AI Engineer Europe 2026

Reverse engineering a Viking VoIP phone protocol with Claude Code — Boris Starkov, ElevenLabs

Loading slides…

Read the talk

Reverse engineering a Viking VoIP phone protocol with Claude Code

A phone-booth voice agent depended on configuring a legacy handset from a Mac. Port probing found the commands; capturing the vendor software’s traffic revealed how to make settings survive a reboot.

From a talk by Boris Starkov

Before you start: Basic familiarity with TCP connections, virtual machines and temporary versus persistent memory will help; no telephony or reverse-engineering background is required.

Pick up the phone and talk to an agent

Pick up the red telephone in the summit’s third-floor booth and an AI agent answers in Michael Caine’s voice. Boris Starkov, who works on voice agents at ElevenLabs, describes the voice as approved for the demo. The agent asks five questions about British AI history; answering at least one correctly earns ElevenLabs swag. The booth, introduced with a photograph of organizer swyx, is the visitor-facing result. Getting the phone configured was the harder project.

A red corded telephone beside an ElevenLabs sign in a red booth, with the presenter inset at lower left.Open full source frame
The red telephone inside the ElevenLabs demo booth.
0:150:32
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:15 · section reference included

A Windows configuration tool meets a Mac-only team

Under the retro exterior, the phone has an electronic controller. Its configuration software was compatible with Windows XP, while the ElevenLabs team had Macs. An attempt to use a virtual environment also encountered driver problems. The obstacle was access to the device’s configuration interface, before any voice-agent integration could begin.

The San Francisco team had bought the phone for an event a year earlier. Starkov reports that three senior software engineers working with ChatGPT could not get it configured, leaving it unused until the London project. After the phone arrived in London, he tried Claude Code as a partner for reverse engineering the hardware.

Slide titled “The Phone” shows a red handset over exposed electronics and lists the Viking K-1900D-IP’s interface and configuration constraints.Open full source frame
The Viking phone has no screen, buttons or web UI and uses Windows-only configuration software.

The project conversation also became the source for the presentation: Starkov supplied it to Claude Code and asked it to generate the slides. The walkthrough follows that investigation from finding an interface to replacing the vendor’s configuration workflow.

1:271:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:27 · section reference included

Find a configuration interface

The intended call path was Viking phone → Twilio → ElevenLabs agent. Starkov describes the handset as placing a call to a phone number, with Twilio handling the SIP-domain complexity between the phone and the agent. That was the booth’s chosen architecture, not a requirement for every ElevenLabs telephony integration.

To configure the handset from his Mac, Starkov connected the laptop and phone through a router. Claude Code began exploring the exposed network services with Nmap. Port 1001 was an early false lead: it turned out to be an electronics tunnel. Further probing found a port that could actually communicate with the phone.

Sending a probe string produced a response. That established a useful starting point: the device had a reachable interface whose behavior could be tested. Starkov could not find publicly searchable documentation for its protocol, so the next step was to infer the command structure from those responses.

3:203:28
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:20 · section reference included

Use error responses to enumerate commands

An invalid string came back prefixed with ER, followed by the submitted string. That gave the investigation a recognizable rejection response. Further experiments suggested that the phone accepted two-letter command codes: some pairs returned the familiar error, while others returned something different.

Once the command width was known, guessing could become enumeration. Claude Code generated a program that tried every two-letter combination and separated ordinary errors from interesting responses. The core of that search can be expressed in Python without assuming the phone’s packet framing:

python

from itertools import product


def command_candidates(alphabet: str):
    for letters in product(alphabet, repeat=2):
        yield "".join(letters).encode("ascii")


def classify_response(response: bytes) -> str:
    if not response:
        return "no response"
    if response.startswith(b"ER"):
        return "rejected"
    return "inspect"

The alphabet determines the candidate set. A non-error reply earns inspection; it does not, by itself, explain what a command does. Starkov reports that eight two-letter combinations returned non-error responses, which the investigation treated as valid commands.

Command names offered clues, but not complete semantics. Starkov tentatively associates SA with status. With a set of responsive commands in hand, the investigation moved from discovery to writing the credentials the phone would need to place a call.

5:225:38
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

5:22 · section reference included

Accepted settings are not necessarily saved settings

The configuration commands appeared to work. The phone accepted the settings intended to connect it to the Twilio SIP trunk. Then a reboot erased them: the writes had changed temporary memory without making the configuration persistent. A successful write response was not the success condition; surviving a reboot was.

TestObserved resultWhat it established
Write the settingsAcceptedThe configuration could be changed temporarily
Reboot the phoneSettings disappearedPersistence was still missing
Search additional commandsNo solution foundMore command guessing did not reveal the save operation

The search for persistence took hours and expanded to three-letter combinations and plausible words. None supplied the missing behavior. Starkov identifies this as roughly the point where the earlier San Francisco effort had stopped.

7:157:25
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:15 · section reference included

Capture a working implementation

The next move changed the source of evidence. Claude proposed running the vendor’s configuration software in a Windows virtual machine. Starkov encountered a Wi-Fi bridging and connectivity obstacle in his particular macOS setup, so Claude introduced a TCP proxy on the Mac. The proxy would provide a route to the phone and a place to observe the conversation.

