MarketCue reads the Indian market continuously through the trading day and turns it into one decision, not another chart. Four Claude-powered agents form the view, manage it, and grade it. Every milestone below is drawn from merged pull requests on main.
Every number on screen is computed by the data pipeline. The models only form and explain the view. Each agent has a specification of record in the database: model, schedule, tools, guardrails, evaluation.
The analyst. Sets the day's context at 08:48, forms the view of record at 09:35, manages it at five hourly checkpoints against a code-checked invalidation, and grades itself after the close.
Reads the Global to India transmission engine and writes the narrative for the GlobalCue tab. A second model call judges every paragraph before it publishes.
Clusters headlines into events, scores each one, and feeds the score into both the Nifty and Sensex verdicts as a bounded component.
The pipeline that computes every figure: PCR, max pain, IV, straddle, expected move, targets. Haiku writes only the short pre-market headline and recap on top of those numbers.
The first product: a rules-based strategy engine and a dashboard built for a phone screen, with GIFT Nifty framed as a predictor of the Nifty open.
Trading strategy logic iterated five times in three days, mobile responsiveness alongside.
1 to 5Market data dashboard launched with mobile performance work and a refreshed brand.
6 to 16“Today” tied to the real IST trade date, not the latest post.
19GIFT Nifty gap as a Nifty-open predictor, raw comparison against previous close.
20 to 22The read leaves the app. Pre-market at 08:55 and post-market at 21:30 IST go to the site, Telegram and X, and eight evergreen pages start pulling in search traffic.
Daily reports to Telegram and X, cron moved to 08:55 and 21:30 IST.
23 to 29Journal with screenshot uploads and a compact saved-notes log.
25SEO rounds P0 to P3: crawlable playbook, schema markup, sitemap, IndexNow, eight evergreen “today” pages.
31 to 42llms-full.txt with real markdown tables for AI crawlers.
35Repositioned as an AI-Agentic Option Intelligence Platform, sitewide except the playbook.
44 to 45Upstox timeout fix in market-data-sync, the pipeline's first hardening.
48The product gets accounts, a token-based design system on IBM Plex, dark as the default, and a dashboard that conforms to a written spec. Two mobile audits land.
Google sign-in on login and signup, confirmation links straight to the dashboard.
51 to 54Design system phases 1 to 4: tokens, IBM Plex, components, a formatter, a /design specimen route.
57 to 59Archive with real table semantics, month grouping, bias filter, pagination.
60Dashboard conforms to DASHBOARD-SPEC, every screen ported to the anatomy.
64 to 73Dark is the default theme, warm dark ramp, brand amber restored.
71, 74, 85The ledger treatment: unboxed content, retuned spacing and type, signed-value colour.
84 to 98Browse any past trading session.
89An intelligence layer sits over the playbook. It states a view, checks it hourly against its own invalidation, writes its narrative under a publish gate, and reads global markets and news as scored inputs.
Rule-becomes-agent layer on Claude Sonnet 4.5, one forced tool, a spec-of-record row per agent in the database.
105 to 117Checkpoint state and code-checked invalidation, the agent manages its own view through the session.
111Every surface reads the agent's view, strategy as the answer, capped text.
113, 116Chart screen after Verdict, from Kite one-minute candles.
121GlobalCue screen, free quotes and RSS headlines.
124Global to India engine, phase 1 and 2a: history analytics, measured transmission, quantitative events.
129, 133Sonnet-written narrative behind a Sonnet judge, a publish gate that stores rejection reasons.
140, 144, 149, 153Five-minute news agent, clustered events scored into both verdicts, score bounded.
145 to 148Timeframes, fullscreen and a standalone tab, then a plug-in framework that let twelve indicators ship in four days, each with its own reference page.
Timeframes, PHLC and OI walls, fullscreen, plain background, Indicators panel.
138 to 150Indicator framework: legend stack, catalogue, templates, Intraday and Chart levels.
154Futures volume feed powering VWAP and a volume pane.
155Standalone full-page chart tab.
157OI wall visibility and default-view fixes, crosshair toggle.
158 to 166Volume Profile, CVD proxy, FII/DII flow as indicators, plus per-indicator SEO pages.
167 to 173Market Profile (TPO), Power Scanner, Market Pulse, Derivatives Positioning.
179 to 185The one piece of the stack that needs a long-running process: a persistent Kite WebSocket worker solving implied volatility per strike every five seconds. Open-interest snapshots start logging every minute.
Nifty Gamma Exposure tab: zero-gamma flip, call wall, put wall, levels withheld under 60% strike coverage.
170, 172, 190GEX architecture documented, compute worker on Railway, singleton row in Supabase.
174One-minute OI snapshot logging, isolated from market-data-sync.
176OI History Session Map tab, wired to real snapshot data.
177, 191Per-strike option-chain history, three-minute cadence, isolated.
178Options Print and Greeks at a glance.
189The product starts running itself: an admin panel with pipeline health, a changelog that writes on every merge, and a home page rebuilt against the final prototype.
Admin panel: Users, Changelog, Pipeline health.
193 to 194Auto-changelog triggered from GitHub on merge.
196Depth pass across home, login, signup, docs and article pages.
195 to 201Home rebuilt: “One read” panel, session journey as a ledger, ticker that fills the strip.
203 to 207One clock and one profile menu in the dashboard topbar.
209Draft items for review. Nothing here is shipped.
Verify both new pipelines on live Monday data after the weekend deploy.
plannedOwn sentiment engines for Global to India transmission, event-driven news, the model explains only.
plannedMobile audit of the chart and premium tabs.
planned