easy-stock: An AI Research Workbench for the China A-Share Market

A local-first desktop workbench for market discovery, stock research, portfolio inspection, and AI portfolio optimization

_{English | 简体中文}

Let AI actually understand the market — and make every judgment traceable to evidence.
Turn intraday observation, post-market review, and long-term insight into one continuously evolving research system.

Download · Why · Core Features · How AI Helps · Architecture · Getting Started · Discussions · Contributing · License

---

Why easy-stock

An AI-era stock terminal shouldn't just bolt a chat box onto a legacy quote app.

Traditional market software is good at showing prices, changes, and turnover. What is actually hard about A-share trading is understanding the structure behind the numbers: What is today's main theme? Is the theme broadening? Did the limit-up ladder extend higher? Did yesterday's strongest stocks get follow-through? Is the trend still intact? Is market sentiment repairing or fading?

easy-stock is an AI-native workbench that speaks this language. It unifies quotes, themes, limit-up pools, news, and post-market reviews into a single local-first research pipeline:

AI-native capability	How easy-stock does it
Sense the market	Aggregates Kaipanla, Eastmoney, Sina, Cailianpress and other sources into one unified model of quotes, K-lines, themes, limit-ups, and news
Understand structure	Turns trends, themes, limit-up ladders, sentiment, volume/price, and relative strength into a computable domain model
Execute tasks	Scheduled jobs, a built-in browser, and agents that discover articles, deduplicate, archive, and distill opinions
Build memory	Keeps articles, summaries, sentiment history, analysis records, model sessions, and research caches on your machine
Verify conclusions	Preserves scoring dimensions, theme sources, original URLs, timestamps, and degradation state — every conclusion can be traced back to evidence

easy-stock does not make decisions for you. It gives you better perception, faster research, and a judgment system that compounds over time.

Note: The app UI and docs are currently Chinese-first (the product targets the China A-share market, where data sources and market microstructure are highly localized). English localization is on the roadmap; the codebase, architecture, and this README are English-friendly.

---

Core Features

Compared with v1.5.0, the current main branch includes fixes for news and earnings-disclosure evidence assessment, portfolio optimization recovery, and score presentation. These fixes are not included in the v1.5.0 installers.

01 · Automated Post-Market Review Digest

Let AI collect and organize what multiple market writers actually said

Post-market reviews from well-known A-share commentators are scattered across Xueqiu, TaoGuba, and WeChat. Manually visiting each profile, filtering for today's articles, and copying text is a time sink. easy-stock organizes all of it into a unified review timeline:

Once collection finishes, one click generates a "today's consensus" report — distilling shared focus areas, key disagreements, market facts, and conditions to verify in the next session:

Scheduled reviews use the Shenzhen Stock Exchange calendar and run on the calendar day before each trading session, at 22:00 Beijing time by default (configurable). The article window extends from the latest completed close to the scheduled time. Select Feishu and/or DingTalk for complete Markdown analysis, including every author's views but excluding article bodies and evidence details; long reports are split automatically. Keep the app running and the computer awake, and sync articles beforehand. Missed runs catch up only before 09:30 on the target trading day and wait for any active review. The calendar refreshes daily, with complete cached months usable for up to seven days during outages; without a usable calendar, scheduling pauses and retries.

The goal is not "AI guessing what rises tomorrow" — it is turning dozens of unstructured articles into a readable, verifiable opinion map you can keep tracking.

02 · Limit-Up Ladder & Sentiment Cycle Analysis

See the ladder, the promotion rate, and where you are in the sentiment cycle

A-share daily price limits vary by board and stock status. Reading the limit-up ladder—how many stocks reached the limit, for how many consecutive sessions, and which rung failed—is central to short-term market analysis. easy-stock puts the limit-up pool, consecutive limit-up ladder, yesterday's follow-through, promotion rates, and sentiment history in one view, with a built-in algorithm for sentiment-cycle staging:

03 · Theme Radar

Identify the market's real leading themes from sector moves

Aggregates theme rankings, momentum strength, money flow, market breadth, and streak duration; breaks down industry chains, concept nodes, and sub-directions through a theme map; and combines AI reads of the market trend, theme stage, and conditional entry points for trend stocks.

