📈 AI Trading Intelligence Framework — MCP Server

The most complete AI-powered trading toolkit for Claude and MCP clients. Backtesting + Live Sentiment + Yahoo Finance + 30+ Technical Analysis Tools — all in one MCP server.

⭐ If this tool improves your workflow, please star the repo and consider sponsoring — it keeps the project alive and growing!

🚀 Don't want to fight uv / pandas / Python on Windows?
pro.cryptosieve.com — same 30+ tools,
one connector URL into Claude.ai, ChatGPT, Copilot, Cursor. $9/month with 7-day free trial.
Self-hosting is free and always will be; the hosted version is just for
folks who'd rather skip the ops dance.

🎥 Framework Demo

https://github-production-user-asset-6210df.s3.amazonaws.com/67838093/478689497-4a605d98-43e8-49a6-8d3a-559315f6c01d.mp4

🆕 What's New

Stability & Strategy Expansion (May 2026)

• 9 backtest strategies (up from 6) — added rsi_pullback, keltner_breakout, and triple_ema, covering trend-pullback, ATR-normalized breakout, and SMA200-filtered EMA cross edges. compare_strategies now ranks the full 9.
• Resilience layer — automatic retry + 60-second TTL cache on the TradingView screener provider, eliminating transient "Expecting value" errors on combined_analysis and multi_timeframe_analysis. (PR #32 — merged)
• Financial news service rebuild — replaces deprecated Reuters RSS endpoints with Yahoo Finance, MarketWatch, and CNBC. Fixes the long-standing count: 0 bug on financial_news. (PR #33 — merged)
• TA throttle — caps concurrent tradingview_ta calls (default 4) + min 0.8s spacing between starts. Prevents parallel bursts of combined_analysis / multi_timeframe_analysis from hitting TradingView's empty-body rate-limit cliff. Tunable via env vars. (PR #34 — merged)
• Walk-forward backtesting (walk_forward_backtest_strategy) — train/test split with overfitting verdict (ROBUST / MODERATE / WEAK / OVERFITTED).
• Hourly (1h) timeframe support across backtest_strategy, compare_strategies, and walk_forward_backtest_strategy.
• Full trade log + equity curve outputs (include_trade_log=True, include_equity_curve=True).

🏗️ Architecture

✨ Why tradingview-mcp?

| Feature | tradingview-mcp | Traditional Setups | Bloomberg Terminal | |---------|-------------------|--------------------|--------------------| | Setup Time | 5 minutes | Hours (Docker, Conda...) | Weeks (Contracts) | | Cost | Free & Open Source | Variable | $30k+/year | | Backtesting | ✅ 9 strategies + Walk-forward + Sharpe | ❌ Manual scripting | ✅ Proprietary | | Live Sentiment | ✅ Reddit + RSS news | ❌ Separate setup | ✅ Terminal | | Market Data | ✅ Live / Real-Time | Historical / Delayed | Live | | API Keys | None required | Multiple (OpenAI, etc.) | N/A |

🚀 Quick Start (5 Minutes)

Install via pip

pip install tradingview-mcp-server

Claude Desktop Config (claude_desktop_config.json)

Note: On macOS, GUI apps like Claude Desktop may not have ~/.local/bin in their PATH. Use the full path to uvx to avoid "command not found" errors.

{
  "mcpServers": {
    "tradingview": {
      "command": "/Users/YOUR_USERNAME/.local/bin/uvx",
      "args": ["--from", "tradingview-mcp-server", "tradingview-mcp"]
    }
  }
}

On Linux, replace /Users/YOUR_USERNAME with /home/YOUR_USERNAME. On Windows, use %USERPROFILE%\.local\bin\uvx.exe.

Codex Plugin Config

This repository also includes mcp-only Codex plugin metadata:

• .codex-plugin/plugin.json
• .codex-mcp.json

The plugin uses the same PyPI package entrypoint:

{
  "mcpServers": {
    "tradingview": {
      "command": "uvx",
      "args": ["--from", "tradingview-mcp-server", "tradingview-mcp"]
    }
  }
}

After installing or enabling the Codex plugin, restart Codex so the MCP server is loaded in the next session. Depending on your Codex version, codex mcp list may show registered MCP servers, but tool availability should be verified in a fresh Codex session.

