# Context First, or the Model Is Guessing *Published 2026-06-01 | updated 2026-08-31 | tags: Product* > Why Stocklake's MCP tools start with market and sector context before touching a single stock - and why most tools default to recent data, not everything. Canonical: https://www.stocklake.dev/blog/why-context-comes-first [Last post](https://www.stocklake.dev/blog/stocklake-origin-story) covered where Stocklake came from - a decade-old Google Sheet that turned into a datalake, an app I never shipped, and an MCP server that became the actual product. This one is about why the tools are shaped the way they are. ## Every analysis starts with the market, not the stock The first thing I built out wasn't stock-level data at all - it was the market itself. Here's the reasoning: ask any model to evaluate a single stock without telling it whether the broader market is risk-on or risk-off, and it's guessing. A 15% drawdown in a stock means something completely different in a bull tape than it does three days into a VIX spike. Without that context, "should I buy this" isn't a real question - it's a coin flip with extra steps. So the first tools in Stocklake aren't about any individual company. `get_market_pulse` gives a live snapshot in one call - VIX level, Fear & Greed, market breadth (what % of the universe is oversold vs. overbought), and where the major indices sit. `get_market_assessment` goes one layer deeper: macro regime (RISK_OFF / CAUTIOUS / NEUTRAL / AGGRESSIVE) and market outlook (POSITIVE / NEUTRAL / NEGATIVE) with the reasoning behind each. get_market_pulse VIX 14.2 Fear & Greed 61 · Greed Breadth 58% Oversold / Overbought 12% / 22% OversoldOverbought S&P 500 +0.4% NASDAQ +0.6% DOW +0.2% Russell 2K -0.3% get_market_pulse - VIX, Fear & Greed, breadth, indices in one call ## Sectors are the next layer down Broad market context alone isn't enough either. A NEUTRAL market can still have Energy ripping and Tech bleeding - sector rotation is often the more useful signal than the index level. `get_sector_intelligence` covers all 11 GICS sectors: which ones are LEADING, STRONG, NEUTRAL, WEAK, or LAGGING right now, plus the underlying stats (RSI distribution, breadth, 1-week/1-month performance) that back up the label. Macro plus sector, together, is the foundation almost everything else in Stocklake sits on top of. ## History matters as much as the snapshot A live number on its own is context-free. Knowing the market is NEUTRAL today doesn't tell you if it just became NEUTRAL after two weeks of POSITIVE, or if it's been stuck there for a month. So the tools carry historical depth, not just a point-in-time read - sector signals, market outlook, and regime assessments all track how they've moved over recent history, not just where they sit right now. That said - deliberately *not* unlimited depth. Which brings me to a design choice I want to be upfront about. ## Why the default window is recent, not everything Early on, I gave a tool wide latitude - pull as much daily history as you want. An agent took that literally: it queried a full year of daily data for a routine check, dragged that entire year into its reasoning, and the output drifted somewhere I never intended. Not because the data was wrong - because nothing constrained how much of it the model decided was relevant, and different models handle that differently. Some stay disciplined with a large context window; some don't. So most Stocklake tools default to a recent window - days or weeks, not a year - unless you explicitly ask for more. It's a guardrail, not a limitation: the goal is to keep the model reasoning over what's actually relevant to now, not quietly averaging in six months of noise it was never asked to consider. News data has the same problem, worse - an agent reasoning about a stock using six-month-old headlines as if they're current is its own failure mode. More on that in a later post. On Pro and hit a gap here - history you need that the defaults don't surface? Tell us at [/contact](https://www.stocklake.dev/contact). This tuning is ongoing, not finished. ## Raw context isn't the end of it - synthesize it, then keep it Handing an agent the raw numbers is step one. Step two was asking: why make every caller re-derive the same conclusion from the same data? So the same macro and sector context that feeds `get_market_assessment` and `get_sector_intelligence` also runs through an AI synthesis pass - regime, outlook, sector signal, with the reasoning attached - and that output gets written back into Stocklake, not just handed to whoever called the tool and forgotten. Raw data AI synthesis pass Regime + outlook + reasoning Written back to Stocklake Raw market & sector data → AI synthesis pass → written back to Stocklake That's the part I actually care about most: it's not a one-off API response. It lands back in the same datalake, so the next call - mine or yours - starts from that synthesis instead of re-deriving it from scratch. One synthesis layer, built once, reused everywhere. ## What this foundation unlocks With macro, sector, and historical context loaded, the tools go deeper from there - into individual stocks. `get_stock` for price, fundamentals, and technicals in one call (with a built-in 0-10 rating on Pro), `get_screener` for filtering the universe by RSI, sector, market cap, or valuation. All of it sitting on top of the same market context, not floating disconnected from it. News and sentiment are their own layer on top of that - and their own set of lessons about what happens when an agent doesn't know how old a signal actually is. That's one of the next blog posts.