Daily, weekly, and monthly theme-index candles with volume help track the sector as a whole. The interface distinguishes provider indices from equal-weight reference indices.

View the theme-index candlestick example

04 · Market Overview

One traceable map of indices, money flow, leaderboards, and research signals

Core indices, news flashes, industry trend strength, sector money flow, themes, per-stock inflows/outflows, the Dragon-Tiger list, announcement radar, institutional views, and industry research — unified into a single research entry point. Every page keeps its data source and fetch time, and you can hand the current context straight to AI for interpretation, instead of hopping between quote terminals and news pages.

05 · Tiered Stock Research

Choose your research purpose, horizon, and depth

Select observation, a potential new position, or an existing holding; choose a short, swing, or medium-term horizon. Cost is optional for existing holdings.

Research mode	Use case
Quantitative preview	Inspect prices, trends, and a quantitative baseline without calling an AI model
Quick AI assessment	Summarize the main thesis and questions that need checking
Standard AI assessment	Compare business drivers, catalysts, supporting evidence, and counter-evidence
Deep AI research	Investigate core disagreements further and develop scenarios and follow-up checks

Reports distinguish source statements, third-party opinions, and research inferences. They retain citations and evidence timestamps alongside the business thesis, disagreements, counter-evidence, confirmation conditions, and invalidation conditions. Completed research stages are saved and can be resumed when recovery requirements are met.

06 · Portfolio AI Inspection

Evaluate individual research in the context of the whole portfolio

Save multiple portfolio plans: click + above the setup form to add one, then switch between plans to edit their separate stocks, weights, costs, and research settings. Plans can be renamed or deleted, and changes are saved locally.

DingTalk and Feishu robot settings support stock AI research and manual portfolio inspection notifications. Both send complete Markdown results without evidence or source details, splitting long reports automatically. With failure alerts enabled, failed or incomplete tasks send their status and available results; tasks without a report send a status explanation.

Each plan supports scheduled inspections every 3 days, weekly, monthly, or every N days/weeks/months, with a start date and execution time in Beijing time. Select DingTalk and/or Feishu to receive complete Markdown inspection results through enabled robots in Settings, excluding evidence and source details and splitting long reports into numbered messages; leave both unchecked to save reports only. Failure alerts follow the system notification setting. Schedules use the holdings and research settings saved with them; save the schedule again after changing a plan. The app must be running and the computer awake. Missed runs are consolidated into one catch-up inspection, active inspections finish before the next starts, and monthly dates beyond the end of a month use its last day. Deleting a plan stops its future scheduled inspections.

Choose an aggressive, balanced, or steady style, set a holding horizon, and add up to 10 stocks. Total assets default to CNY 500,000 and are editable. Costs default to a valid quote fetched when adding a stock and can be overridden. Whole-share quantities are estimated from assets, weights, and costs; the form shows current prices, market values, and per-stock and aggregate profit/loss. Unallocated funds remain cash. Inspection reuses valid AI stock reports from the past 24 hours, researches missing stocks, and produces a portfolio conclusion. Individual stock research can also be refreshed separately.

View portfolio setup: plans, assets and P/L, trading style, weights, and costs

The setup screenshot shows a saved plan with market value, remaining cash, and P/L at the top, followed by trading style, research settings, and per-stock weights and costs. Share quantities and market values are estimates based on the inputs and quotes.

AI explains four scoring dimensions; the application recomputes and validates their fixed-weight total: holding thesis 35%, portfolio structure 25%, risk management 25%, and strategy fit 15%. Scoring evaluates the stock portfolio itself. Being fully invested, the cash percentage, and total exposure do not add or subtract points; concentration uses weights within the stock portfolio.

The inspection shown covers 4 holdings, reusing 3 reports and researching 1 stock anew. It scores 65 with medium risk. The four cards explain the holding thesis, portfolio structure, risk management, and strategy fit, with expandable scoring evidence and sources.

