SignalDesk Local

A private, local-first, AI-powered market research & trading-journal dashboard. Runs entirely on your laptop via Docker. No broker, no live trading, no paid infrastructure, no multi-user accounts — a personal research tool that blends TradingView-style charts, a Notion-like reading view, a simplified Bloomberg-terminal feel, and an AI research assistant.

Not a SaaS product, not a platform, not a trading-execution system.

Screenshots

Dashboard (charts + signals)	Daily briefing + provenance

Watchlist (grouped, drag-reorder)	Compare view (2 charts)

Signals (overlays + legend toggles)	Macro tab (FRED, deterministic)

News tab (RSS + descriptive brief)	Hidden-asset search picker

The idea: a 3-tier "brain"

A reviewer pipeline, not parallel opinions — each tier has a distinct job and everything is provenance-stamped so you always know who said what:

Tier 0 — Signals (pure Python, deterministic)
   trend / RSI / ATR / Bollinger / 52w / volume / VIX regime / flags
   → can't hallucinate, always works, the trustworthy floor
        │
Tier 1 — Auto-draft (free LLM, scheduled 10:30 ET, push)
   segmented daily briefing + per-asset notes; deterministic risk-on/off
   call the model EXPLAINS (never decides); templated fallback if no key
        │
Tier 2 — Deep review (strong model via MCP, on demand, pull)
   Claude / Codex read the real local data and ANNOTATE the draft
   (append-only — the Tier-1 draft is never overwritten)

The AI narrates deterministic facts; it never invents numbers or prices. Conclusions (the risk call, the "signals to watch" list) are computed, not the model's opinion.

Features (built)

• Dashboard shell — persistent watchlist rail + 1–2 chart workspace, selection in the URL (bookmarkable), single-screen feel.
• Watchlist — grouped (Equities & Indices / Crypto), drag-to-reorder, live price + % change + trend per row.
• Charts — TradingView lightweight-charts candlesticks + volume, with toggleable overlays (MA20/50/100, Bollinger, 52w hi/lo) via a clickable legend (per-signal + master on/off), 3M/6M/1Y ranges, compare 2 assets.
• Tier-0 signals — pure deterministic engine; trend, %, RSI, ATR, 52-week distance, volume flags, VIX regime.
• Tier-1 briefing — automatic daily digest segmented by asset class, deterministic risk-on/off call, "signals to watch", day-over-day deltas; free-LLM provider chain (Groq → OpenRouter → templated fallback); scheduled by a dedicated cron container; provenance-badged in the UI.
• Archive — browse past briefings by date; Tier-2 annotations layered beneath the draft with their own provenance badges.
• MCP server — one client-agnostic stdio server (Claude Desktop / Claude CLI / Codex); the strong model reads local data and writes one consolidated Tier-2 annotation onto the daily briefing (the single Tier-2 surface), attributed per client.
• Desk dock — a trading journal (real / paper / observation) with thesis-now / lesson-on-close and computed P/L (never typed), plus Notion-like markdown notes (pin, per-asset). Bottom dock mirrors the briefing bar: the desk speaks at the top, you write back at the bottom.
• Macro & News — the top dock is a 3-tab desk: Briefing | Macro | News. Macro = deterministic FRED series (yields, curve, Fed funds, unemployment, CPI YoY) — the same numbers are handed to the briefing to narrate (never invent). News = free RSS headlines (CNBC / WSJ / Fed) with an on-demand, descriptive LLM brief that is deliberately walled off from the deterministic risk stance.
• Crypto Fear & Greed — fourth tile on the briefing card alongside TREND / BREADTH / VOL. Real values from alternative.me (free, no key), 6 h read-through cached. Display-only and explicitly walled off from the deterministic risk stance (sentiment ≠ stance).
• Hidden assets + search-to-add — every asset row has an enabled flag. The sidebar's + button opens a search popover over the hidden pool (MSTR, MSFT, GOOGL, META, AMZN, AMD, NFLX, COIN, GLD, SLV, EGP/USD, OIL, XMR-USD); click to add, hover-× on any row to remove. Hidden rows incur zero compute cost — every analysis path filters on enabled_only=True, so they're inert metadata until you promote them.

Tech stack

Next.js + TypeScript + TailwindCSS · TradingView lightweight-charts · FastAPI · SQLite (WAL) · Python MCP server · Docker Compose · yfinance (equities/index) + Binance public REST (crypto) · free/OSS only.

API on :8081, frontend on :3000.

Quick start

Requires Docker Desktop running (provides the Linux VM on macOS).

cp .env.example .env        # optional: add a free GROQ_API_KEY for real AI
docker compose up --build

• Dashboard: http://localhost:3000
• API + docs: http://localhost:8081/docs

Without an LLM key everything still works — the briefing uses its deterministic templated fallback. Add a free Groq key to .env for real AI briefings (the cron + the ↻ regenerate button use it). A free FRED_API_KEY (fred.stlouisfed.org) enables the Macro tab; everything degrades gracefully without it.

Run modes

• Dev (default) — docker compose up --build. Turbopack + hot reload; edit a file and it updates live.

• Prod (opt-in, fast) — precompiled, ~50–150ms page loads, no hot reload. Use when you just want to use the app:

  docker compose -f docker-compose.yml -f docker-compose.prod.yml up --build

  (Needs Docker Compose ≥ 2.24 for the !override volume tag.)

Connect the AI (Tier 2)

See mcp/README.md for one-time setup of Claude Desktop, Claude CLI, and Codex. The server is identical for every client; only each client's config differs (an MCP_AGENT env var stamps provenance).

Architecture & design

Full design rationale and the phased build plan are in ARCHITECTURE.md.

Key principle: one source of truth (SQLite); the FastAPI backend is the only process that opens it. The dashboard, the cron, and the MCP server are all just clients of that backend.

Project layout

backend/    FastAPI app — data adapters, signal engine, briefing, cron
frontend/   Next.js dashboard (App Router)
mcp/        client-agnostic MCP server (Tier-2 bridge) + setup docs
data/       SQLite DB (gitignored — host-visible for GUI inspection)
ARCHITECTURE.md   design + phased plan

Notes

• Local-first & single-user by design. .env, the SQLite DB, and node_modules/venvs are gitignored.
• Free equity data is delayed (~15 min); this is a research/journaling tool, not a real-time trading system.