata = { ...event.user.app_metadata, device_id: deviceId };
};

Deploy this Action and enable TLS termination with client‑certificate forwarding on your reverse proxy (e.g., Nginx proxy_set_header X-Client-Cert $ssl_client_cert;). Auth0 will reject any login that lacks a valid device‑trust certificate.

For more information about how to use Auth0’s mTLS authentication flow to validate device trust certificates, review their official mTLS docs.

Note: For services that may not directly support your Identity provider or may not be exposed externally, you can utilize a simple nginx proxy to validate a device trust certificate before allowing users to login.

Best Practices for Production Deployments

AreaRecommendation
Certificate LifetimeUse short lifetimes (30 days) and automate renewal via a provisioner like ACME or SCEP
Key ProtectionStore private keys in credential manager, KMS, Key Vault, or secrets store. Never write them to disk unencrypted.
RevocationPublish CRLs or use OCSP stapling; IdPs should query the revocation endpoint on each login.
Logging & AuditingInclude tls.client.subject_dn and certificate fingerprint in SIEM logs. Enable audit‑log retention for at least 90 days.
Fail‑Open vs Fail‑CloseDefault to fail‑close: if the client cert cannot be validated, deny access.
Device InventoryKeep a synchronized inventory service (e.g., CMDB) that tracks active device fingerprints; automate de‑provisioning when devices leave the fleet.

Testing & Validation

  1. OpenSSL verification – From a client machine with the device cert:
   openssl s_client -connect idp.acme.local:443 \
     -cert device.crt -key device.key -CAfile device-trust-ca.pem

You should see Verify return code: 0 (ok) and the TLS handshake succeed.

  1. Token request – Use curl with the client cert:
   curl -k https://idp.acme.local/auth/realms/myrealm/protocol/openid-connect/token \
     -E device.crt --key device.key \
     -d "grant_type=client_credentials&client_id=myapp"

If the certificate is invalid or missing, the response will be 401 Unauthorized.

  1. Audit log check – In Keycloak’s admin console go to Events → Config and enable Login events. Verify that each successful login entry contains client_certificate_fingerprint.

Common Pitfalls & How to Avoid Them

SymptomRoot CauseFix
Handshake fails with “unknown ca”Device‑trust CA not added to truststore or wrong file format (PEM vs PKCS12)Convert to PKCS12 and import correctly.
Token issued despite missing certReverse proxy terminates TLS before forwarding to IdP, losing client‑cert headerEnable proxy_ssl_verify and forward X-Client-Cert or use end‑to‑end mTLS (no TLS termination at proxy).
Frequent revocation failuresCRL/OCSP endpoint unreachable from IdP and user deviceHost a OCSP responder or cache CRLs; ensure network connectivity.
Certificate renewal breaks sessionsApplications cache the old cert and do not reload new filesUse cronjob or sidecar containers that check the certificate and automate renewal

Conclusion

By integrating device‑bound certificates with your identity provider, you transform authentication from a just a 2 factor authentication model and combine it into a robust multi factor with verifiable trust paradigm. The configurations shown for Keycloak, Authentik, Okta, OneLogin, and Auth0 prove that enabling Certificate ClientAuth is straightforward, often just a few lines of YAML or JSON plus the import of a trusted CA.

When you adopt this pattern:

  • Security posture improves dramatically – compromised passwords no longer grant access.
  • Compliance becomes easier – many regulations (e.g., NIST 800‑63B, PCI DSS) encourage strong mutual authentication.
  • Operational overhead stays low – automated short‑lived cert issuance and rotation eliminate manual key management.

If you’re looking to future‑proof your applications against credential‑theft attacks, make the shift today: empower your identity provider with device trust and let mTLS do the heavy lifting.