Daily News Briefing

Every morning this app pulls the day's news from several sources, has an LLM curate it and rewrite it into a spoken briefing, and then phones me. A conversational voice agent reads the briefing and answers follow-up questions about any story. It runs on Google Cloud, so it doesn't depend on my laptop.

The briefing has four segments, all configured in feeds.yaml:

How it works

flowchart TD
    Sched[Cloud Scheduler · 11:30 AM ET] --> Job[Cloud Run Job → python -m agent.sync --call]
    Job --> Loop{for each feed in feeds.yaml}
    Loop --> Fetch[sources/ · fetch + merge stories]
    Fetch --> Cur[curation/ · pick + rank]
    Cur --> Enr[enrich thin stories with article text]
    Enr --> Rew[rewrite · Gemini 2.5 Pro]
    Rew --> Loop
    Loop -->|all segments| Doc[agent/context.py · narrative + source stories]
    Doc --> KB[ElevenLabs agent · knowledge base + anchor persona]
    KB --> Call[outbound call via Twilio number → my phone]

1. Fetch. Each feed merges stories from its sources into a common Story shape. If one source fails, it's skipped and the rest still run.
2. Curate. A curator ranks and trims the candidates. agentic is a Google ADK agent that can fetch articles and check past digests; deterministic ranks by recency and keywords. If the agentic curator errors, it falls back to deterministic.
3. Rewrite. Gemini turns each segment into speech-ready prose at temperature 0.2. The prompts in prompts/ stop it from inventing facts.
4. Sync + call. The composed narrative and the grounding stories are uploaded as the agent's knowledge doc for the day. ElevenLabs then rings USER_PHONE_NUMBER.

The original MP3-by-email path (main.py) still works: it rewrites the briefing, synthesizes it with Cloud TTS (Chirp3-HD), and emails the clip. It's no longer what production runs, but it's useful locally.

For the full picture (diagrams, patterns, how to extend), see ARCHITECTURE.md.

Quick start (local)

python -m venv .venv && .venv/bin/pip install -r requirements.txt
cp .env.example .env                      # fill in Gmail, GCP project, ElevenLabs keys
gcloud auth application-default login     # Vertex AI + TTS credentials

Then run whichever entry point you need:

# Phone-agent path (what production runs)
.venv/bin/python -m agent.sync --dry-run          # build + print today's knowledge doc, touch nothing
.venv/bin/python -m agent.sync                    # refresh the ElevenLabs agent's knowledge
.venv/bin/python -m agent.sync --call             # refresh, then ring USER_PHONE_NUMBER
.venv/bin/python -m agent.sync --call-only        # ring without rebuilding (telephony test)

# MP3-by-email path
DRY_RUN=1 .venv/bin/python main.py                # print the script, no audio/email
.venv/bin/python main.py                          # full: TTS + email

# Tests
.venv/bin/python -m pytest tests -q

GMAIL_ADDRESS / GMAIL_APP_PASSWORD are only needed for the MP3-by-email path (main.py), which validates them when it sends; the phone-agent and MCP paths build their config without them.

MCP server

The news engine (fetch → curate → rewrite) is also exposed as an MCP server, so any MCP client (Claude Desktop, Cursor, …) can pull AI/news content. It runs locally and uses your own credentials — nothing is sent to a hosted service. (The voice/phone-call feature is deliberately not exposed over MCP.)

.venv/bin/python mcp_server.py           # speaks MCP over stdio

Wire it into Claude Desktop (claude_desktop_config.json) — keys stay on your machine:

{
  "mcpServers": {
    "daily-news": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["/absolute/path/to/Daily_News/mcp_server.py"],
      "env": { "GOOGLE_CLOUD_PROJECT": "your-gcp-project" }  // only needed for get_briefing
    }
  }
}

list_topics and get_stories need no credentials or extra packages. get_briefing additionally requires the briefing extra (the Gemini/Vertex stack) plus gcloud auth application-default login:

pip install "daily-news-mcp[briefing]"    # adds google-genai + google-adk for get_briefing

Without the extra, get_briefing returns a clear "install the briefing extra" message; the keyless tools still work. Once published to PyPI, uvx daily-news-mcp runs the server without a local checkout (uvx "daily-news-mcp[briefing]" to include get_briefing).

Deploy

export USER_PHONE_NUMBER='+1...'   # who the agent calls (kept out of the repo)
bash deploy.sh

deploy.sh enables the APIs, builds the image with Cloud Build, stores the Gmail and ElevenLabs secrets in Secret Manager, grants the vertex-ai-runner@ service account its roles (including cloudtrace.agent), deploys the Cloud Run Job with its entrypoint overridden to python -m agent.sync --call, and creates the Scheduler trigger (default 30 11 * * *, America/New_York).

You also need an ElevenLabs Conversational AI agent, and a Twilio number imported into ElevenLabs. ElevenLabs places the call with that number's credentials, so the app never calls Twilio directly.

gcloud run jobs execute daily-news --region us-central1            # run now (refresh + call)
gcloud run jobs executions list --job daily-news --region us-central1

Configuration

Feeds, sources, word budgets, and story caps live in feeds.yaml. To add a topic, add a feed there plus a prompt in prompts/; no code changes are needed. Everything else is an environment variable (see .env.example). The main ones:

Evals

evals/ grades segments against a frozen golden set: the day's captured sources, so a score change reflects a code change rather than different news. There are two kinds of grading.

• Reference-free (no answer key needed): code checks for length, structure, and TTS safety, plus LLM judges for faithfulness, coverage, and relevance, and ADK's hallucinations_v1.
• Reference-based: graded against a hand-written expectations.md answer key, both as a whole and rule by rule as ADK rubrics.

python -m evals.capture --name YYYY-MM-DD                          # freeze today's sources
python -m evals.run --golden YYYY-MM-DD --curator agentic           # own judges
python -m evals.run --golden YYYY-MM-DD --curator agentic --adk-metrics   # + Google's ADK judges
python -m evals.compare --golden YYYY-MM-DD                         # deterministic vs agentic

Observability

Each run is a single trace in Cloud Trace. ADK emits GenAI spans for the curator's model and tool calls, raw google-genai calls are auto-instrumented, and observability.py adds spans for the pipeline's own stages (gather, curate, enrich, rewrite, sync, call), with token counts on every model call. No LangSmith or LangChain is involved.

To watch the curator agent turn by turn: adk web adk_apps.

Project layout

STATE.md is my running dev log. docs/field-notes.html holds longer notes.