Micdrop WebSocket Protocol Between Client and Server
WebSocket Protocol
Micdrop uses a simple custom protocol over WebSocket for real-time communication between the client and server.
sequenceDiagram
participant W as Client
participant B as Server
Note left of W: Audio setup
W -->> B: Create call
B -->> W: Message (First assistant message)
Note right of B: Start TTS : Generate voice for first message
B -->> W: Audio chunk (First assistant message)
Note left of W: Play assistant speech
B -->> W: Audio chunk (First assistant message)
Note right of B: Stop TTS - Sent audio for first message
loop
Note left of W: Wait until user speaks
W -->> B: StartSpeaking
Note right of B: Start STT : Transcribe user speech
W -->> B: Audio chunk (User speech)
W -->> B: Audio chunk (User speech)
Note left of W: Silence
W -->> B: StopSpeaking - User stops speaking
Note right of B: Stop STT : User speech transcribed
B -->> W: Message (Transcribed user message)
Note right of B: Start Agent : Generate answer
Note right of B: Start TTS : Generate voice for answer
B ->> W: Audio chunk (Assistant answer)
B ->> W: Audio chunk (Assistant answer)
B -->> W: Message (Assistant answer)
Note right of B: Stop Agent : Finished answering
B ->> W: Audio chunk (Assistant answer)
B ->> W: Audio chunk (Assistant answer)
Note right of B: Stop TTS : Sent audio for answer
Note left of W: Play assistant speech
end
Why WebSocket?
Micdrop talks to its server over a WebSocket because a browser calling your own backend needs just the server you already deploy. The browser detects speech and sends audio only while someone speaks, which keeps bandwidth low. WebRTC is worth its extra servers for group calls, video, phone lines and users on weak mobile networks.
Calls without a voice
A call configured without a text to speech carries no audio chunk back. The
answer arrives as a Message, followed by SkipAnswer, which is what tells the
client to stop waiting and listen again. A call configured without an agent
sends SkipAnswer right after the transcribed user message.
See Dictation and text-only calls.
Client Commands
The client can send the following commands to the server:
MicdropClientCommands.StartSpeaking- The user starts speakingMicdropClientCommands.StopSpeaking- The user stops speakingMicdropClientCommands.Mute- The user mutes the microphone
Server Commands
The server can send the following commands to the client:
MicdropServerCommands.Message- A settled message, from the user or the assistant.MicdropServerCommands.PartialAssistantMessage- The answer being written, before it is finished. Sent whenpartialMessagesis on.MicdropServerCommands.CancelLastUserMessage- Cancel the last user message.MicdropServerCommands.SkipAnswer- Notify that the last generated answer was ignored, itβs listening again.MicdropServerCommands.EndCall- End the call.MicdropServerCommands.ToolCall- A tool the agent called, with its result.MicdropServerCommands.Classification- What the classifier made of the user turn, as JSON. Sent when the server has a classifier andclassifierOptions.sendToClientis on.