Briefly: A Multi-Protocol Nano Agent powered by Gemma 4.

The Edge-native intelligence layer for creative agencies.

Briefly monitors the AI industry in real-time and makes that intelligence available to humans via chat, to AI assistants via MCP, and to other agents via A2A.

---

What is a Nano Agent?

A nano agent is an AI agent with a deliberately narrow scope and zero infrastructure overhead. Unlike monolithic AI platforms, a nano agent:

• Does one job — Briefly surfaces AI industry intelligence for creative agencies. It doesn't try to be a general assistant, a code editor, or a research tool.
• Runs at the edge — No servers, no databases to manage. Briefly lives entirely on Cloudflare Workers and Durable Objects, globally distributed with sub-millisecond cold starts.
• Is composable — Briefly speaks MCP (Model Context Protocol) and A2A (Agent-to-Agent) natively. Any other AI agent or LLM tool can call it as a sub-agent, delegating the "what's new in AI" question to the one agent built for it.
• Stays lean — The entire backend is ~600 lines of TypeScript. No heavy frameworks, no ORMs. Complexity is the enemy of reliability.

The nano agent philosophy: build the smallest thing that's genuinely useful, make it fast, and make it interoperable.

---

What Briefly Does

Briefly monitors five live intelligence sources on a rolling 6-hour schedule:

Source	Content
The Rundown AI	Daily AI newsletter digest
The Neuron Daily	Startup and product AI news
Futurepedia	Curated AI tools directory
Product Hunt	Newly launched AI products
Arena.ai Leaderboard	Live model performance rankings

That intelligence is surfaced through three interfaces:

1. Chat UI — Ask questions like "what's the best AI video tool for a campaign brief?" and get sourced, grounded answers from Gemma 4.
2. MCP endpoint — Claude Desktop, Cursor, or any MCP-compatible client can add Briefly as a tool server and call its 4 tools directly.
3. A2A endpoint — Other agents can delegate tasks to Briefly via the Agent-to-Agent protocol, polling for results or streaming via SSE.

---

Architecture

┌─────────────────────────────────────────────────────────┐
│                   Cloudflare Worker                      │
│                                                         │
│   Hono Router                                           │
│   ├── /            → React SPA (Workers Assets)         │
│   ├── /mcp         → MCP Streamable HTTP (stateless)    │
│   ├── /a2a/*       → A2A task protocol                  │
│   ├── /api/*       → REST feed / search API             │
│   └── /agents/*    → Agent WebSocket routing            │
│                                                         │
│   Durable Objects                                       │
│   ├── BrieflyAgent   → Conversational AI + briefs       │
│   └── ScraperAgent   → Source scraping + SQLite store   │
│                                                         │
│   Cloudflare Workers AI  (Gemma 4)                      │
│   Cloudflare KV          (A2A task state, 24h TTL)      │
└─────────────────────────────────────────────────────────┘

briefly/
├── src/
│   ├── server.ts            # Hono router — all routes
│   ├── types.ts             # Shared TypeScript interfaces
│   ├── agents/
│   │   ├── BrieflyAgent.ts  # Conversational agent with Gemma 4
│   │   └── ScraperAgent.ts  # Background scraper (6h schedule)
│   ├── mcp/
│   │   └── server.ts        # Stateless MCP server factory
│   ├── a2a/
│   │   ├── agentCard.ts     # Agent Card (/.well-known/agent-card.json)
│   │   ├── taskManager.ts   # Task lifecycle backed by KV
│   │   └── handler.ts       # A2A HTTP routes
│   ├── scrapers/            # One scraper per source
│   └── utils/               # agentFetch helper, HTML parsing
└── src/components/          # React UI (dark glassmorphism)

Key Design Decisions

Stateless MCP — Instead of a Durable Object per MCP session, Briefly creates a fresh McpServer per request. This avoids DO hibernation edge cases and correctly exposes all tools on the first tools/list call without any warmup.

