commit a42b3e8d737063441571daa51c3eb127c70ec0da
Author: Jeuner <62662523+Jeuners@users.noreply.github.com>
Date:   Mon Sep 7 13:39:51 2026 +0200

    feat: initial commit of astra local voice agent
    
    Lokaler deutscher Sprachagent für Apple Silicon: Pipecat-Pipeline mit
    Nemotron-ASR (MLX), Qwen über natives Ollama /api/chat, und Pocket TTS.
    Loopback-only WebRTC-Server mit Origin/Host-Härtung.
    
    Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
    Claude-Session: https://claude.ai/code/session_01LVgSHNHdRx3UNTBodFmhRA

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..f3e0a3e
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,10 @@
+.venv/
+.cache/
+.runtime/
+__pycache__/
+.pytest_cache/
+.ruff_cache/
+.coverage
+htmlcov/
+*.log
+.DS_Store
diff --git a/astra/__init__.py b/astra/__init__.py
new file mode 100644
index 0000000..e387b61
--- /dev/null
+++ b/astra/__init__.py
@@ -0,0 +1 @@
+"""Astra: a local German voice companion for Apple Silicon."""
diff --git a/astra/core.py b/astra/core.py
new file mode 100644
index 0000000..27406c6
--- /dev/null
+++ b/astra/core.py
@@ -0,0 +1,80 @@
+"""Configuration and pure request policy, independent of audio hardware."""
+
+import os
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class Settings:
+    model: str = "qwen3.5:latest"
+    ollama_url: str = "http://127.0.0.1:11434"
+    stt_model: str = "mlx-community/nemotron-3.5-asr-streaming-0.6b-8bit"
+    tts_language: str = "german"
+    voice: str = "alba"
+    port: int = 7860
+    context_tokens: int = 4096
+
+    @classmethod
+    def from_env(cls):
+        return cls(
+            model=os.getenv("ASTRA_MODEL", cls.model),
+            ollama_url=os.getenv("ASTRA_OLLAMA_URL", cls.ollama_url).rstrip("/"),
+            stt_model=os.getenv("ASTRA_STT_MODEL", cls.stt_model),
+            tts_language=os.getenv("ASTRA_TTS_LANGUAGE", cls.tts_language),
+            voice=os.getenv("ASTRA_VOICE", cls.voice),
+            port=int(os.getenv("ASTRA_PORT", cls.port)),
+        )
+
+
+SYSTEM_PROMPT = (
+    "Du bist Astra, ein freundlicher deutschsprachiger Gesprächsassistent. "
+    "Antworte natürlich und knapp, normalerweise in ein bis drei kurzen Sätzen. "
+    "Deine Antwort wird vorgelesen: kein Markdown, keine Sternchen, keine Listen. "
+    "Sprich Zahlen und Abkürzungen verständlich aus. Stelle bei Bedarf eine kurze Rückfrage. "
+    "Du hast keine Werkzeuge, keinen Internetzugang und keinen Zugriff auf Dateien oder Apps. "
+    "Behaupte nicht, Aktionen ausgeführt zu haben."
+)
+
+
+def trim_messages(messages: list[dict], max_chars: int = 10000) -> list[dict]:
+    """Retain recent whole turns within a conservative context character budget."""
+    system = [dict(m) for m in messages if m["role"] == "system"][:1]
+    if system:
+        system[0]["content"] = system[0]["content"][: max_chars // 2]
+    budget = max_chars - sum(len(m["content"]) for m in system)
+    turns = []
+    for message in reversed(messages):
+        if message["role"] not in ("user", "assistant"):
+            continue
+        content = message.get("content")
+        if not isinstance(content, str) or not content:
+            continue
+        if len(content) > budget:
+            if not turns:
+                turns.append({"role": message["role"], "content": content[-budget:]})
+            break
+        turns.append({"role": message["role"], "content": content})
+        budget -= len(content)
+    turns.reverse()
+    while turns and turns[0]["role"] != "user":
+        turns.pop(0)
+    return system + turns
+
+
+def build_request(settings: Settings, messages: list[dict]) -> dict:
+    return {
+        "model": settings.model,
+        "messages": trim_messages(messages),
+        "think": False,
+        "stream": True,
+        "keep_alive": -1,
+        "options": {
+            "num_ctx": settings.context_tokens,
+            "num_predict": 256,
+            "temperature": 0.6,
+        },
+    }
+
+
+def local_origin_allowed(origin: str, port: int = 7860) -> bool:
+    return origin in {f"http://localhost:{port}", f"http://127.0.0.1:{port}"}
diff --git a/astra/inference.py b/astra/inference.py
new file mode 100644
index 0000000..f1f81fa
--- /dev/null
+++ b/astra/inference.py
@@ -0,0 +1,111 @@
+"""Thread-confined model inference. No audio is stored on disk."""
+
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+
+import numpy as np
+
+from astra.core import Settings
+
+
+async def on_executor(executor, function, *args):
+    """Finish an in-flight native call before allowing its owner to be reused."""
+    future = asyncio.get_running_loop().run_in_executor(executor, function, *args)
+    try:
+        return await asyncio.shield(future)
+    except asyncio.CancelledError:
+        await asyncio.shield(future)
+        raise
+
+
+def drain_stream(stream):
+    """Let Pocket TTS join its internal threads before reusing the model.
+
+    Its upstream generator only joins on normal exhaustion, not generator.close().
+    Discard the unplayed remainder while the audio transport stops immediately.
+    """
+    for _ in stream:
+        pass
+
+
+class Models:
+    def __init__(self, settings: Settings):
+        self.settings = settings
+        self.stt_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="astra-stt")
+        self.tts_executor = ThreadPoolExecutor(max_workers=1, thread_name_prefix="astra-tts")
+        self.stt = None
+        self.tts = None
+        self.voice = None
+
+    def load_stt(self):
+        from mlx_audio.stt import load
+
+        self.stt = load(self.settings.stt_model)
+        # Ensure this installed MLX version really exposes incremental microphone APIs.
+        from mlx_audio.stt.models.nemotron_asr.audio import StreamingLogMelSpectrogram  # noqa: F401
+        from mlx_audio.stt.models.nemotron_asr.streaming import (
+            ConformerStreamingState,  # noqa: F401
+        )
+
+    def load_tts(self):
+        from pocket_tts import TTSModel
+
+        self.tts = TTSModel.load_model(language=self.settings.tts_language)
+        self.voice = self.tts.get_state_for_audio_prompt(self.settings.voice)
+
+    async def close(self):
+        await asyncio.to_thread(self.stt_executor.shutdown, wait=True, cancel_futures=True)
+        await asyncio.to_thread(self.tts_executor.shutdown, wait=True, cancel_futures=True)
+
+
+class Recognizer:
+    """One utterance's incremental mel, encoder and RNN-T decoder state.
+
+    All methods run on Models.stt_executor, including construction. Encoder and
+    decoder caches survive microphone chunks and are discarded at utterance end.
+    """
+
+    def __init__(self, model):
+        from mlx_audio.stt.models.nemotron_asr.audio import StreamingLogMelSpectrogram
+        from mlx_audio.stt.models.nemotron_asr.streaming import ConformerStreamingState
+
+        self.model = model
+        self.mel = StreamingLogMelSpectrogram(model.preprocessor_config)
+        self.encoder = ConformerStreamingState(model.encoder, att_context_size=[56, 3])
+        self.last_token = model.blank_id
+        self.hidden = None
+        self.tokens = []
+        self.closed = False
+
+    def push(self, pcm: bytes, final: bool = False) -> str:
+        import mlx.core as mx
+        from mlx_audio.stt.models.nemotron_asr import tokenizer
+
+        if self.closed:
+            raise RuntimeError("Utterance already closed")
+        samples = mx.array(np.frombuffer(pcm, dtype="<i2").astype(np.float32) / 32768.0)
+        mel = self.mel.push(samples, final=final)
+        for encoded in self.encoder.push(mel, final=final):
+            prompted = self.model.apply_prompt(encoded, "de-DE")
+            self.encoder.materialize(prompted)
+            for frame_index in range(prompted.shape[1]):
+                feature = prompted[:, frame_index : frame_index + 1]
+                for _ in range(self.model.max_symbols or 10):
+                    token = (
+                        mx.array([[self.last_token]], dtype=mx.int32)
+                        if self.last_token != self.model.blank_id
+                        else None
+                    )
+                    output, (h, c) = self.model.decoder(token, self.hidden)
+                    prediction = int(
+                        mx.argmax(self.model.joint(feature, output.astype(feature.dtype)))
+                    )
+                    if prediction == self.model.blank_id:
+                        break
+                    self.last_token = prediction
+                    self.hidden = (h.astype(feature.dtype), c.astype(feature.dtype))
+                    mx.eval(*self.hidden)
+                    if not tokenizer.is_special_token(prediction, self.model.vocabulary):
+                        self.tokens.append(prediction)
+        self.closed = final
+        return tokenizer.decode(self.tokens, self.model.vocabulary).strip()
diff --git a/astra/prepare.py b/astra/prepare.py
new file mode 100644
index 0000000..1d02b03
--- /dev/null
+++ b/astra/prepare.py
@@ -0,0 +1,95 @@
+"""Download and warm local models; never opens the microphone."""
