🔍 DeepLook

Free Bloomberg Terminal for AI Agents — open-source MCP server that researches any company in ~10 seconds.

LLMs hallucinate financial data. Other finance MCP servers return raw data from a single source — you still do the research yourself. DeepLook runs the full workflow: 10 sources in parallel, cross-referenced, with a structured bull/bear verdict. One call, ~10 seconds, no API keys needed.

---

⚡ Connect in 30 Seconds

1. Claude.ai → Settings → Connectors → Add MCP Server
2. Paste: https://mcp.deeplook.dev/mcp
3. Try: "Use DeepLook to research NVIDIA"

Works with Claude Desktop, Cursor, Windsurf, or any MCP-compatible client.

---

What You Get

NVIDIA Corporation — $181.93 | EXPANDING / ACCELERATING
Key Signals:

🟢 Jensen Huang projects $1T AI chip revenue by 2027
🟢 Vera Rubin platform with 7 new chips in production
🔴 Earnings surprise: -55.03%

Verdict: Mega-cap AI leader with 73% revenue growth, $1T opportunity

🟢 Revenue +73.2% YoY, earnings +95.6%, $58.1B FCF
🔴 RSI 37.2 oversold, $4.42T valuation limits upside
⏳ Wait for: Q1 FY2027 earnings on 2026-05-20

Embedded structured JSON with precise metrics, peer comparison, technicals → AI clients auto-render as interactive dashboards

---

Features

• 10+ data sources in parallel (yfinance, news, CoinGecko, DeFiLlama, SEC EDGAR, Wikipedia, YouTube, etc.)
• Works for public stocks, crypto, and private companies
• Dual output: human-readable summary + structured JSON for AI agents
• Bull/bear verdict with catalyst timeline
• Peer comparison with financial metrics
• ~10 second research time
• Two tools: deeplook_research (full report) and deeplook_lookup (quick snapshot)

Supported Entity Types

Public Equity · Crypto/DeFi · Private Companies · Exchanges · VCs · Foundations

---

Self-Host

1. Clone and install:

git clone https://github.com/OSOJDJD/deeplook.git
cd deeplook
python3 -m venv venv && source venv/bin/activate
pip install -e .
cp .env.example .env   # add at least one LLM key

2. Run as HTTP MCP server:

python -m deeplook.mcp_server --http --port 8819

3. Or add to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "deeplook": {
      "command": "/full/path/to/deeplook/venv/bin/python",
      "args": ["-m", "deeplook.mcp_server"],
      "cwd": "/full/path/to/deeplook",
      "env": { "ANTHROPIC_API_KEY": "sk-ant-..." }
    }
  }
}

CLI (no MCP):

python -m deeplook "NVIDIA"
python -m deeplook "Aave"
python -m deeplook "Anthropic"

---

API Keys

Pick at least one LLM provider:

Variable	Provider
ANTHROPIC_API_KEY	Claude — Haiku + Sonnet (recommended)
OPENAI_API_KEY	GPT-4o-mini
GEMINI_API_KEY	Gemini 2.0 Flash Lite
DEEPSEEK_API_KEY	DeepSeek Chat

Optional (for deeper research):

Variable	Description
TAVILY_API_KEY	Search fallback when DDG is rate-limited
COINGECKO_API_KEY	CoinGecko Pro for crypto data
ROOTDATA_SKILL_KEY	RootData for crypto project data

Cost per report: ~$0.02–0.05 (Anthropic) · ~$0.01–0.03 (OpenAI) · ~$0.01–0.02 (Gemini) · ~$0.005–0.01 (DeepSeek)

---

Data Sources

Source	Used For
yFinance	Price, financials, analyst targets, technicals
DuckDuckGo News	Recent signals, headlines
Wikipedia	Company background
YouTube	Earnings calls, CEO interviews
CoinGecko	Token price, market cap, volume
RootData	Crypto funding, team data
DefiLlama	TVL, chain metrics
SEC EDGAR	10-K, 10-Q, 8-K filings
Finnhub	Earnings, news, sentiment
Website	Investor relations, product pages

---

How It Works

Company Name
    ↓
Entity Type Router  (public equity / crypto / private / exchange / VC / foundation)
    ↓
10 Parallel Fetchers  (DDG News, yFinance, CoinGecko, SEC EDGAR, ...)
    ↓
3-Call LLM Pipeline:  Extract (Haiku)  →  Judge (Sonnet)  →  Act (Sonnet)
    ↓
Structured Report + Embedded JSON

---

Eval

Tested across 58 companies (US mega-cap, growth stocks, crypto, pre-IPO, international, edge cases):

Metric	Score
Overall	3.78 / 5.0
Risk detection	4.36 / 5.0
Signal quality	3.94 / 5.0
Actionability	3.38 / 5.0

Eval framework ships in /eval — run it yourself, contribute ground truth data.

---

License

MIT — use it however you want.

Built by @OSOJDJD · Open an issue if something breaks or a report looks wrong.