Elastic News — AI-Powered Multi-Agent Newsroom

A multi-agent newsroom where five AI agents collaborate to research, write, edit, and publish news articles. Built on the A2A protocol, FastMCP, Tavily web search, and Elasticsearch.

How It Works

You submit a story topic through the React UI. From there, the system runs autonomously:

UI  →  News Chief assigns story
           ↓
       Reporter generates an outline and research questions
           ↓
       Researcher (Tavily web search) ──┐
       Archivist (Elasticsearch archive) ┘ → results sent back to Reporter
           ↓
       Reporter writes article from research + archive context
           ↓
       Editor reviews for grammar, tone, length
           ↓
       Publisher indexes to Elasticsearch, deploys, notifies subscribers
           ↓
       Article viewable in UI with source links

Each agent runs as its own A2A server. They discover each other via agent cards and communicate over JSON-RPC. The MCP server provides shared tools (research, outline generation, article writing, editing, tag generation) that agents call directly.

Agents

Agent	Port	Role
News Chief	8080	Coordinates workflow, routes stories between agents
Reporter	8081	Writes articles, calls Researcher + Archivist in parallel
Editor	8082	Reviews drafts, suggests edits, approves for publication
Researcher	8083	Answers research questions via Tavily + Claude synthesis
Publisher	8084	Indexes to Elasticsearch, saves markdown, triggers CI/CD mock

Supporting services: MCP Server (8095), Event Hub for SSE (8090), Article API (8085), React UI (3001).

External: An Archivist agent on Elastic Cloud (built with Agent Builder) searches the news_archive index for historical context.

Quick Start

Docker (recommended)

cp .env.docker .env
# Edit .env — add ANTHROPIC_API_KEY and TAVILY_API_KEY at minimum
docker-compose up -d
open http://localhost:3001

Local

cp env.example .env
# Edit .env with your keys (see docs/configuration-guide.md)
pip install -r requirements.txt
cd react-ui && npm install && cd ..
make start-ui        # agents + React UI with hot reload

Then open http://localhost:3001, fill in a topic, and hit Assign Story.

Configuration

All config lives in .env. The required keys:

Variable	Required	Purpose
ANTHROPIC_API_KEY	Yes	Claude API for content generation
TAVILY_API_KEY	Yes	Web search for research
ELASTICSEARCH_ENDPOINT	For publishing	Elasticsearch cluster URL
ELASTICSEARCH_API_KEY	For publishing	Elasticsearch auth
ELASTIC_ARCHIVIST_AGENT_URL	For archive search	Archivist A2A endpoint
ELASTIC_ARCHIVIST_API_KEY	For archive search	Archivist auth
ANTHROPIC_MODEL	No	Defaults to claude-sonnet-4-6
LOG_FORMAT	No	text (default) or json for structured logs

Elasticsearch

The Publisher indexes articles to an news_archive index. The mapping supports full-text search, semantic_text fields for vector search, and nested research_sources with real URLs from Tavily.

To create the index:

python scripts/create_elasticsearch_index.py

The Article API (/articles/search) supports three search modes: keyword, semantic, and hybrid (RRF). The Publisher also supports bulk indexing via bulk_publish.

Project Structure

agents/              A2A agents (news_chief, reporter, editor, researcher, publisher)
agents/base_agent.py Shared utilities: A2A client, Anthropic calls, MCP, event publishing
mcp_servers/         FastMCP tool server (Tavily research, Claude writing/editing, mocks)
services/            Event Hub (SSE), Article API (FastAPI)
utils/               Logging (text + JSON), config, MCP client, env loader
react-ui/            React frontend with live workflow visualization
scripts/             Start scripts, ES index creation, Docker entrypoint
tests/               Pytest suite with mock Anthropic, Tavily, and Elasticsearch
docs/                Architecture, configuration, Archivist integration guides

Tech Stack

A2A SDK 0.3.26, FastMCP 3.2.4, Anthropic Claude, Tavily 0.7.23, Elasticsearch 9.x, Starlette, React, Uvicorn, httpx.

Common Commands

# Docker
docker-compose up -d          # start
docker-compose logs -f        # tail logs
docker-compose down           # stop

# Local
make start-ui                 # agents + UI
make stop                     # stop all
make test                     # run tests (no API keys needed)
make logs-color               # colorized log tail

Docs

Detailed guides live in docs/: configuration, architecture, Archivist setup, testing, Docker.

License

MIT