+
+import asyncio
+import json
+import time
+from pathlib import Path
+
+import httpx
+
+from astra.core import Settings, build_request
+from astra.inference import Models, Recognizer, on_executor
+
+
+async def main():
+    import nltk
+
+    nltk_path = Path(__file__).resolve().parent.parent / ".cache" / "nltk"
+    nltk_path.mkdir(parents=True, exist_ok=True, mode=0o700)
+    if not nltk.download("punkt_tab", download_dir=str(nltk_path), quiet=True):
+        raise RuntimeError("Satzsegmentierung konnte nicht heruntergeladen werden")
+    settings = Settings.from_env()
+    models = Models(settings)
+    try:
+        print("Lade Nemotron für deutsche Streaming-Erkennung …", flush=True)
+        await on_executor(models.stt_executor, models.load_stt)
+        recognizer = await on_executor(models.stt_executor, Recognizer, models.stt)
+        await on_executor(models.stt_executor, recognizer.push, bytes(16000), True)
+        print("Nemotron bereit. Lade deutsche Pocket-TTS-Stimme …", flush=True)
+        await on_executor(models.tts_executor, models.load_tts)
+
+        def synthesize():
+            start = time.monotonic()
+            chunks = []
+            first_ms = None
+            for chunk in models.tts.generate_audio_stream(
+                models.voice,
+                "Hallo, ich bin Astra. Ich laufe vollständig auf deinem Mac.",
+                copy_state=True,
+            ):
+                if first_ms is None:
+                    first_ms = round((time.monotonic() - start) * 1000)
+                chunks.append(chunk.numpy())
+            import numpy as np
+
+            samples = np.concatenate(chunks)
+            return samples, {
+                "first_audio_ms": first_ms,
+                "seconds": round(len(samples) / models.tts.sample_rate, 2),
+                "generation_ms": round((time.monotonic() - start) * 1000),
+            }
+
+        samples, metrics = await on_executor(models.tts_executor, synthesize)
+        print(f"Pocket TTS: {json.dumps(metrics)}", flush=True)
+        import numpy as np
+        from scipy.signal import resample_poly
+
+        audio16 = resample_poly(samples, 2, 3)
+        pcm = (np.clip(audio16, -1, 1) * 32767).astype("<i2").tobytes()
+        recognizer = await on_executor(models.stt_executor, Recognizer, models.stt)
+        started = time.monotonic()
+        partials = []
+        for offset in range(0, len(pcm), 10240):
+            text = await on_executor(
+                models.stt_executor, recognizer.push, pcm[offset : offset + 10240], False
+            )
+            if text and (not partials or text != partials[-1]):
+                partials.append(text)
+        text = await on_executor(models.stt_executor, recognizer.push, b"", True)
+        print(f"ASR-Rücktest: {text}", flush=True)
+        print(
+            f"ASR: {len(partials)} Live-Zwischenstände, {round((time.monotonic() - started) * 1000)} ms",
+            flush=True,
+        )
+        if "astra" not in text.lower() or "mac" not in text.lower():
+            raise RuntimeError("Deutscher Sprach-Rücktest fehlgeschlagen")
+        async with httpx.AsyncClient(timeout=180) as client:
+            payload = build_request(
+                settings, [{"role": "user", "content": "Sage kurz Hallo auf Deutsch."}]
+            )
+            response = await client.post(f"{settings.ollama_url}/api/chat", json=payload)
+            response.raise_for_status()
+            events = [json.loads(line) for line in response.text.splitlines() if line]
+            if any(e.get("message", {}).get("thinking") for e in events):
+                raise RuntimeError("Thinking ist nicht deaktiviert")
+            answer = "".join(e.get("message", {}).get("content", "") for e in events)
+            if not answer or not events[-1].get("done"):
+                raise RuntimeError("Ollama hat keine vollständige Antwort geliefert")
+            print(f"Ollama ohne Thinking: {answer}", flush=True)
+        print("Alle drei lokalen Modelle erfolgreich geprüft.", flush=True)
+    finally:
+        await models.close()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/astra/server.py b/astra/server.py
new file mode 100644
index 0000000..f18c7e3
--- /dev/null
+++ b/astra/server.py
@@ -0,0 +1,322 @@
+"""Loopback-only WebRTC voice app. One live session shares the warm models."""
+
+import asyncio
+import contextlib
+import os
+import sys
+from collections import deque
+from contextlib import asynccontextmanager
+from pathlib import Path
+from typing import Literal
+
+import httpx
+import uvicorn
+from fastapi import FastAPI, HTTPException, Request
+from fastapi.responses import FileResponse, JSONResponse
+from fastapi.staticfiles import StaticFiles
+from loguru import logger
+from pydantic import BaseModel, Field
+
+from astra.core import SYSTEM_PROMPT, Settings, build_request, local_origin_allowed
+from astra.inference import Models, on_executor
+
+ROOT = Path(__file__).resolve().parent.parent
+
+
+class Offer(BaseModel):
+    sdp: str = Field(min_length=10, max_length=65536)
+    type: Literal["offer"]
+
+
+class Disconnect(BaseModel):
+    pc_id: str = Field(max_length=100)
+
+
+async def run_voice(connection, models, config):
+    from pipecat.audio.vad.silero import SileroVADAnalyzer
+    from pipecat.audio.vad.vad_analyzer import VADParams
+    from pipecat.frames.frames import (
+        BotStartedSpeakingFrame,
+        BotStoppedSpeakingFrame,
+        InterruptionFrame,
+        LLMFullResponseStartFrame,
+        VADUserStartedSpeakingFrame,
+    )
+    from pipecat.observers.base_observer import BaseObserver
+    from pipecat.pipeline.pipeline import Pipeline
+    from pipecat.pipeline.worker import PipelineParams, PipelineWorker
+    from pipecat.processors.aggregators.llm_context import LLMContext
+    from pipecat.processors.aggregators.llm_response_universal import (
+        LLMContextAggregatorPair,
+        LLMUserAggregatorParams,
+    )
+    from pipecat.transports.base_transport import TransportParams
+    from pipecat.transports.smallwebrtc.transport import SmallWebRTCTransport
+    from pipecat.workers.runner import WorkerRunner
+
+    from astra.services import LocalPocketTTSService, NativeOllamaService, NemotronSTTService
+
+    notify = connection.send_app_message
+
+    class UIObserver(BaseObserver):
+        def __init__(self):
+            super().__init__()
+            self.seen = deque(maxlen=128)
+
+        async def on_push_frame(self, data):
+            frame = data.frame
+            relevant = (
+                BotStartedSpeakingFrame,
+                BotStoppedSpeakingFrame,
+                InterruptionFrame,
+                LLMFullResponseStartFrame,
+                VADUserStartedSpeakingFrame,
+            )
+            if not isinstance(frame, relevant) or frame.id in self.seen:
+                return
+            self.seen.append(frame.id)
+            if isinstance(frame, BotStartedSpeakingFrame):
+                notify({"type": "state", "state": "speaking"})
+            elif isinstance(frame, LLMFullResponseStartFrame):
+                notify({"type": "state", "state": "responding"})
+            else:
+                notify({"type": "state", "state": "listening"})
+
+    transport = SmallWebRTCTransport(
+        connection,
+        TransportParams(
+            audio_in_enabled=True,
