Trendwatch

(Trendwatch — July 2025)

---

1 · Live Server Endpoints

Purpose	URL
Web UI (minimal search page)	https://trendwatch-1095464065298.us-east1.run.app
MCP SSE Endpoint	https://trendwatch-1095464065298.us-east1.run.app/sse

The server is deployed on Google Cloud Run from the included Dockerfile. If you’re self‑hosting, replace the URLs above with your own domain.

---

2 · Prerequisites

• Deployed Trendwatch FastMCP server (see Dockerfile + app/server.py).
• Publicly‑reachable HTTPS URL (Cloud Run or any reverse proxy).
• (Optional) Bearer token set via API_TOKEN env var for authenticated access.

---

3 · Connecting From Popular AI Chat Providers

3.1 · OpenAI ChatGPT (Deep Research Mode)

1. Open ⋯ Settings → Beta features → Connectors and click ➕ Add connector.

2. Paste the following JSON, then save:

   {
     "server_label": "Trendwatch Shorts",
     "server_url": "https://trendwatch-1095464065298.us-east1.run.app",
     "allowed_tools": ["search", "fetch"],
     "require_approval": "never" // or "per-search" if you want manual approval
   }
   

3. Start a new chat, select Deep Research model (“o3‑mini‑deep‑research” or newer) and pick Trendwatch Shorts in the Tools dropdown. 4. Test with: “Find shorts about AI makeup tutorials.” 5. Inspect the trace link to confirm tool calls succeed.

Tip: Add "require_approval": "auto" to force silent calls while still recording traces.

Programmatic Usage

from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
    model="o3-mini-deep-research",
    tools=[{
        "type": "mcp",
        "server_label": "trendwatch",
        "server_url": "https://trendwatch-1095464065298.us-east1.run.app",
        "allowed_tools": ["search", "fetch"],
        "require_approval": "never",
    }],
    input=[{"role": "user", "content": [{"type": "input_text", "text": "Top skateboarding shorts uploaded this week"}]}],
)
print(resp.choices[0].message.content[0].text)

---

3.2 · Anthropic Claude 3 (Messages API)

Claude 3 Opus and later support MCP via the anthropic-beta: mcp-client-2025-04-04 header.

curl https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-beta: mcp-client-2025-04-04" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-3-opus-2025-07-15",
    "system": "You are a research assistant.",
    "tools": [{
      "type": "mcp",
      "server_url": "https://trendwatch-1095464065298.us-east1.run.app",
      "allowed_tools": ["search", "fetch"]
    }],
    "messages": [{"role":"user","content":"Give me viral cat shorts from the past 48 hours"}]
  }'

Anthropic automatically infers when to call search/fetch based on your prompt.

---

3.3 · ElevenLabs 11.ai (Voice & Chat Agents)

1. Open Dashboard → Agents → Tools (Beta).

2. Click Add MCP Server, then enter:

   • Label: Trendwatch
   • Server URL: https://trendwatch-1095464065298.us-east1.run.app
   • Transport: SSE
   • Allowed tools: search, fetch

3. Attach the toolset to your voice agent and deploy.

4. Speak: “Tell me about trending Pokémon shorts.” 11.ai converts the request to text, triggers the MCP calls, and returns spoken results.

ElevenLabs limits MCP calls to 30 req/min per agent. Use Cloud Run’s concurrency settings to scale accordingly.

---

4 · Testing & Troubleshooting

Symptom	Likely Cause	Fix
401 Unauthorized	Missing/incorrect API_TOKEN header	Supply Authorization: Bearer <token> or unset API_TOKEN.
ToolNotFound: search	Connector allowed‑tools mismatch	Ensure allowed_tools includes both search and fetch.
Network timeout	Cloud Run cold start or egress block	Warm the service; check VPC egress rules.
Empty result list	Query too narrow	Confirm index contains matching titles/descriptions.

Check the raw endpoint manually:

curl "https://trendwatch-1095464065298.us-east1.run.app/search?query=ai"

---

5 · Security Checklist

• Input sanitization – FastMCP escapes HTML; still validate prompt‑generated IDs.
• Rate limiting – Apply Cloud Armor or FastMCP’s built‑in limits.
• Token scopes – Issue least‑privilege API keys (read‑only access to this corpus).
• Prompt‑injection review – Monitor traces for attacker‑controlled instructions.
• HTTPS everywhere – Cloud Run terminates TLS; do not expose plain HTTP.

---

6 · Local Development Recap

# 1 – Install deps
pip install -r requirements.txt

# 2 – Run the pipeline & create a parquet
python -m app.pipeline --query "AI" --days 5 --out trend.csv

# 3 – Serve locally (SSE on :8000)
DATA_PATH=trend.csv python -m app.server

If DATA_PATH is unset or the file is missing, the server will create a small example dataset automatically.

Then register http://localhost:8000 as an MCP server in your test provider.

---

7 · References & Further Reading

• FastMCP Docs: https://github.com/jlowin/fastmcp
• OpenAI Cookbook – MCP examples: https://cookbook.openai.com/examples/partners/mcp_powered_voice_agents/mcp_powered_agents_cookbook
• Building MCP Servers for Deep Research (PDF) – see repo docs.

---