@@ -67,32 +67,33 @@ A sample for building real-time voice agents that maintain two parallel tracks s
 | [main.py](src/duplex-live-agent/main.py) | The whole agent — `@app.ws_handler` opens a Voice Live session and runs two pumps. |
 | [azure.yaml](azure.yaml) | Hosted-agent runtime config (`invocations_ws`, 1 CPU / 1 Gi). |
 | [azure.yaml](azure.yaml) | `azd ai agent init` manifest. |
-| [Dockerfile](src/duplex-live-agent/Dockerfile) | python:3.12-slim, installs dependencies from `requirements.txt`. |
-| [requirements.txt](src/duplex-live-agent/requirements.txt) | Pins `azure-ai-agentserver-invocations==1.1.0` + `azure-ai-voicelive[aiohttp]` + `azure-identity` + `python-dotenv`. |
+| [Dockerfile](src/duplex-live-agent/Dockerfile) | Builds a frozen uv environment on `python:3.12-slim`. |
+| [pyproject.toml](src/duplex-live-agent/pyproject.toml) | Runtime dependencies and the local-only `websockets` development group. |
+| [uv.lock](src/duplex-live-agent/uv.lock) | Fully resolved runtime and development dependency graph. |
 | [.env.example](src/duplex-live-agent/.env.example) | Required Voice Live env vars. |
 | [e2e_local.py](src/duplex-live-agent/e2e_local.py) | Headless E2E that streams a 1 kHz tone in and asserts audio + events come back. |
 
 ## Prerequisites
 
-1. Python 3.10 or later.
-2. Azure CLI logged in (`az login`).
-3. An Azure AI Services / Voice Live resource and access to a realtime model (e.g. `gpt-realtime`).
-4. The `Foundry User` role on that resource for your user (or the hosted agent's managed identity, post-deploy).
+1. Python 3.12 or later.
+2. [uv](https://docs.astral.sh/uv/getting-started/installation/) 0.11.7 or later.
+3. Azure CLI logged in (`az login`).
+4. An Azure AI Services / Voice Live resource and access to a realtime model (e.g. `gpt-realtime`).
+5. The `Foundry User` role on that resource for your user (or the hosted agent's managed identity, post-deploy).
 
 ## Running locally
 
 ```bash
-cd duplex-live-agent
-
-python -m venv .venv
-source .venv/bin/activate
-pip install -r requirements.txt websockets
+cd samples/python/hosted-agents/bring-your-own/invocations_ws/duplex-live-agent
+python -m pip install uv
+cd src/duplex-live-agent
+uv sync --frozen
 
 cp .env.example .env
 # Fill in AZURE_VOICELIVE_ENDPOINT and AZURE_VOICELIVE_MODEL.
 export $(grep -v '^#' .env | xargs)
 
-python main.py
+uv run --no-sync python main.py
 # → Uvicorn running on http://0.0.0.0:8088
 ```
 
@@ -101,9 +102,8 @@ python main.py
 In a second terminal:
 
 ```bash
-cd duplex-live-agent
-source .venv/bin/activate
-python e2e_local.py
+cd samples/python/hosted-agents/bring-your-own/invocations_ws/duplex-live-agent/src/duplex-live-agent
+uv run --no-sync python e2e_local.py
 # [e2e] session_started:   True
 # [e2e] audio_bytes recvd: 230400
 # [e2e] result:            PASS
@@ -118,11 +118,11 @@ spoken utterance — server-VAD is bypassed entirely.
 
 ### Browser test (standalone client)
 
-The sample ships a tiny single-file browser client in [`chat_client/`](chat_client/index.html).
+The sample ships a tiny single-file browser client in [`chat_client/`](src/duplex-live-agent/chat_client/index.html).
 With the agent running on `:8088`, just serve the folder:
 
 ```bash
-cd chat_client
+cd samples/python/hosted-agents/bring-your-own/invocations_ws/duplex-live-agent/src/duplex-live-agent/chat_client
 python -m http.server 8080
 ```
 
@@ -183,7 +183,8 @@ this on a WebSocket, so the standalone `chat_client/index.html` is for
 local dev only):
 
 ```bash
-python e2e_local.py \
+cd src/duplex-live-agent
+uv run --no-sync python e2e_local.py \
   --foundry "https://<account>.services.ai.azure.com/api/projects/<project>" \
   --agent duplex-live-agent
 ```
 
@@ -1,12 +1,15 @@
+FROM ghcr.io/astral-sh/uv:0.11.7 AS uv
 FROM python:3.12-slim
 
+COPY --from=uv /uv /uvx /bin/
+
 WORKDIR /app
 
 ENV PYTHONDONTWRITEBYTECODE=1 \
     PYTHONUNBUFFERED=1
 
-COPY requirements.txt ./
-RUN pip install --no-cache-dir -r requirements.txt
+COPY pyproject.toml uv.lock uv.toml ./
+RUN uv sync --frozen --no-dev --no-install-project
 
 COPY . ./
 
@@ -22,7 +25,7 @@ EXPOSE 8088
 # container start.
 # PYTHONDONTWRITEBYTECODE is cleared for these commands only: it blocks
 # bytecode *writes*, not *reads*, so images that set it still benefit.
-RUN PYTHONDONTWRITEBYTECODE= python -m compileall -q $(python -c "import sysconfig as s; print(s.get_paths()['stdlib'], s.get_paths()['purelib'])") || true; \
-    PYTHONDONTWRITEBYTECODE= python -m compileall -q .
+RUN PYTHONDONTWRITEBYTECODE= .venv/bin/python -m compileall -q $(.venv/bin/python -c "import sysconfig as s; print(s.get_paths()['stdlib'], s.get_paths()['purelib'])") || true; \
+    PYTHONDONTWRITEBYTECODE= .venv/bin/python -m compileall -q -x '[.]venv' .
 
-CMD ["python", "main.py"]
+CMD ["/app/.venv/bin/python", "main.py"]
@@ -0,0 +1,21 @@
+[project]
+name = "duplex-live-agent"
+version = "0.1.0"
+description = "A Microsoft Foundry hosted agent using the Invocations WebSocket protocol."
+requires-python = ">=3.12"
+dependencies = [
+    "agent-framework-core==1.10.0",
+    "agent-framework-foundry==1.10.0",
+    "agent-framework-orchestrations==1.0.0",
+    "azure-ai-agentserver-invocations==1.1.0",
+    "azure-ai-voicelive[aiohttp]>=1.0.0b1",
+    "azure-identity>=1.20.0",
+    "debugpy",
+    "github-copilot-sdk==0.3.0",
+    "python-dotenv>=1.0.0",
+]
+
+[dependency-groups]
+dev = [
+    "websockets>=14",
+]