+            audio_out_enabled=True,
+            audio_in_sample_rate=16000,
+            audio_out_sample_rate=24000,
+        ),
+    )
+    context = LLMContext([{"role": "system", "content": SYSTEM_PROMPT}])
+    aggregators = LLMContextAggregatorPair(
+        context,
+        user_params=LLMUserAggregatorParams(
+            vad_analyzer=SileroVADAnalyzer(
+                params=VADParams(
+                    confidence=0.7,
+                    start_secs=0.1,
+                    stop_secs=0.7,
+                    min_volume=0.6,
+                )
+            ),
+        ),
+    )
+    stt = NemotronSTTService(models, notify)
+    llm = NativeOllamaService(config, notify)
+    tts = LocalPocketTTSService(models)
+    pipeline = Pipeline(
+        [
+            transport.input(),
+            stt,
+            aggregators.user(),
+            llm,
+            tts,
+            transport.output(),
+            aggregators.assistant(),
+        ]
+    )
+    worker = PipelineWorker(
+        pipeline,
+        params=PipelineParams(audio_in_sample_rate=16000, audio_out_sample_rate=24000),
+        enable_rtvi=False,
+        idle_timeout_secs=None,
+        observers=[UIObserver()],
+    )
+
+    @transport.event_handler("on_client_connected")
+    async def connected(transport, client):
+        notify({"type": "state", "state": "listening"})
+
+    @transport.event_handler("on_client_disconnected")
+    async def disconnected(transport, client):
+        await worker.cancel()
+
+    @aggregators.user().event_handler("on_user_turn_stopped")
+    async def user_turn(aggregator, strategy, message):
+        if message.content:
+            notify({"type": "transcript", "role": "user", "text": message.content})
+
+    @aggregators.assistant().event_handler("on_assistant_turn_stopped")
+    async def assistant_turn(aggregator, message):
+        if message.content:
+            notify(
+                {
+                    "type": "transcript",
+                    "role": "assistant",
+                    "text": message.content,
+                    "interrupted": message.interrupted,
+                }
+            )
+
+    @worker.event_handler("on_pipeline_error")
+    async def error(worker, frame):
+        notify({"type": "error", "message": frame.error})
+
+    runner = WorkerRunner(handle_sigint=False)
+    await runner.add_workers(worker)
+    await runner.run()
+
+
+def create_app(config=None, *, load_models=True):
+    config = config or Settings.from_env()
+    models = Models(config)
+    status = {"ready": False, "stage": "starting", "error": None}
+    sessions = {}
+    lock = asyncio.Lock()
+
+    async def warmup():
+        try:
+            status["stage"] = "Gesprächssteuerung wird vorbereitet"
+
+            def prepare_pipeline():
+                import nltk
+                from pipecat.audio.turn.smart_turn.local_smart_turn_v3 import (
+                    LocalSmartTurnAnalyzerV3,
+                )
+                from pipecat.audio.vad.silero import SileroVADAnalyzer
+
+                from astra import services  # noqa: F401
+
+                nltk.data.path.insert(0, str(ROOT / ".cache" / "nltk"))
+                nltk.data.find("tokenizers/punkt_tab")
+                nltk.sent_tokenize("Hallo. Alles bereit.")
+                LocalSmartTurnAnalyzerV3()
+                SileroVADAnalyzer()
+
+            await asyncio.to_thread(prepare_pipeline)
+            status["stage"] = "Spracherkennung wird geladen"
+            await on_executor(models.stt_executor, models.load_stt)
+            status["stage"] = "Deutsche Stimme wird geladen"
+            await on_executor(models.tts_executor, models.load_tts)
+            status["stage"] = "Qwen wird vorbereitet"
+            async with httpx.AsyncClient(timeout=180) as client:
+                payload = build_request(config, [{"role": "user", "content": "Sage Hallo."}])
+                payload["stream"] = False
+                payload["options"]["num_predict"] = 12
+                response = await client.post(f"{config.ollama_url}/api/chat", json=payload)
+                response.raise_for_status()
+                result = response.json()
+                if result.get("error") or not result.get("message", {}).get("content"):
+                    raise RuntimeError(f"Ollama: {result.get('error', 'Keine Antwort')}")
+                if result["message"].get("thinking"):
+                    raise RuntimeError("Ollama hat Thinking nicht ausgeschaltet")
+            status.update(ready=True, stage="Bereit")
+            logger.info("Astra bereit auf http://localhost:{}", config.port)
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            logger.exception("Modelle konnten nicht geladen werden")
+            message = str(exc) or type(exc).__name__
+            if isinstance(exc, httpx.TimeoutException):
+                message = "Ollama antwortet nicht rechtzeitig. Bitte den lokalen Modelldienst prüfen."
+            status.update(ready=False, stage="Start fehlgeschlagen", error=message)
+
+    @asynccontextmanager
+    async def lifespan(app):
+        warming = asyncio.create_task(warmup()) if load_models else None
+        try:
+            yield
+        finally:
+            if warming:
+                warming.cancel()
+                with contextlib.suppress(asyncio.CancelledError):
+                    await warming
+            for connection, task in list(sessions.values()):
+                await connection.disconnect()
+                task.cancel()
+                with contextlib.suppress(asyncio.CancelledError):
+                    await task
+            await models.close()
+
+    app = FastAPI(lifespan=lifespan, docs_url=None, redoc_url=None, openapi_url=None)
+    app.state.status = status
+
+    @app.middleware("http")
+    async def local_only(request: Request, call_next):
+        if request.url.hostname not in {"localhost", "127.0.0.1"}:
+            return JSONResponse({"detail": "Nur lokal erreichbar"}, status_code=403)
+        if request.method == "POST" and not local_origin_allowed(
+            request.headers.get("origin", ""), config.port
+        ):
+            return JSONResponse({"detail": "Ungültiger Ursprung"}, status_code=403)
+        response = await call_next(request)
+        response.headers["Cache-Control"] = "no-store"
+        response.headers["X-Content-Type-Options"] = "nosniff"
+        response.headers["Referrer-Policy"] = "no-referrer"
+        response.headers["Content-Security-Policy"] = (
+            "default-src 'self'; script-src 'self'; style-src 'self'; "
+            "connect-src 'self'; media-src 'self' blob:; img-src 'self' data:; frame-ancestors 'none'"
+        )
+        return response
+
+    @app.get("/")
+    async def index():
+        return FileResponse(ROOT / "web" / "index.html")
+
+    @app.get("/api/status")
+    async def health():
+        return {**status, "busy": bool(sessions), "model": config.model, "thinking": False}
+
+    @app.post("/api/offer")
+    async def offer(body: Offer):
+        if not status["ready"]:
+            raise HTTPException(503, status["error"] or status["stage"])
+        async with lock:
+            if sessions:
+                raise HTTPException(
+                    409, "Ein Gespräch läuft bereits. Beende es im anderen Fenster."
