Content Greenlight Copilot

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Content Greenlight Copilot

Inspiration

As a college student, I spend a lot of time scrolling YouTube trying to find content that's actually useful — for internships, coding skills, home fitness, or learning to cook on my own. Algorithms optimize for watch time, not usefulness, so I'm often misguided toward popular content rather than genuinely helpful content. Creators face the flip side of the same problem: they don't know what to make next without relying on guesswork.

What it does

Content Greenlight Copilot is a multi-agent system with three specialized agents:

• Analytics Agent — connects live to a ClickHouse Cloud database via the official ClickHouse MCP server, answering "what has historically proven to be useful?" using real SQL queries, not guesses.
• Trend Scout Agent — searches the live web to answer "what's relevant right now?", using a custom function tool.
• Coordinator Agent — calls both agents and merges their answers into one recommendation, like a producer combining data with current context before greenlighting content.

Together, they help both viewers find genuinely useful content and creators decide what to make next — backed by data, not algorithmic guesswork.

How we built it

Built using Google's Agent Development Kit (ADK) and Gemini models. Analytics connects to ClickHouse Cloud through the official mcp-clickhouse MCP server. Trend Scout uses a custom Python function tool with the free duckduckgo-search library. Coordinator uses ADK's AgentTool to call both agents and synthesize their outputs into one grounded recommendation.

Challenges we ran into

Several real integration issues came up: an mcp library version mismatch broke tool loading; a custom subprocess environment silently wiped the system PATH, causing the ClickHouse MCP server to fail with zero tools loaded; and persistent 429 rate-limit errors that survived even fresh API keys and fresh Google Cloud projects — eventually traced to the discovery that Gemini's built-in Google Search grounding isn't available on the AI Studio free tier at all. We solved this by replacing it with a free, keyless DuckDuckGo-based custom tool.

What we learned

The difference between a model and an agent, how MCP standardizes tool integration versus custom API code, and how to debug distributed multi-process systems by isolating components (testing the MCP server standalone helped pinpoint the PATH issue).

What's next for Content Greenlight Copilot

Extending it with a Comment Ecosystem Analyzer agent that classifies real YouTube comments (sentiment, toxicity, intent, spam) using the YouTube Data API, to compare behavioral patterns across content categories.

Built With

• agentic-ai
• clickhouse
• clickhouse-cloud
• duckduckgo-search
• gemini
• gemini-api
• google-adk
• google-cloud
• mcp
• model-context-protocol
• multi-agent-systems
• python