The report also retains per-stock judgments, common drivers, historical correlations, research sources, and confirmation and invalidation conditions. Existing stock reports open directly. Tasks run in the background, with reports and recovery state saved locally.

07 · AI Portfolio Optimization

Compare the original and proposed portfolios—and explain where the capital goes

Start optimization from an inspection report. Candidate screening, stock research, investment comparisons, allocation search, and independent review produce target weights, funding sources, and entry and exit conditions. The process compares adding to existing holdings with introducing new stocks.

• Capital constraints: total stock exposure and cash remain fixed. Cumulative replacement is capped at 70% of the optimization chain's initial stock allocation, with at most 2 new stocks and 10 target holdings.
• Paired review: original and proposed portfolios are independently scored using the same evidence and market snapshot. Improvement uses that review's scores. The earlier inspection score remains separate; two AI evaluations may differ.
• Substantive improvement: investment value, structure, trading eligibility, and risk constraints are checked together. A higher score alone does not qualify a proposal. Results may be conditional, a qualified alternative, or no adjustment.
• Saved progress: stock judgments, investment comparisons, and review blocks are saved separately. Recovery continues missing work; restarting can still reuse valid research.

The screenshot retains the original inspection score of 65 and separately shows optimization review scores of 64 for the original holdings and 74 for the proposal: a 10-point improvement in the paired review. The side-by-side view shows reductions, unchanged weights, and additions. The pending-entry notice means the target requires price or other conditions to be met; the proposed allocation has not been executed.

Applying a proposal starts a new inspection of its target holdings. It does not place orders or change actual positions. Scores evaluate current research evidence; they are not expected returns or win probabilities.

08 · Trading Wisdom Library & AI Copilot

Turn hard-won trading experience into durable, re-readable knowledge — and let AI learn how you research

The Daily Research Loop

Stage	What easy-stock provides
Intraday discovery	Market overview, trend themes, live quotes, theme map, limit-up ladder, and data source status
Stock research	Quantitative preview and tiered AI research, business drivers, evidence and counter-evidence, conditions, and scenarios
Portfolio inspection	Valid report reuse, four-dimensional scoring, concentration and correlation risks, per-stock judgments, and sources
Portfolio optimization	Candidate screening, investment comparisons, allocation search, paired review, and conditional proposals
Post-market review	Sentiment timeline, yesterday's follow-through, promotion structure, and limit-up ladder analysis
Information gathering	Automatic profile sync from commentators, article import, text cleaning, and local archiving
AI distillation	Per-article summaries, per-author synthesis, cross-author consensus, disagreements, and next-session watch conditions
Long-term accumulation	Article library, trading wisdom, stock analysis records, AI sessions, and local research memory

---

How AI Augments A-Share Research

1. From "looking at data" to "understanding market structure"

easy-stock first organizes themes, ladders, limit-up reasons, trends, relative strength, and sentiment history into a domain model — then hands that structure to AI. The model never faces isolated numbers; it faces market evidence with A-share semantics.

2. From "manual browsing" to "agents that act"

Desktop Browser Bridges first reuse persistent Electron sessions to visit subscribed Xueqiu or TaoGuba profiles and read articles. Without a desktop bridge, the selected agent can collect content through agent-browser. The Go review service normalizes metadata, deduplicates, archives, and distills the articles.

3. From "single summaries" to "an opinion network"

The system first synthesizes each author's core views, then aggregates across authors for the day's consensus and disagreements — separating market facts, minority expectations, shared focus, and conditions to verify next session.

4. From "generic chat" to "an A-share research copilot"

AI sessions use a local shared agent service with a choice of Hermes or Codex. The runtimes share model profiles, reasoning settings, Skills, and MCP configuration while keeping their native sessions separate. Business tasks bind their runtime and model configuration for model calls, session continuation, tools, and task context. OpenAI, DeepSeek, Qwen, Moonshot, Anthropic, and custom compatible endpoints can be configured; Codex requires a Responses connection.

