DEV Community

Rajaram > aarna
Rajaram > aarna

Posted on

Stop Building Crypto Data Pipelines. We Built the Fix.

#webdev #cryptocurrency #ai #javascript

Over the past few months we kept hearing the same story from teams building crypto predictive models and AI agents. It did not matter if they were ML engineers, quant researchers, or indie developers shipping crypto tools -- the first chapter of every project sounded identical.

Connect to an exchange. Normalize OHLCV candles. Compute indicators. Handle gaps. Resample. Repeat for every token.
Every single team.

The problem we kept hearing

The engineers we spoke to were not complaining about the modeling. They were complaining about the infrastructure that came before it.

One team told us they spent three weeks just getting clean, consistent RSI and MACD data across six tokens before they could train a single model. Another said their model runs varied between experiments not because the model changed -- but because the data pipeline produced slightly different outputs each time.

The insight that stuck with us:

Building a crypto predictive model is a data problem before it is a modeling problem.

What we built

We built aTars - the featurization and dataset access layer for teams building predictive models on crypto markets.

aTars is a free MCP server that exposes a pre-computed, continuously refreshed crypto feature dataset across 9 major tokens as queryable tools - accessible to any AI agent or model in a single endpoint call.

40+ pre-computed indicators
1-min resolution · 90-day history · ~129K candles per token
BULLISH / BEARISH / NEUTRAL signal summaries
Resample on query to 1min, 1h, 4h or 1d

Tokens covered: BTC, ETH, SOL, XRP, BNB, ADA, DOGE, TRX, BCH
The five tools your agent gets:

atars:get_latest_features -- latest indicator snapshot for any token
atars:get_category_features -- indicators by category (trend, momentum,
volatility, volume)
atars:get_sentiment_symbols -- sentiment overview across all 9 tokens
atars:get_latest_sentiment -- current sentiment for a specific token
atars:get_sentiment_trend -- sentiment trend over time for a token

How to connect it to Claude desktop
You need Node.js installed. Check with:

node --version
Enter fullscreen mode Exit fullscreen mode

Then find or create your Claude desktop config file:

Mac:

~/Library/Application Support/Claude/claude_desktop_config.json
Enter fullscreen mode Exit fullscreen mode

Windows:

C:\Users\YOUR_NAME\AppData\Roaming\Claude\claude_desktop_config.json
Enter fullscreen mode Exit fullscreen mode

Add this:

{
  "mcpServers": {
    "atars": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@0.1.38",
        "https://mcp.aarna.ai/mcp"
      ]
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Fully quit Claude and reopen it. Then ask:

Is BTC bullish right now? Show me the full indicator breakdown.
Enter fullscreen mode Exit fullscreen mode

You should see this:

🔴 ATARS BTC Analysis
Technical Indicators → BEARISH

Indicator        Signal
─────────────────────────
MACD             ✅ Bullish
RSI (14)         ⚪ Neutral
EMA Cross        🔴 Bearish
ADX Direction    🔴 Bearish
Enter fullscreen mode Exit fullscreen mode

Screenshot of atars MCP answer

Live data. Pre-computed. No pipeline.

What developers told us they use it for

Based on the conversations that shaped aTars, three use cases came up repeatedly:

Predictive model training - pulling a clean feature matrix for supervised learning without maintaining a Binance pipeline. One query returns 720 rows of hourly data across 42 features, ML-ready.

AI trading agents - agents that need to reason about market conditions before making decisions. One tool call returns a structured signal verdict instead of raw numbers the agent has to interpret itself.

Quant research - consistent, reproducible indicator data across tokens and timeframes. Same implementation every run. No variation because the data layer varies.

What it is not

This came up in our developer conversations too so worth being direct about it.

aTars is not for live execution strategies on 5-minute timeframes across multiple exchanges. The hourly refresh and single-exchange coverage make it the wrong fit for that use case. If that is what you are building, it will not serve you well today.

It is infrastructure for agents and predictive models — not sub-hourly execution.

The endpoint

https://mcp.aarna.ai/mcp
Enter fullscreen mode Exit fullscreen mode

Free. No API key. No account required. Compatible with Claude, Cursor, Windsurf, Cline and any MCP-enabled framework.

If you are one of the developers who told us about your pipeline pain -- this is what we built from those conversations. Would love to hear if it solves it.

And if you hit a limitation, tell us. The feedback we got from developers is what shaped v1. The same feedback will shape what comes next.

Top comments (0)