+                )
+            from pipecat.transports.smallwebrtc.connection import SmallWebRTCConnection
+
+            connection = SmallWebRTCConnection(ice_servers=[], connection_timeout_secs=20)
+            try:
+                await connection.initialize(body.sdp, body.type)
+            except Exception:
+                await connection.disconnect()
+                raise
+
+            async def session():
+                try:
+                    await run_voice(connection, models, config)
+                except asyncio.CancelledError:
+                    raise
+                except Exception as exc:
+                    logger.exception("Gespräch fehlgeschlagen")
+                    connection.send_app_message({"type": "error", "message": str(exc)})
+                finally:
+                    await connection.disconnect()
+                    sessions.pop(connection.pc_id, None)
+
+            task = asyncio.create_task(session())
+            sessions[connection.pc_id] = (connection, task)
+            return connection.get_answer()
+
+    @app.post("/api/disconnect")
+    async def disconnect(body: Disconnect):
+        pair = sessions.get(body.pc_id)
+        if pair:
+            connection, task = pair
+            await connection.disconnect()
+            task.cancel()
+            with contextlib.suppress(asyncio.CancelledError):
+                await task
+        return {"ok": True}
+
+    app.mount("/static", StaticFiles(directory=ROOT / "web"), name="static")
+    return app
+
+
+def main():
+    os.environ.setdefault("HF_HUB_OFFLINE", "1")
+    logger.remove()
+    logger.add(sys.stderr, level=os.getenv("ASTRA_LOG_LEVEL", "INFO"))
+    config = Settings.from_env()
+    uvicorn.run(create_app(config), host="127.0.0.1", port=config.port, access_log=False)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/astra/services.py b/astra/services.py
new file mode 100644
index 0000000..1429fe8
--- /dev/null
+++ b/astra/services.py
@@ -0,0 +1,262 @@
+"""Pipecat adapters for native Ollama, live Nemotron ASR, and Pocket TTS."""
+
+import asyncio
+import json
+import time
+from contextlib import aclosing
+
+import httpx
+from loguru import logger
+from openai.types.chat import ChatCompletionChunk
+from pipecat.frames.frames import (
+    CancelFrame,
+    EndFrame,
+    ErrorFrame,
+    Frame,
+    InputAudioRawFrame,
+    InterimTranscriptionFrame,
+    StartFrame,
+    TranscriptionFrame,
+    VADUserStartedSpeakingFrame,
+    VADUserStoppedSpeakingFrame,
+)
+from pipecat.processors.frame_processor import FrameDirection
+from pipecat.services.openai.llm import OpenAILLMService
+from pipecat.services.settings import STTSettings, TTSSettings
+from pipecat.services.stt_service import STTService
+from pipecat.services.tts_service import TTSService
+from pipecat.utils.time import time_now_iso8601
+
+from astra.core import Settings, build_request
+from astra.inference import Models, Recognizer, drain_stream, on_executor
+
+
+class NativeOllamaService(OpenAILLMService):
+    """Use /api/chat so think=False cannot be lost in compatibility translation."""
+
+    supports_developer_role = False
+
+    def __init__(self, config: Settings, notify):
+        super().__init__(
+            api_key="local",
+            base_url=f"{config.ollama_url}/v1",
+            settings=self.Settings(model=config.model),
+        )
+        self.config = config
+        self.notify = notify
+
+    async def get_chat_completions(self, context):
+        payload = build_request(self.config, context.get_messages())
+        context.set_messages(payload["messages"])
+
+        async def chunks():
+            start = time.monotonic()
+            first = True
+            complete = False
+            async with httpx.AsyncClient(timeout=httpx.Timeout(90, connect=5)) as client:
+                async with client.stream(
+                    "POST", f"{self.config.ollama_url}/api/chat", json=payload
+                ) as response:
+                    response.raise_for_status()
+                    async for line in response.aiter_lines():
+                        if not line:
+                            continue
+                        event = json.loads(line)
+                        if event.get("error"):
+                            raise RuntimeError(event["error"])
+                        message = event.get("message", {})
+                        if message.get("thinking"):
+                            raise RuntimeError("Ollama liefert Thinking trotz think=false.")
+                        content = message.get("content", "")
+                        if content:
+                            if first:
+                                self.notify(
+                                    {
+                                        "type": "metric",
+                                        "name": "llm_ms",
+                                        "value": round((time.monotonic() - start) * 1000),
+                                    }
+                                )
+                                first = False
+                            yield ChatCompletionChunk(
+                                id="local",
+                                object="chat.completion.chunk",
+                                created=int(time.time()),
+                                model=self.config.model,
+                                choices=[
+                                    {
+                                        "index": 0,
+                                        "delta": {"content": content},
+                                        "finish_reason": None,
+                                    }
+                                ],
+                            )
+                        if event.get("done"):
+                            complete = True
+                    if not complete:
+                        raise RuntimeError("Ollama hat den Antwortstream vorzeitig geschlossen.")
+
+        return chunks()
+
+
+class NemotronSTTService(STTService):
+    def __init__(self, models: Models, notify):
+        super().__init__(
+            sample_rate=16000,
+            audio_passthrough=True,
+            ttfs_p99_latency=1.0,
+            settings=STTSettings(model=models.settings.stt_model, language="de-DE"),
+        )
+        self.models = models
+        self.notify = notify
+        self.queue = asyncio.Queue(maxsize=100)
+        self.worker = None
+        self.active = False
+        self.preroll = bytearray()
+        self.pending = bytearray()
+        self.failed = False
+
+    async def run_stt(self, audio):
+        # Audio is fed by process_audio_frame to avoid blocking VAD on inference.
+        if False:
+            yield None
+
+    async def start(self, frame: StartFrame):
+        await super().start(frame)
+        self.worker = self.create_task(self._worker())
+
+    async def stop(self, frame: EndFrame):
+        if self.worker:
+            await self.cancel_task(self.worker)
+            self.worker = None
+        await super().stop(frame)
+
+    async def cancel(self, frame: CancelFrame):
+        if self.worker:
+            await self.cancel_task(self.worker)
+            self.worker = None
+        await super().cancel(frame)
+
+    def _enqueue(self, item):
+        if self.failed:
+            return
+        try:
+            self.queue.put_nowait(item)
+        except asyncio.QueueFull:
+            self.failed = True
+            self.notify(
+                {"type": "error", "message": "Spracherkennung überlastet. Bitte neu verbinden."}
+            )
+
+    async def process_audio_frame(self, frame: InputAudioRawFrame, direction: FrameDirection):
+        if frame.sample_rate != 16000 or frame.num_channels != 1:
+            raise ValueError("Nemotron benötigt 16 kHz Mono-PCM")
+        if self.active:
+            self.pending.extend(frame.audio)
+            while len(self.pending) >= 10240:
+                self._enqueue(("audio", bytes(self.pending[:10240])))
+                del self.pending[:10240]
+        else:
+            self.preroll.extend(frame.audio)
+            del self.preroll[:-6400]  # 200 ms, includes VAD start delay.
