Every AI finance app wants your data. I didn’t trust that — so I built my own. Offline.

What I Built

If you spend any time browsing AI finance tools right now, you'll quickly realize that half of them are just shiny, sleek UI wrappers designed to do one thing: beg you to paste your portfolio data into their cloud servers.

I don't know about you, but the idea of handing my financial watchlist over to a random startup's black box just didn't sit right with me. So, I decided to ignore the SaaS products and build my own setup from scratch.

What I ended up with is a fully local, AI-powered market assistant. It’s a hybrid beast—part retro terminal CLI, part clean web dashboard. It silently tracks live stock prices in the background, runs deep sentiment analysis on market news, and actively pings me on Telegram and WhatsApp the second it thinks it's time to hold or buy.

The entire core idea revolves around absolute, uncompromising privacy. The whole "brain" of this operation runs directly on my laptop. None of my data, my watchlists, or my trading logic ever leaves my machine. There are no expensive monthly API subscriptions to OpenAI or Anthropic, and no cloud servers analyzing my trades. It's just a local pipeline crunching numbers in the background, and honestly, it feels incredibly cool to have that kind of autonomy.

For those of you who live in the terminal, here is exactly what the CLI output looks like when you point it at a watchlist and let it do its thing:

$ openclaw scan --watchlist AAPL,TSLA,NVDA

[09:31:02] Fetching live data...
[09:31:04] Running analysis (local)...
[09:31:07] Scanning 3 tickers...

  NVDA  │ BUY   │ Score: 87/100 │ Volume spike +23%
  AAPL  │ HOLD  │ Score: 54/100 │ Consolidation phase
  TSLA  │ HOLD  │ Score: 41/100 │ Mixed signals

✅ Notifications sent → Telegram + WhatsApp

How It Works

I wanted the pipeline to be ruthless in its efficiency but straightforward enough that I could actually explain it to someone. It breaks down into five distinct phases:

• Data collection: First things first, I need the numbers. I pull live prices, volume metrics, and moving averages from free financial APIs like Alpha Vantage and Finnhub. To keep things fast and avoid getting blocked, everything gets cached straight into a local SQLite database.
• Sentiment analysis: Numbers only tell half the story. The system reaches out and grabs recent news headlines for every ticker on my watchlist. It bundles all that text up and feeds it to the local LLM essentially to read the "vibe" or mood of the market that day.
• Correlation: This is the part I think is actually really cool. The system actively forces the LLM to correlate the soft, fuzzy sentiment data (how the news sounds) with the hard, undeniable technical data (actual price action and volume spikes).
• Decision: Based on that correlation, the LLM isn't allowed to write a summary. It has to output a strict, no-nonsense verdict: BUY, HOLD, or SELL, alongside a hard confidence score from 0 to 100.
• Reporting: If a signal manages to cross my pre-set confidence threshold, it triggers an alert. A cleanly formatted message gets pushed directly to my phone via Telegram and WhatsApp bots, while the web dashboard quietly logs the history for me to review later.

My Experience

I want to be completely honest here: I am not a trader. Not even close. I went into this project knowing almost nothing about how real markets operate.

Surprisingly, that ended up being my biggest advantage. I didn't have any bad trading habits to unlearn, and I wasn't trying to force the AI to fit some preconceived trading strategy.

The coding part was an absolute blast, but the actual learning curve was humbling. I quickly realized I couldn't just type "tell me if this stock is good" into a prompt. I had to sit down, read up on what moving averages actually represent, and understand why volume matters before I could even construct a prompt that made sense to the AI.

The biggest problem I faced by far was hallucinations. Early on in the build, the local LLM confidently told me to buy a specific stock because of a "massive volume spike." I got excited, checked the raw data myself, and realized there was no spike. The AI had completely fabricated the number just to sound authoritative and confident.

That was a massive reality check. I had to go back and build an entirely separate Python verification layer just to double-check the AI's reasoning against the actual math before I allowed any notification to reach my phone. It felt like building guardrails for a very enthusiastic but slightly delusional robot.

I built this because I wanted to master data pipelines, but mostly I built it because I genuinely believe you shouldn't have to trade your personal financial data just to get a simple alert that a stock is moving.

OpenClaw Features Used

I tried building this with pure Python at first, and the code quickly turned into spaghetti. OpenClaw ended up being the absolute glue that held this entire chaotic pipeline together. Here is exactly how I used it:

• Multi-agent pipeline: Instead of one massive script, I used OpenClaw to separate my logic into distinct, focused agents. I have a "Data Fetcher" agent whose only job is getting numbers, a "Quant Analyzer" agent to do the math, and a "Notifier" agent. OpenClaw manages the handoffs and communication between them seamlessly.
• Local LLM (Ollama): I am running a local model via Ollama. Instead of spending days writing custom boilerplate to handle streaming, timeouts, and prompt routing, I just used OpenClaw's native integration. I point it at my local model, and it just works.
• Structured decision system: LLMs love to talk. They want to explain their reasoning in paragraphs. I needed strict data. I used OpenClaw's structured output constraints to essentially wrestle the AI into returning exact JSON formats ({"signal": "BUY", "confidence": 85}). If the LLM tries to get chatty, OpenClaw flat-out rejects the response and forces a retry.

Try It Yourself

If you want to run your own completely private, local quant desk just for the fun of it, the code is entirely open-source. I warn you, it might be a little janky in places—but that's the beauty of building in public.

You can check out the repository, read through the setup guides, and get it running on your own machine right here:
OpenClaw Finance Assistant on GitHub