TNT-Likely Releases PanWatch: Self-Hosted AI Trading Assistant

What Happened

TNT-Likely released PanWatch, a self-hosted AI trading assistant that wraps the TradingAgents multi-agent investment decision framework in a monitoring and notification layer. The project covers A-shares, Hong Kong equities, and US markets, providing real-time monitoring with multi-channel push (e.g., messaging platforms, webhooks). It reached 175 GitHub stars within the first day of public availability.

Why It Matters

Multi-agent investment reasoning has largely lived in research papers and closed platforms; PanWatch packages it as a deployable, self-hosted service. For builders constructing finance-focused agent stacks, this provides an inspectable reference implementation covering the full loop: data ingestion, agent deliberation, decision synthesis, and event-driven alerting. It also addresses a specific operational gap — most existing TradingAgents forks stop at notebook-level inference, leaving operators to build scheduling, state, and delivery infrastructure themselves. The self-hosted posture matters for teams handling portfolio data or targeting non-US markets where third-party API custody is a compliance liability. Anyone evaluating whether multi-agent consensus adds edge over single-model prompting now has a cheap way to instrument that question in production-like conditions.

Technical Details

PanWatch builds on the TradingAgents pattern: specialized agents (fundamental, technical, sentiment, bull/bear debate, risk management) whose outputs are aggregated into a final decision rather than emitted independently. It adds a monitoring daemon that polls market data on configurable intervals, triggers agent runs on watchlist conditions, and routes results through push adapters. Market coverage spans A-shares, HK, and US, implying dependency on region-specific data sources — likely akshare, yfinance, or equivalent libraries — with differing latency and rate-limit profiles. Self-hosting requires Python runtime, persistent storage for watchlists and run history, and outbound access to both LLM APIs and market data endpoints. Cost and performance are dominated by LLM token consumption per decision cycle; no published benchmarks for latency or decision quality were included in the release.

Operational Impact

Builders can now fork PanWatch instead of assembling a monitoring skeleton from scratch, collapsing days of integration work into configuration. The monitoring-plus-push architecture shifts agent usage from on-demand querying to event-driven execution, which changes cost modeling: token spend becomes a function of watchlist size and polling frequency rather than user sessions. This makes per-decision cost the primary tuning knob, and teams will need rate limiting, deduplication, and decision caching to avoid redundant agent runs on unchanged data. For operators running personal or small-fund portfolios, the practical result is a lower barrier to running structured, multi-perspective analysis overnight across three markets. Distribution defaults — push adapters, market calendars, timezone handling — become the day-to-day friction points rather than model selection.