+
+    async def process_frame(self, frame: Frame, direction: FrameDirection):
+        if isinstance(frame, VADUserStartedSpeakingFrame) and not self.active:
+            self.active = True
+            self._enqueue(("start", bytes(self.preroll)))
+            self.preroll.clear()
+        elif isinstance(frame, VADUserStoppedSpeakingFrame) and self.active:
+            self.active = False
+            self._enqueue(("final", bytes(self.pending)))
+            self.pending.clear()
+        await super().process_frame(frame, direction)
+
+    async def _worker(self):
+        recognizer = None
+        previous = ""
+        try:
+            while True:
+                kind, pcm = await self.queue.get()
+                logger.debug("Nemotron chunk: {}, {} bytes", kind, len(pcm))
+                if kind == "start":
+                    recognizer = await on_executor(
+                        self.models.stt_executor, Recognizer, self.models.stt
+                    )
+                    previous = ""
+                if recognizer is None:
+                    continue
+                text = await on_executor(
+                    self.models.stt_executor, recognizer.push, pcm, kind == "final"
+                )
+                logger.debug("Nemotron result: {}, {} characters", kind, len(text))
+                if kind == "final":
+                    if text:
+                        await self.push_frame(
+                            TranscriptionFrame(
+                                text=text,
+                                user_id="local",
+                                timestamp=time_now_iso8601(),
+                                finalized=True,
+                            )
+                        )
+                    self.notify({"type": "partial", "text": ""})
+                    recognizer = None
+                elif text and text != previous:
+                    await self.push_frame(
+                        InterimTranscriptionFrame(
+                            text=text, user_id="local", timestamp=time_now_iso8601()
+                        )
+                    )
+                    self.notify({"type": "partial", "text": text})
+                    previous = text
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            self.failed = True
+            self.notify({"type": "error", "message": f"Spracherkennung: {exc}"})
+            await self.push_error_frame(ErrorFrame(error=f"Nemotron: {exc}", fatal=True))
+
+
+def next_chunk(stream):
+    return next(stream, None)
+
+
+class LocalPocketTTSService(TTSService):
+    def __init__(self, models: Models):
+        super().__init__(
+            push_start_frame=True,
+            push_stop_frames=True,
+            settings=TTSSettings(model="pocket-tts", voice=models.settings.voice, language="de"),
+        )
+        self.models = models
+
+    async def run_tts(self, text: str, context_id: str):
+        import torch
+
+        stream = self.models.tts.generate_audio_stream(self.models.voice, text, copy_state=True)
+
+        async def audio():
+            try:
+                while True:
+                    chunk = await on_executor(self.models.tts_executor, next_chunk, stream)
+                    if chunk is None:
+                        break
+                    yield (chunk.clamp(-1, 1) * 32767).to(torch.int16).numpy().tobytes()
+            finally:
+                await on_executor(self.models.tts_executor, drain_stream, stream)
+
+        try:
+            async with aclosing(audio()) as audio_stream:
+                async with aclosing(
+                    self._stream_audio_frames_from_iterator(
+                        audio_stream,
+                        in_sample_rate=self.models.tts.sample_rate,
+                        context_id=context_id,
+                    )
+                ) as frames:
+                    async for frame in frames:
+                        yield frame
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            yield ErrorFrame(error=f"Sprachausgabe: {exc}")
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 0000000..4a5d407
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,25 @@
+[project]
+name = "astra-local-voice"
+version = "0.1.0"
+description = "Lokaler deutscher Sprachagent für Apple Silicon und Ollama"
+requires-python = ">=3.12,<3.13"
+dependencies = [
+  "pipecat-ai[silero,webrtc,pocket-tts]",
+  "mlx-audio",
+  "fastapi",
+  "uvicorn",
+  "httpx",
+]
+
+[dependency-groups]
+dev = ["pytest", "pytest-asyncio", "pytest-cov", "playwright", "ruff"]
+
+[tool.pytest.ini_options]
+asyncio_mode = "auto"
+testpaths = ["tests"]
+
+[tool.ruff]
+line-length = 100
+
+[tool.ruff.lint]
+select = ["E4", "E7", "E9", "F", "I", "B", "UP"]
diff --git a/scripts/browser_check.py b/scripts/browser_check.py
new file mode 100644
index 0000000..b0ca3c4
--- /dev/null
+++ b/scripts/browser_check.py
@@ -0,0 +1,92 @@
+"""Exercise actual WebRTC with synthetic German mic audio, never the real mic."""
+
+import asyncio
+import json
+from pathlib import Path
+
+import numpy as np
+from playwright.async_api import async_playwright, expect
+from scipy.io.wavfile import write
+
+ROOT = Path(__file__).resolve().parent.parent
+ARTIFACTS = ROOT / ".runtime"
+
+
+def make_fixture():
+    from pocket_tts import TTSModel
+
+    model = TTSModel.load_model(language="german")
+    voice = model.get_state_for_audio_prompt("alba")
+    samples = model.generate_audio(voice, "Hallo Astra. Wie heißt du?").numpy()
+    # The initial silence lets the pipeline finish establishing WebRTC.
+    audio = np.concatenate([np.zeros(24000 * 15), samples, np.zeros(24000 * 35)])
+    ARTIFACTS.mkdir(exist_ok=True)
+    path = ARTIFACTS / "test-microphone.wav"
+    write(path, 24000, (np.clip(audio, -1, 1) * 32767).astype(np.int16))
+    return path
+
+
+async def main():
+    fixture = ARTIFACTS / "test-microphone.wav"
+    if not fixture.exists():
+        fixture = await asyncio.to_thread(make_fixture)
+    async with async_playwright() as playwright:
+        browser = await playwright.chromium.launch(
+            executable_path="/Applications/Google Chrome.app/Contents/MacOS/Google Chrome",
+            headless=True,
+            args=[
+                "--use-fake-ui-for-media-stream",
+                "--use-fake-device-for-media-stream",
+                f"--use-file-for-fake-audio-capture={fixture}%noloop",
+                "--autoplay-policy=no-user-gesture-required",
+            ],
+        )
+        page = await browser.new_page(viewport={"width": 1440, "height": 1000})
+        errors = []
+        page.on("pageerror", lambda error: errors.append(str(error)))
+        await page.goto("http://localhost:7860")
+        await expect(page.locator("#connect")).to_be_enabled(timeout=120000)
+        await page.screenshot(path=str(ARTIFACTS / "desktop.png"))
+        await page.get_by_role("button", name="Gespräch starten").click()
+        try:
+            await page.locator(".message.assistant").wait_for(timeout=60000)
+            transcript = await page.locator("#messages").inner_text()
+            audio_stats = await page.evaluate("""async () => {
+                const reports = await peer.getStats();
+                return [...reports.values()].filter(report =>
+                    report.type === 'inbound-rtp' && report.kind === 'audio'
+                ).map(report => ({
+                    packets: report.packetsReceived,
+                    energy: report.totalAudioEnergy,
+                    duration: report.totalSamplesDuration,
+                }));
+            }""")
+            assert any(report.get("energy", 0) > 0 for report in audio_stats), audio_stats
+            assert await page.evaluate("!output.paused && output.currentTime > 0")
+            await page.get_by_role("button", name="Mikrofon pausieren").click()
+            assert await page.locator("#mute").get_attribute("aria-pressed") == "true"
+            await page.get_by_role("button", name="Mikrofon aktivieren").click()
+            await page.get_by_role("button", name="Gespräch beenden").click()
+            assert await page.locator("#mute").is_hidden()
+            await page.get_by_role("button", name="Verlauf leeren").click()
+            assert await page.locator(".message").count() == 0
+            await page.set_viewport_size({"width": 390, "height": 844})
+            await page.screenshot(path=str(ARTIFACTS / "mobile.png"))
+            assert await page.evaluate("document.documentElement.scrollWidth <= innerWidth")
+            assert not errors, errors
+            print(
+                json.dumps(
+                    {"passed": True, "transcript": transcript,
+                     "audio": audio_stats, "js_errors": errors},
+                    ensure_ascii=False,
+                )
+            )
+        finally:
+            await page.screenshot(path=str(ARTIFACTS / "last-test.png"))
+            print("Browser status:", await page.locator("#status").inner_text())
+            print("Browser error:", await page.locator("#error").inner_text())
+            await browser.close()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/tests/test_core.py b/tests/test_core.py
new file mode 100644
index 0000000..655a703
--- /dev/null
+++ b/tests/test_core.py
@@ -0,0 +1,51 @@
+import unittest
+
+from astra.core import Settings, build_request, local_origin_allowed, trim_messages
+
+
+class CoreTests(unittest.TestCase):
+    def test_thinking_is_disabled_in_every_native_request(self):
+        request = build_request(Settings(), [{"role": "user", "content": "Hallo"}])
+        self.assertIs(request["think"], False)
+        self.assertIs(request["stream"], True)
+        self.assertEqual(request["model"], "qwen3.5:latest")
+        self.assertNotIn("think", request["options"])
+        self.assertEqual(request["options"]["num_ctx"], 4096)
+
+    def test_history_stays_bounded_and_preserves_system_and_latest_user(self):
+        messages = [{"role": "system", "content": "Deutsch"}]
+        for i in range(40):
+            messages.extend(
+                [
+                    {"role": "user", "content": f"Frage {i}"},
+                    {"role": "assistant", "content": "Antwort " * 100},
+                ]
+            )
+        messages.append({"role": "user", "content": "Neueste Frage"})
+        trimmed = trim_messages(messages, max_chars=5000)
+        self.assertEqual(trimmed[0], messages[0])
+        self.assertEqual(trimmed[1]["role"], "user")
+        self.assertEqual(trimmed[-1], messages[-1])
+        self.assertLessEqual(sum(len(m["content"]) for m in trimmed), 5000)
+        self.assertEqual(len(messages), 82)
+
+    def test_rejects_foreign_web_origins(self):
+        self.assertTrue(local_origin_allowed("http://localhost:7860"))
+        self.assertTrue(local_origin_allowed("http://127.0.0.1:7860"))
+        self.assertFalse(local_origin_allowed("https://evil.example"))
+        self.assertFalse(local_origin_allowed("http://localhost.evil.example:7860"))
+        self.assertFalse(local_origin_allowed("null"))
+
+    def test_huge_last_message_is_bounded(self):
+        trimmed = trim_messages(
+            [
+                {"role": "system", "content": "Deutsch"},
+                {"role": "user", "content": "x" * 20000},
+            ],
+            max_chars=5000,
+        )
+        self.assertLessEqual(sum(len(m["content"]) for m in trimmed), 5000)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_inference.py b/tests/test_inference.py
new file mode 100644
index 0000000..89fbfea
--- /dev/null
+++ b/tests/test_inference.py
@@ -0,0 +1,55 @@
+import asyncio
+import threading
+from concurrent.futures import ThreadPoolExecutor
+
+import pytest
+
+from astra.inference import drain_stream, on_executor
+
+
+@pytest.mark.asyncio
+async def test_cancel_waits_for_native_inference_before_releasing_model():
+    started = threading.Event()
+    release = threading.Event()
+    finished = threading.Event()
+
+    def inference():
+        started.set()
+        release.wait(timeout=5)
+        finished.set()
+        return "audio"
+
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        task = asyncio.create_task(on_executor(executor, inference))
+        await asyncio.to_thread(started.wait, 3)
+        task.cancel()
+        await asyncio.sleep(0.01)
+        assert not task.done()
+        release.set()
+        with pytest.raises(asyncio.CancelledError):
+            await task
+        assert finished.is_set()
+
+
+@pytest.mark.asyncio
+async def test_native_exception_is_visible():
+    def fail():
+        raise RuntimeError("Model failed")
+
+    with ThreadPoolExecutor(max_workers=1) as executor:
+        with pytest.raises(RuntimeError, match="Model failed"):
+            await on_executor(executor, fail)
+
+
+def test_abandoned_tts_stream_reaches_normal_join_before_reuse():
+    joined = []
+
+    def threaded_stream():
+        yield b"first audio"
+        yield b"remaining audio"
+        joined.append(True)  # Represents Pocket TTS's join after its yield loop.