The vendor application already knew how to configure the device. Instead of guessing more commands, the investigation could watch that application perform the operation:

  1. Run the phone’s configuration software inside the Windows VM.
  2. Direct its connection to the TCP proxy on the Mac.
  3. Relay the traffic from the proxy to the Viking phone and return the replies.
  4. Log the bytes in both directions to identify the missing protocol behavior.

This made the proxy both a connectivity workaround and an observation point. The Windows application still performed the known workflow; the capture exposed what it sent over the network.

Slide titled “The Proxy” shows Windows VM, Mac TCP proxy and Viking Phone boxes connected by arrows, with a note about logging bytes in both directions.Open full source frame
A Mac proxy relays and logs traffic between the Windows VM and Viking phone.
8:549:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:54 · section reference included

Infer the binary payload, then test the inference

The capture revealed a TS command carrying a binary payload that Claude had not previously understood. Most of its parameters became interpretable, but the checksum remained unexplained. The checksum occupied one byte, making a bounded search practical. Captured traffic supplied the transmitted data and the corresponding observed result, constraining the possible calculation.

Starkov describes the arithmetic in terms of a simple subtraction or addition involving a one-byte value. Although he also calls it encryption, the mechanism described here is a checksum calculation, not evidence of defeating a cryptographic protection scheme. The important experimental step came after proposing the calculation: Claude ran additional values through it to confirm the inferred behavior. The hypothesis was checked against more observations, rather than accepted because it fit one capture. At that point, Starkov considered the protocol substantially understood.

10:3310:50
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

10:33 · section reference included

Turn the recovered protocol into repeatable setup

Understanding the vendor application’s traffic supplied more than a command list. It revealed the sequence needed to make the configuration persistent—the operation that ordinary setting writes had failed to accomplish. Starkov also reports decoding a 256-byte memory area in the phone. Together, the command behavior, persistence sequence and memory information provided a basis for programming the device directly.

The next goal was to stop launching Windows for every configuration change. With the protocol understood, Starkov could factory-reset the phone and configure it directly. He packaged the recovered knowledge as an open-source Claude Code skill, so a subsequent setup could use that knowledge without repeating the VM-based discovery process.

The resulting phone connected through Twilio to ElevenLabs. Starkov describes that final integration as straightforward, though the walkthrough concentrates on the handset rather than the telephony implementation. Starkov estimates that reverse engineering the phone took a couple of days.

12:1712:24
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:17 · section reference included

From an inaccessible interface to a reusable skill

The recovered protocol had two layers: the text commands encountered through probing and the binary behavior exposed by capture. Starkov credits the simple one-byte checksum as a favorable condition; a harder protection scheme would have made the investigation much more difficult. The intermediate server between the Windows VM and phone was the decisive tool for understanding persistence.

The final slide describes AI as making reverse engineering ten times faster, but Starkov’s spoken explanation makes a different claim: for him, it made this demo possible. He was a software engineer without a security-engineering background, unfamiliar even with the 0x notation he encountered. There is no measured speedup comparison here; the result is a project he says he could not otherwise have completed.

Five conclusions describe text and binary protocol layers, a single-byte checksum, persistence, a MITM capture and the claim that AI made reverse engineering ten times faster.Open full source frame
The complete five-point “What We Learned” slide.

The lasting output was therefore both a working booth and a reusable configuration skill. Starkov proposes applying the same approach to other hardware whose vendor interface is inconvenient or inaccessible. In this case, the vendor software remained essential during discovery; the recovered protocol removed that dependency from subsequent configuration. The generalizable move is to turn observed device behavior into knowledge that can be reused.

14:2214:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:22 · section reference included

The human supplied the hands—and the physical connection

The audience first asks how many tokens the project consumed. Starkov gives a spending estimate instead: he estimates the project used $10–$100 of model spend through ElevenLabs’ corporate account. He does not provide a token count or an itemized usage measurement.

The next question asks whether Claude controlled the Windows software running in UTM. Starkov explains that he operated it manually, following Claude’s instructions. The same division of work applied to physical observations: Claude asked him to lift the handset and count the beeps. He initially reported three, then corrected the answer to two after Claude indicated that the expected possibilities were two or four. Claude orchestrated the investigation, while Starkov supplied actions and observations it could not obtain directly. Automating the VM with computer use was a possibility he considered, not something implemented in this project.

Asked where he had to unblock Claude, Starkov distinguishes physical intervention from protocol reasoning. He could reboot the phone and control the virtual machine. He says he could not independently resolve the algorithmic reverse-engineering problems. The working loop therefore combined model-directed experiments with human operation of the software and hardware.

The final question returns to the physical connection. The phone’s black cable carried Power over Ethernet (PoE): Ethernet connectivity and power from an adapter were combined to supply the handset. The phone connected through a router, and Starkov’s laptop joined the same router over Wi-Fi. With both devices on that network, Claude checked the connected devices to locate the phone before probing its ports. That reachable network connection was the foundation for every later experiment.

16:5717:01
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:57 · section reference included

Resources

Updates since the talk

Transcript download
Full transcript

Loading transcript…