ScraperAgent as the single source of truth — All four MCP tools, all A2A skills, and the chat agent read from the same ScraperAgent Durable Object. One scrape cycle populates everything.

Gemma 4 for inference — @cf/google/gemma-4-26b-a4b-it runs natively on Cloudflare's GPU fleet. No API keys, no per-token billing beyond Workers AI usage.

---

MCP Integration

Connect any MCP-compatible client to:

https://briefly.info-693.workers.dev/mcp

Available tools:

Tool	Description
search_ai_tools	Find AI tools matching a project description or use case
get_industry_news	Latest AI headlines from The Rundown AI and The Neuron Daily
get_model_leaderboard	Current Arena.ai model performance rankings
generate_brief	Full innovation brief as structured JSON (powered by Gemma 4)

Example: Add to Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "briefly": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/client-cli",
        "https://briefly.info-693.workers.dev/mcp"
      ]
    }
  }
}

---

A2A Integration

Briefly implements the Agent-to-Agent (A2A) protocol.

Agent card (auto-discovery):

GET https://briefly.info-693.workers.dev/.well-known/agent-card.json

Task endpoints:

POST   /a2a/tasks              Create task (202 Accepted or SSE stream)
GET    /a2a/tasks/:id          Poll task status
GET    /a2a/tasks/:id/stream   SSE stream for real-time updates
DELETE /a2a/tasks/:id          Cancel task

Skills:

Skill	Input
search_tools	{ query?, category?, limit? }
industry_summary	{ source?, limit? }
model_leaderboard	{ limit? }
generate_brief	{ topic }

Example:

curl -X POST https://briefly.info-693.workers.dev/a2a/tasks \
  -H "Content-Type: application/json" \
  -d '{"message": {"skill": "model_leaderboard", "input": {"limit": 5}}}'

---

Setup & Deploy

Prerequisites

• Node.js 18+
• Wrangler CLI authenticated with your Cloudflare account

1. Install dependencies

npm install

2. Create the KV namespace for A2A tasks

wrangler kv namespace create A2A_TASKS

Copy the output id and replace REPLACE_WITH_KV_NAMESPACE_ID in wrangler.toml.

3. Local development

npm run dev

Open http://localhost:5173.

Note: On first run, ScraperAgent will begin fetching from all 5 sources. The feed may be empty for the first minute.

4. Deploy to Cloudflare

npm run build
cd dist/briefly && npx wrangler deploy --config wrangler.json

The GitHub Actions workflow (.github/workflows/deploy.yml) handles KV namespace creation and deployment automatically on push to main.

---

Bindings

Binding	Type	Purpose
AI	Workers AI	Gemma 4 inference
BRIEFLY_AGENT	Durable Object	Conversational agent
SCRAPER_AGENT	Durable Object	Background scraper + SQLite
A2A_TASKS	KV Namespace	A2A task storage (24h TTL)
ASSETS	Static Assets	React SPA

---

Security

• No API keys in code — all bindings use Cloudflare's secure binding system
• User inputs sanitized and length-capped (500 chars max for queries)
• Prompt injection protection — user input is separated from system prompt at all times
• A2A webhooks — only HTTPS URLs accepted; Bearer tokens supported
• MCP tool inputs validated with Zod schemas before execution
• CORS open only on interoperability endpoints (/mcp, /a2a/*, /.well-known/*)

---

Stack

• Runtime: Cloudflare Workers + Durable Objects (SQLite)
• Routing: Hono
• UI: React + Vite
• AI: Cloudflare Workers AI — Gemma 4 (@cf/google/gemma-4-26b-a4b-it)
• Agent SDK: agents v0.9.0 (Cloudflare)
• MCP: @modelcontextprotocol/sdk v1.29.0
• A2A: Custom implementation following the Agent-to-Agent spec
• Build: Vite + @cloudflare/vite-plugin

---

MIT License · Built by Eduardo Arana