+
+    stream = threaded_stream()
+    next(stream)
+    drain_stream(stream)
+    assert joined == [True]
diff --git a/tests/test_ollama.py b/tests/test_ollama.py
new file mode 100644
index 0000000..339f136
--- /dev/null
+++ b/tests/test_ollama.py
@@ -0,0 +1,48 @@
+import json
+from unittest.mock import patch
+
+import httpx
+import pytest
+from pipecat.processors.aggregators.llm_context import LLMContext
+
+from astra.core import Settings
+from astra.services import NativeOllamaService
+
+
+@pytest.mark.asyncio
+async def test_native_stream_sends_think_false_and_returns_content():
+    requests = []
+
+    def handler(request):
+        requests.append(json.loads(request.content))
+        return httpx.Response(200, text='{"message":{"content":"Hallo"}}\n{"done":true}\n')
+
+    client = httpx.AsyncClient(transport=httpx.MockTransport(handler))
+    service = NativeOllamaService(Settings(), lambda message: None)
+    with patch("astra.services.httpx.AsyncClient", return_value=client):
+        stream = await service.get_chat_completions(LLMContext([{"role": "user", "content": "Hi"}]))
+        chunks = [chunk async for chunk in stream]
+    assert requests[0]["think"] is False
+    assert chunks[0].choices[0].delta.content == "Hallo"
+    await service._client.close()
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "body",
+    [
+        '{"message":{"thinking":"secret reasoning"}}\n',
+        '{"error":"out of memory"}\n',
+        '{"message":{"content":"incomplete"}}\n',
+    ],
+)
+async def test_bad_ollama_streams_fail_explicitly(body):
+    client = httpx.AsyncClient(
+        transport=httpx.MockTransport(lambda request: httpx.Response(200, text=body))
+    )
+    service = NativeOllamaService(Settings(), lambda message: None)
+    with patch("astra.services.httpx.AsyncClient", return_value=client):
+        stream = await service.get_chat_completions(LLMContext([{"role": "user", "content": "Hi"}]))
+        with pytest.raises(RuntimeError):
+            _ = [chunk async for chunk in stream]
+    await service._client.close()
diff --git a/tests/test_server.py b/tests/test_server.py
new file mode 100644
index 0000000..892bd41
--- /dev/null
+++ b/tests/test_server.py
@@ -0,0 +1,45 @@
+from fastapi.testclient import TestClient
+
+from astra.server import create_app
+
+
+def test_status_and_static_ui_work_before_models_are_ready():
+    with TestClient(create_app(load_models=False), base_url="http://localhost:7860") as client:
+        status = client.get("/api/status")
+        assert status.status_code == 200
+        assert status.json()["thinking"] is False
+        assert status.json()["ready"] is False
+        page = client.get("/")
+        assert "Einfach aussprechen" in page.text
+        assert "frame-ancestors 'none'" in page.headers["content-security-policy"]
+        assert client.get("/static/app.js").status_code == 200
+
+
+def test_other_websites_cannot_open_a_microphone_session():
+    with TestClient(create_app(load_models=False), base_url="http://localhost:7860") as client:
+        for origin in ["https://evil.example", "null", "http://localhost.evil.example:7860"]:
+            response = client.post("/api/offer", headers={"Origin": origin}, json={})
+            assert response.status_code == 403
+        assert client.post("/api/offer", json={}).status_code == 403
+
+
+def test_unready_returns_actionable_status_and_invalid_sdp_is_rejected():
+    with TestClient(create_app(load_models=False), base_url="http://localhost:7860") as client:
+        headers = {"Origin": "http://localhost:7860"}
+        assert client.post("/api/offer", headers=headers, json={}).status_code == 422
+        response = client.post(
+            "/api/offer", headers=headers, json={"sdp": "a" * 20, "type": "offer"}
+        )
+        assert response.status_code == 503
+        assert response.json()["detail"]
+
+
+def test_disconnect_is_idempotent_and_host_is_checked():
+    with TestClient(create_app(load_models=False), base_url="http://localhost:7860") as client:
+        response = client.post(
+            "/api/disconnect",
+            headers={"Origin": "http://localhost:7860"},
+            json={"pc_id": "already-gone"},
+        )
+        assert response.json() == {"ok": True}
+        assert client.get("/api/status", headers={"Host": "evil.example"}).status_code == 403
diff --git a/tests/test_stt.py b/tests/test_stt.py
new file mode 100644
index 0000000..a3105d6
--- /dev/null
+++ b/tests/test_stt.py
@@ -0,0 +1,28 @@
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from pipecat.processors.frame_processor import FrameDirection
+
+from astra.core import Settings
+from astra.inference import Models
+from astra.services import NemotronSTTService
+
+
+@pytest.mark.asyncio
+async def test_recognition_failure_reaches_pipeline_upstream_and_notifies_browser():
+    models = Models(Settings())
+    notices = []
+    service = NemotronSTTService(models, notices.append)
+    service.push_frame = AsyncMock()
+    try:
+        service.queue.put_nowait(("start", b""))
+        with patch("astra.services.Recognizer", side_effect=RuntimeError("Model failed")):
+            await service._worker()
+        assert service.failed
+        assert notices[0]["type"] == "error"
+        assert "Model failed" in notices[0]["message"]
+        frame, direction = service.push_frame.call_args.args
+        assert frame.fatal
+        assert direction == FrameDirection.UPSTREAM
+    finally:
+        await models.close()
diff --git a/web/app.js b/web/app.js
new file mode 100644
index 0000000..461e168
--- /dev/null
+++ b/web/app.js
@@ -0,0 +1,178 @@
+"use strict";
+const $ = (id) => document.getElementById(id);
+const labels = {listening: "Ich höre dir zu.", responding: "Einen Moment …", speaking: "Astra spricht. Du kannst jederzeit unterbrechen."};
+let peer = null, stream = null, channel = null, pcId = null, heartbeat = null;
+let connecting = false, disconnecting = false, muted = false, ready = false;
+const output = new Audio();
+output.autoplay = true;
+
+function state(value) {
+  document.body.dataset.state = value;
+  $("status").textContent = muted ? "Mikrofon pausiert" : (labels[value] || value);
+}
+function showError(message) {
+  $("error").textContent = message;
+  $("error").hidden = false;
+}
+function addMessage(event) {
+  $("messages").querySelector(".empty")?.remove();
+  const article = document.createElement("article");
+  article.className = `message ${event.role === "user" ? "user" : "assistant"}`;
+  const speaker = document.createElement("span");
+  speaker.className = "speaker";
+  speaker.textContent = event.role === "user" ? "Du" : "Astra";
+  const text = document.createElement("p");
+  text.textContent = event.text;
+  article.append(speaker, text);
+  if (event.interrupted) {
+    const note = document.createElement("small");
+    note.textContent = "Unterbrochen";
+    article.append(note);
+  }
+  $("messages").append(article);
