gemini-bright-agent

A research-analyst agent built on Google Cloud Agent Builder (ADK), Gemini 2.5, and the Bright Data MCP server. Submission for the Bright Data AI Agents Web Data Hackathon on lablab.ai (build window 2026-05-25 → 2026-05-30).

What it does

The user gives the agent a research question that needs fresh web data. The agent walks the Bright Data MCP tools — SERP search, Web Unlocker scrape, text extraction, and structured-dataset lookup — and answers with cited URLs, verbatim quotes, and a confidence note.

Tool surface

The agent uses the standard Bright Data MCP tool surface — same as the official @brightdata/mcp npm package — so the stub here is one env-var swap away from a real Bright Data account.

• search_engine(query, engine) — SERP API (Google / Bing / DuckDuckGo)
• scrape_page(url) — Web Unlocker, anti-bot bypass, rendered HTML + text
• extract_text(url, css_selector) — clean text from a previously-scraped page
• web_data_lookup(dataset, key) — structured datasets (LinkedIn companies, Amazon products, etc.)

Architecture

┌──────────────────────┐    ┌───────────────────────┐   ┌────────────────────────────┐
│ Streamlit dashboard  │──▶ │  ADK LlmAgent          │──▶│  Bright Data MCP server     │
│ on Cloud Run         │    │  Gemini 2.5 on Vertex  │   │  (stub for demos,           │
│                      │    │  AI                    │   │   real account via          │
│ "research my Q ..."  │    │                        │   │   BRIGHTDATA_API_TOKEN)     │
└──────────────────────┘    └───────────────────────┘   └────────────────────────────┘

Output contract

The system prompt requires EXACTLY these labeled sections per answer:

ANSWER:      one-two sentences, every number/date/version copied verbatim from tools.
SOURCES:     bulleted URLs the agent actually consulted (not the whole SERP).
KEY QUOTES:  2-4 verbatim quotes pulled from the scraped pages, each tagged with its URL.
CONFIDENCE:  high / medium / low, with a one-sentence reason grounded in source quality.
NEXT STEP:   one concrete follow-up search.

Strict rule: every quantitative claim must come from a tool result. The agent cites byte-for-byte, never paraphrases inside KEY QUOTES.

Try it against a real Bright Data account

export BRIGHTDATA_API_TOKEN="brd_..."
pip install -e .
streamlit run app/dashboard.py

Untick "Use stub Bright Data MCP" in the sidebar. The agent now spawns the official @brightdata/mcp server via npx.

Tests

pip install -e ".[dev]"
pytest -q

The suite pins the research-chain contract: the top SERP result for the canned query points to a URL scrape_page recognises, the verbatim "22% latency drop" claim appears in both the SERP snippet and the scraped page, and structured-dataset lookups return the expected employee count for Anthropic.

License

Apache 2.0. Standalone repo created during the Bright Data Web Data Hackathon contest period (2026-05-25 → 2026-05-30).