5. From "one-off Q&A" to "a long-term research flywheel"

Perceive	Understand	Act	Remember	Verify
Live quotes, themes, articles	AI identifies structure, views, disagreements	Auto-sync, analyze, draft plans	Local articles, history, sessions	Verify next session against prices and sources

Every verification becomes context for the next round of research. Over time the system accumulates not just more data, but a research process and a body of judgment that fits the way you work.

---

AI-Native Architecture

The Go backend collects market data, computes quantitative baselines, organizes evidence, and runs business workflows. A shared agent service selects Hermes or Codex from the application settings. Data services and AI reasoning are separate dependencies: quantitative previews need no model call, while deep stock research builds questions, retrieves evidence through Go, generates judgments and conditions, and validates references.

Layer	Responsibility
AI research workbench	React + TypeScript: market overview, trend themes, limit-up ladder, tiered stock research, portfolio inspection and optimization, review digest, trading wisdom, and AI copilot
Local Go API	HTTP queries and background job polling; WebSocket quotes and AI events; local authentication, request validation, and request logging
Domain services and orchestration	Theme fusion, ladder and sentiment calculations, inflection evaluation, stock research, portfolio report reuse and scoring, candidate screening and optimization review, viewpoint consensus, next-day verification, and scheduled sync
Data and local evidence	Source adapters and fallbacks, unified market models, theme attribution, versioned snapshots with source and time metadata; SQLite stores reviews, research jobs, portfolio inspections and optimization checkpoints, sentiment history, and theme caches
Shared agent service	Shared model profiles, reasoning settings, Skills, and MCP; Hermes or Codex selection; task-bound configuration, session resumption, approvals, clarification, and token usage. Codex uses its native App Server and requires a Responses connection
Electron desktop host	Bundled Go / Hermes / Codex / Python; local ports and startup tokens; Preload / IPC; Xueqiu and TaoGuba Browser Bridges, agent-browser fallback, and WeChat link parsing; logs, updates, and backups before installation
External ecosystem	Market and research data sources, content platforms, public review feeds and trading wisdom, plus model providers or compatible endpoints called directly by the selected runtime

Stock research saves stage checkpoints and can resume when model configuration and evidence still match. Portfolio inspection first reuses successful stock reports completed within 24 hours, then fills missing research before generating a portfolio report. Portfolio optimization saves stock judgments, investment comparisons, and review checkpoints so recovery can continue missing work. See the current architecture and code entry points (Chinese).

---

Getting Started

Users

No Node.js, Go, or Python required. Grab the installer or archive for your OS from GitHub Releases.

For first-time setup (LLM API keys, Xueqiu/TaoGuba login for content sync), see the User Guide (Chinese). To add or enable local skills, see the Skill Installation Guide (Chinese).

Developers

To run, debug, test, or build from source, see the Developer Guide (Chinese — architecture diagrams and code are language-neutral).

Configure Feishu and DingTalk custom group bots under Settings → Notifications, with signatures, keywords, and test messages. See Notification setup (Chinese) for setup and supported events.

Community & Contributing

• Discussions, research methods, and use cases: GitHub Discussions
• Bugs and data issues: open an issue with your OS, app version, and reproduction steps
• Feature requests: describe the scenario and the outcome you expect
• Code and docs contributions: start with the Contributing Guide and the Roadmap
• Chinese-speaking chat: QQ group 422158208
• Security issues: follow the Security Policy and report privately

---

License

The original backend, frontend, desktop app, and documentation of easy-stock are licensed under the PolyForm Noncommercial License 1.0.0.

• Personal learning, research, experimentation, and other non-commercial use, modification, and distribution are allowed, provided the license and copyright notice are retained.
• Any direct or indirect commercial use — including production deployment, paid services, SaaS, paid consulting or training, integration into commercial products, or resale — requires separate written permission from the author.
• Third-party dependencies, data sources, and bundled materials remain under their original licenses and terms of service.
• easy-stock is source-available software, not OSI open source.

---

Disclaimer

This project is for learning, research, and information organization only. It is not investment advice, a promise of returns, or a trading basis. Markets carry risk, and AI output and third-party data may be delayed, incomplete, or wrong — always verify against original sources, judge independently, and own your decisions.

_{Local first · Evidence based · Human in control}