+  while ($("messages").children.length > 80) $("messages").firstElementChild.remove();
+  $("messages").scrollTop = $("messages").scrollHeight;
+}
+function receive(event) {
+  let message;
+  try { message = JSON.parse(event.data); } catch { return; }
+  if (message.type === "state") state(message.state);
+  if (message.type === "partial") $("partial").textContent = message.text;
+  if (message.type === "transcript") addMessage(message);
+  if (message.type === "error") showError(message.message);
+  if (message.type === "metric" && message.name === "llm_ms") {
+    $("latency").textContent = `Erstes Antwortwort · ${(message.value / 1000).toFixed(2)} s`;
+  }
+}
+async function request(path, body, timeout = 20000) {
+  const response = await fetch(path, {method: "POST", headers: {"Content-Type": "application/json"},
+    body: JSON.stringify(body), signal: AbortSignal.timeout(timeout)});
+  const data = await response.json();
+  if (!response.ok) throw new Error(typeof data.detail === "string" ? data.detail : "Verbindung fehlgeschlagen.");
+  return data;
+}
+async function waitIce(pc) {
+  if (pc.iceGatheringState === "complete") return;
+  await new Promise((resolve, reject) => {
+    const timeout = setTimeout(() => { cleanup(); reject(new Error("Lokaler Verbindungsaufbau dauert zu lange.")); }, 10000);
+    function cleanup() { clearTimeout(timeout); pc.removeEventListener("icegatheringstatechange", check); }
+    function check() { if (pc.iceGatheringState === "complete") { cleanup(); resolve(); } }
+    pc.addEventListener("icegatheringstatechange", check);
+    check();
+  });
+}
+async function connect() {
+  if (connecting || peer) return;
+  connecting = true;
+  $("error").hidden = true;
+  $("connect").disabled = true;
+  state("Mikrofon wird verbunden …");
+  try {
+    if (!navigator.mediaDevices?.getUserMedia) throw new Error("Bitte diese Seite unter http://localhost:7860 öffnen.");
+    stream = await navigator.mediaDevices.getUserMedia({audio: {
+      echoCancellation: true, noiseSuppression: true, autoGainControl: true, channelCount: 1,
+    }, video: false});
+    const pc = new RTCPeerConnection({iceServers: []});
+    peer = pc;
+    stream.getTracks().forEach(track => pc.addTrack(track, stream));
+    channel = pc.createDataChannel("astra");
+    channel.onmessage = receive;
+    channel.onopen = () => {
+      heartbeat = setInterval(() => { if (channel?.readyState === "open") channel.send("ping"); }, 1000);
+    };
+    pc.ontrack = (event) => {
+      output.srcObject = event.streams[0] || new MediaStream([event.track]);
+      output.play().catch(() => showError("Audioausgabe blockiert. Bitte die Audiowiedergabe im Browser erlauben und neu starten."));
+    };
+    pc.onconnectionstatechange = () => {
+      if (peer !== pc) return;
+      if (pc.connectionState === "connected") state("listening");
+      if (["failed", "disconnected", "closed"].includes(pc.connectionState) && !disconnecting) {
+        showError("Verbindung beendet. Du kannst das Gespräch erneut starten.");
+        void disconnect();
+      }
+    };
+    await pc.setLocalDescription(await pc.createOffer());
+    await waitIce(pc);
+    const answer = await request("/api/offer", {sdp: pc.localDescription.sdp, type: "offer"});
+    pcId = answer.pc_id;
+    await pc.setRemoteDescription({sdp: answer.sdp, type: answer.type});
+    $("connect").textContent = "Gespräch beenden";
+    $("mute").hidden = false;
+    $("clear").textContent = "Neues Gespräch";
+    $("hint").textContent = "Sprich frei. Beim Dazwischenreden hält Astra an.";
+  } catch (error) {
+    await disconnect();
+    const messages = {NotAllowedError: "Mikrofonzugriff nicht erlaubt. Bitte in den Browser-Einstellungen freigeben.",
+      NotFoundError: "Kein Mikrofon gefunden. Bitte ein Mikrofon anschließen.",
+      NotReadableError: "Das Mikrofon ist gerade nicht verfügbar."};
+    showError(messages[error.name] || error.message);
+  } finally {
+    connecting = false;
+    $("connect").disabled = !ready;
+  }
+}
+async function disconnect() {
+  if (disconnecting) return;
+  disconnecting = true;
+  const id = pcId;
+  pcId = null;
+  clearInterval(heartbeat);
+  heartbeat = null;
+  stream?.getTracks().forEach(track => track.stop());
+  stream = null;
+  const oldPeer = peer;
+  peer = null;
+  channel?.close();
+  channel = null;
+  oldPeer?.close();
+  output.pause();
+  output.srcObject = null;
+  muted = false;
+  $("mute").hidden = true;
+  $("mute").setAttribute("aria-pressed", "false");
+  $("mute").textContent = "Mikrofon pausieren";
+  $("connect").textContent = "Gespräch starten";
+  $("clear").textContent = "Verlauf leeren";
+  $("partial").textContent = "";
+  $("hint").textContent = "Mikrofon ist aus. Beim nächsten Start beginnt ein neues Gespräch.";
+  state("Bereit, wenn du es bist.");
+  try { if (id) await request("/api/disconnect", {pc_id: id}, 10000); }
+  catch { showError("Der Server hat das Beenden nicht bestätigt. Das Mikrofon ist aus; bitte kurz warten, bevor du neu startest."); }
+  finally { disconnecting = false; }
+}
+$("connect").addEventListener("click", () => { if (peer) void disconnect(); else void connect(); });
+$("mute").addEventListener("click", () => {
+  muted = !muted;
+  stream?.getAudioTracks().forEach(track => { track.enabled = !muted; });
+  $("mute").setAttribute("aria-pressed", String(muted));
+  $("mute").textContent = muted ? "Mikrofon aktivieren" : "Mikrofon pausieren";
+  state("listening");
+});
+$("clear").addEventListener("click", async () => {
+  const reconnect = Boolean(peer);
+  if (reconnect) await disconnect();
+  $("messages").replaceChildren();
+  if (reconnect) await connect();
+});
+window.addEventListener("pagehide", () => {
+  stream?.getTracks().forEach(track => track.stop());
+  peer?.close();
+});
+async function poll() {
+  try {
+    const response = await fetch("/api/status", {signal: AbortSignal.timeout(5000)});
+    if (!response.ok) throw new Error("Status nicht verfügbar");
+    const status = await response.json();
+    ready = status.ready;
+    if (!peer && !connecting) {
+      $("connect").disabled = !ready;
+      state(ready ? "Bereit, wenn du es bist." : status.stage);
+      if (status.error) showError(status.error);
+    }
+  } catch {
+    ready = false;
+    if (!peer) { $("connect").disabled = true; state("Lokaler Server nicht erreichbar."); }
+  } finally { setTimeout(poll, ready ? 5000 : 1500); }
+}
+void poll();
diff --git a/web/index.html b/web/index.html
new file mode 100644
index 0000000..55b3108
--- /dev/null
+++ b/web/index.html
@@ -0,0 +1,30 @@
+<!doctype html>
+<html lang="de">
+<head>
+  <meta charset="UTF-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="theme-color" content="#111820">
+  <title>Astra · Dein lokaler Sprachraum</title>
+  <link rel="stylesheet" href="/static/style.css">
+  <script src="/static/app.js" defer></script>
+</head>
+<body>
+  <main>