Or run from source

git clone https://github.com/atilaahmettaner/tradingview-mcp
cd tradingview-mcp
uv run tradingview-mcp

🛠️ Troubleshooting

🪟 Windows: MCP error -32001: Request timed out on first launch

Symptom — you see this in the Claude Desktop logs shortly after adding the config:

[tradingview] Server started and connected successfully
[tradingview] Message from client: initialize ...
[60 seconds later]
[tradingview] notifications/cancelled — reason: "MCP error -32001: Request timed out"

Why it happens: on Windows with Python 3.14, uvx downloads tradingview-mcp-server, creates a fresh virtualenv, and installs dependencies the first time it runs. Because pandas has no prebuilt wheel for Python 3.14 yet, pip falls back to a source build — which typically exceeds Claude Desktop's 60-second MCP initialization timeout.

Fix — pin to Python 3.13 (has prebuilt pandas wheels):

{
  "mcpServers": {
    "tradingview": {
      "command": "uvx",
      "args": ["--python", "3.13", "--from", "tradingview-mcp-server", "tradingview-mcp"]
    }
  }
}

On macOS use the full path to uvx (see the note in Quick Start). On Windows uvx is typically %USERPROFILE%\.local\bin\uvx.exe.

Alternative — pre-install once, then let Claude Desktop reuse the cache:

# Run in a terminal before launching Claude Desktop
uv tool install --python 3.13 tradingview-mcp-server

After the install finishes, start Claude Desktop with the normal config and the server will come up instantly (cache is already warm).

Credit: @wyh4444 for the original report in #24.

⚠️ Error Envelope Format

Tools that have adopted the structured error format return either their normal payload or an error envelope:

{"error": {"code": "ALL_BATCHES_FAILED", "message": "All 5 batches failed; first error: JSONDecodeError(...)", "batches_attempted": 5, "batches_failed": 5, "first_error": "..."}}

Why: the previous [] / {"error": "Analysis failed: ..."} strings made it impossible to distinguish "no matches today" from "upstream rate-limit cliff." The new envelope is programmatically branchable by code.

Currently adopted by: top_gainers, top_losers, rating_filter, volume_breakout_scanner, smart_volume_scanner. More tools will follow in subsequent PRs.

Detecting an error:

result = volume_breakout_scanner(exchange="KUCOIN")
if isinstance(result, dict) and "error" in result:
    code = result["error"]["code"]
    if code == "ALL_BATCHES_FAILED":
        # Wait + retry, raise alert, fall back to single-batch call, etc.
        ...
else:
    for row in result:
        ...

Stable codes are defined in core/errors.py.

📱 Use via Telegram, WhatsApp & More (OpenClaw)

Connect this server to Telegram, WhatsApp, Discord and 20+ messaging platforms using OpenClaw — a self-hosted AI gateway. Tested & verified on Hetzner VPS (Ubuntu 24.04).

How It Works

OpenClaw routes Telegram messages to an AI agent. The agent uses trading.py — a thin Python wrapper — to call tradingview-mcp functions and return formatted results. No MCP protocol needed between OpenClaw and the server; it's a direct Python import.

Telegram → OpenClaw agent (AI model) → trading.py (bash) → tradingview-mcp → Yahoo Finance

Quick Setup

# 1. Install UV and tradingview-mcp
curl -LsSf https://astral.sh/uv/install.sh | sh && source ~/.bashrc
uv tool install tradingview-mcp-server

# 2. Configure OpenClaw channels
cat > ~/.openclaw/openclaw.json << 'EOF'
{
  channels: {
    telegram: {
      botToken: "YOUR_BOT_TOKEN_HERE",
    },
  },
}
EOF

# 3. Configure gateway + agent
openclaw confi