The core constraint was staying Linux-native. Official paths meant the .NET media SDK, a Windows sidecar, or a third-party meeting bot, so I chose the harder route: understand the Teams wire protocol well enough to implement it directly.

To learn the protocol I took Microsoft’s own binaries apart — decompiling the native and .NET DLLs in Ghidra and ILSpy, live-debugging the Teams client with x64dbg and Frida, and reading the decrypted traffic in Wireshark. I wired the decompiler into an LLM over MCP so I could interrogate thousands of lines of decompiled code and get through the dead ends faster.

Then I rebuilt it on Linux: from-scratch Python implementations of Microsoft’s old binary framing and SOAP-over-the-wire protocols, an ICE agent to negotiate the UDP media path, SRTP key derivation, and a fix for two hidden four-second timers that were silently killing every connection attempt.

The bridge now joins a real meeting, plays audio people can hear, and captures the returning speech byte-for-byte after decryption. For the codec path, I drove Microsoft’s native decoder from Python and validated it against captured frames.

I kept the process disciplined for something this fiddly: I versioned every experiment, kept real captured calls as test fixtures, and refused to ship guesses for parts I couldn’t yet name. And so nothing was blocked on the hard path, I shipped a simpler working voice agent alongside it that any company agent can call to join a meeting and answer by voice.