+    <header><a class="brand" href="/" aria-label="Astra Startseite"><span class="star">✳</span> astra</a><span class="local"><i></i> Nur auf deinem Mac</span></header>
+    <section class="conversation" aria-labelledby="title">
+      <div class="intro"><p class="eyebrow">DEIN LOKALER SPRACHRAUM</p><h1 id="title">Einfach aussprechen.</h1><p class="description">Gedanken sortieren. Fragen stellen. Gemeinsam weiterdenken.</p></div>
+      <div class="orb-wrap" aria-hidden="true"><div class="orbit"></div><div class="orb"><div class="wave"><i></i><i></i><i></i><i></i><i></i><i></i><i></i></div></div></div>
+      <div class="live-status"><span id="state-dot"></span><p id="status" role="status" aria-live="polite">Modelle werden vorbereitet …</p></div>
+      <p id="partial" class="partial" aria-live="off"></p>
+      <p id="error" class="error" role="alert" hidden></p>
+      <div class="controls"><button id="connect" class="primary" disabled><span aria-hidden="true">◉</span> Gespräch starten</button><button id="mute" class="secondary" hidden aria-pressed="false">Mikrofon pausieren</button></div>
+      <p class="hint" id="hint">Mikrofon wird erst nach dem Start freigegeben.</p>
+    </section>
+    <section class="transcript" aria-labelledby="transcript-title">
+      <div class="section-top"><h2 id="transcript-title">Unser Gespräch</h2><button id="clear" class="text-button">Verlauf leeren</button></div>
+      <div id="messages" role="log" aria-live="polite" aria-relevant="additions"><p class="empty">Hier erscheinen deine Worte und Astras Antworten.<br>Der Verlauf bleibt nur für dieses Gespräch im Speicher.</p></div>
+    </section>
+    <footer><span>Deutsch <b>·</b> Lokale Spracherkennung & Stimme</span><span id="latency">Bereit für deine Gedanken</span></footer>
+  </main>
+</body>
+</html>
diff --git a/web/style.css b/web/style.css
new file mode 100644
index 0000000..b51fc94
--- /dev/null
+++ b/web/style.css
@@ -0,0 +1,3 @@
+@font-face{font-family:system;src:local("Avenir Next")}
+:root{color-scheme:dark;--bg:#111820;--panel:#19222b;--text:#eef2ef;--muted:#a3b1b8;--accent:#b4e5cc;--border:#303c44}
+*{box-sizing:border-box}body{margin:0;background:radial-gradient(ellipse at 50% 25%,#1c2d34 0%,var(--bg) 58%);color:var(--text);font-family:system,-apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif;min-height:100vh}main{max-width:1060px;margin:auto;padding:36px 48px 25px}header{display:flex;justify-content:space-between;align-items:center}.brand{text-decoration:none;color:var(--text);font-size:30px;letter-spacing:-1.5px;font-weight:600}.star{color:var(--accent);font-size:35px;vertical-align:middle;margin-right:7px}.local{display:flex;align-items:center;gap:9px;font-size:12px;color:#c3d3ce;border:1px solid var(--border);padding:9px 13px;border-radius:30px}.local i{height:6px;width:6px;background:var(--accent);border-radius:50%}.conversation{text-align:center;padding:60px 0 34px}.eyebrow{font-size:10px;font-weight:600;letter-spacing:2.8px;color:var(--accent);margin-bottom:17px}h1{font-size:clamp(32px,5vw,49px);font-weight:500;letter-spacing:-1.8px;margin:0 0 13px}.description{font-size:14px;color:var(--muted);margin:0}.orb-wrap{position:relative;width:210px;height:210px;display:grid;place-items:center;margin:31px auto 15px}.orb{width:145px;height:145px;border-radius:50%;display:grid;place-items:center;background:radial-gradient(circle at 30% 25%,#a3e1cf77,#4e938a22 55%,#172b3544);border:1px solid #a3d3c35c;box-shadow:inset 0 0 38px #92e1ce19,0 0 65px #84c5b215}.orbit{position:absolute;inset:6px;border:1px solid #9acdb01b;border-radius:50%;box-shadow:0 0 0 16px #98cdb104}.wave{height:40px;display:flex;align-items:center;gap:5px}.wave i{display:block;width:4px;height:13px;border-radius:5px;background:#d0f3e4;transition:height .3s}.wave i:nth-child(2),.wave i:nth-child(6){height:22px}.wave i:nth-child(3),.wave i:nth-child(5){height:32px}.wave i:nth-child(4){height:41px}body[data-state="speaking"] .wave i,body[data-state="listening"] .wave i{animation:voice 1.2s ease-in-out infinite alternate}.wave i:nth-child(2n){animation-delay:-.8s!important}.wave i:nth-child(3n){animation-delay:-.4s!important}@keyframes voice{to{height:8px;opacity:.6}}.live-status{display:flex;justify-content:center;align-items:center;gap:8px;height:27px}.live-status p{font-size:14px;margin:0}#state-dot{height:6px;width:6px;border-radius:50%;background:var(--muted)}body[data-state="listening"] #state-dot,body[data-state="speaking"] #state-dot{background:var(--accent)}.partial{color:var(--muted);font-size:13px;min-height:22px;max-width:640px;margin:13px auto 17px;overflow-wrap:anywhere}.controls{display:flex;justify-content:center;gap:12px;flex-wrap:wrap}button{font:inherit;cursor:pointer;border:0;transition:background .2s,transform .2s}button:active{transform:translateY(1px)}button:focus-visible,a:focus-visible{outline:3px solid #d5f3e5;outline-offset:5px}.primary{background:var(--accent);color:#162c25;border-radius:9px;font-size:14px;font-weight:600;padding:15px 24px;min-width:211px}.primary:hover{background:#cef2df}.primary span{margin-right:9px}.primary:disabled{background:#43594f;color:#bfcbc4;cursor:wait}.secondary{border:1px solid #576b68;border-radius:9px;padding:14px 18px;background:transparent;color:var(--text);font-size:13px}.secondary[aria-pressed="true"]{background:#553a38;border-color:#c2887c}.hint{font-size:11px;color:var(--muted);margin:16px 0 0}.error{color:#ffd5cd;background:#4a2c2a;border:1px solid #8e564f;border-radius:9px;padding:12px;max-width:640px;margin:15px auto;font-size:13px;overflow-wrap:anywhere}.transcript{border:1px solid var(--border);border-radius:14px;background:#19222b9c;padding:22px 26px}.section-top{display:flex;justify-content:space-between;align-items:center;border-bottom:1px solid var(--border);padding-bottom:16px}h2{margin:0;font-size:13px;font-weight:500}.text-button{color:var(--muted);background:none;font-size:11px;padding:7px}.text-button:hover{color:var(--text)}#messages{max-height:330px;overflow:auto;scroll-behavior:smooth;padding-top:5px}.empty{font-size:12px;color:var(--muted);line-height:1.9;text-align:center;padding:18px 0 12px}.message{padding:15px 0 7px}.message .speaker{font-size:10px;letter-spacing:1px;font-weight:600;text-transform:uppercase;color:var(--accent)}.message.user .speaker{color:#b1bdcb}.message p{font-size:14px;line-height:1.65;margin:6px 0;white-space:pre-wrap;overflow-wrap:anywhere}.message small{color:var(--muted);font-size:11px}footer{display:flex;justify-content:space-between;gap:15px;margin-top:21px;color:var(--muted);font-size:10px}footer b{padding:0 6px;font-weight:400}[hidden]{display:none!important}@media(max-width:620px){main{padding:23px 20px}.conversation{padding-top:44px}.description{max-width:270px;margin:auto;line-height:1.6}.local{font-size:10px;padding:8px 10px}.transcript{padding:18px}footer{flex-direction:column;gap:8px}.orb-wrap{margin-top:21px}}@media(prefers-reduced-motion:reduce){*,*::before,*::after{animation:none!important;transition:none!important;scroll-behavior:auto!important}}