Four open-source products · Vibe Research and Vibe AStock ship MCP servers

Financial research toolsbuilt for your agent.

Add our MCP servers and data skills in Claude Code or Codex, and your agent can fetch A-share, US and Hong Kong market data and read reports and evidence directly. Starting research over MCP is A-shares only for now. Reports retain source citations and calculation records for line-by-line review.

phoenixtree · agent_link

PRODUCTS

Four products in two layers, built for your agent

The research systems fetch data, calculate and check the evidence. The data skills give your agent quotes, financials and filings directly.

Where to start

If you want toStart withHow it connects
Have your agent pull A-share quotes, financials, filings and research reports A-Stock Data Load one skill file
Have your agent pull US market data, plus Hong Kong quotes and financials Global Stock Data Load one skill file
Research a company (A-shares, US or Hong Kong) and keep the report and evidence Vibe Research Local workbench; your agent connects over MCP, and starting research over MCP is A-shares only for now
Review the A-share short-term market each trading day and look back on your own trades Vibe AStock Local workbench; your agent connects over MCP, starting research over MCP is A-shares only for now, and MCP does not read your trade journal or holdings

Pick the one that fits the task. You don't need all four.

01 · SYSTEMS

Research systems

Full local workbenches your agent connects to for research and reviews.

● Released
Research workbench · MCP

Vibe Research

A research workbench for A-shares, US and Hong Kong stocks: 168 data endpoints, six-stage company research, backtesting and portfolio checks. Connect it to your agent through MCP to fetch data, start research and read reports and evidence.

v2.2.1 · 2026-10-11 · ★ 2,657 MCP · Claude Code / Codex / Cursor (research runs: A-shares for now)Windows · macOS · Linux · portable builds available
● Released
A-share short-term review · MCP

Vibe AStock

A short-term review and tracking workbench for China A-shares: live watch, sentiment review, AI reviews checked against the next close, a trade journal and 182 data endpoints. Connect it to your agent through MCP to fetch short-term data, start research and read reports and evidence.

v2.1.0 · 2026-10-11 · ★ 678 MCP · Claude Code / Codex / Cursor (research runs: A-shares for now)Windows · macOS · Linux · portable builds available
02 · DATA SKILLS

Data skills

One skill file your agent loads to fetch data directly. We ship code, not data: you fetch data on your own computer under each source's terms.

● Released
A-share data skill

A-Stock Data

A full-stack A-share data skill: 15 layers, 87 endpoints and 34 sources covering quotes, research, fund flows, filings, limit-up pools, macro, futures and convertibles. Load one skill file into Claude Code or Codex and your agent fetches the data directly.

v3.10.2 · 2026-10-11 · ★ 10,784 Skill file · Claude Code / CodexPython 3.9+ · no keys needed except iwencai
● Released
US market data skill

Global Stock Data

A full-stack US market data skill: 13 layers, 30+ endpoints and 11 sources, including CBOE options Greeks and 0DTE flow, FINRA short volume, the SEC filing stream and a market-wide screener, plus Hong Kong quotes and financials. Every source is labelled with whether commercial use is allowed.

v2.0.3 · 2026-07-26 · ★ 1,723 Skill file · Claude Code / CodexPython 3.9+ · requests only · no keys needed

HOW IT WORKS

One call, four steps

From the question your agent asks to a result you can check.

  1. 01 · CONNECT

    Connect your agent

    Add the MCP server in Claude Code, Codex or Cursor, or drop a data skill into your skills folder. You can also connect your own subscription or model API inside the workbench; a setting is saved only after its test passes.

  2. 02 · SCOPE

    Controlled tools only

    Each research stage is a separate session that can call only the product's five controlled tools. The subscription CLI's built-in tools are switched off.

  3. 03 · COMPUTE

    Computed by code, citations checked

    Quotes, metrics and backtests are computed by code, not estimated by a model. Citations in reports are checked.

  4. 04 · DELIVER

    Results you can check

    You get the report, its sources and the calculations. You and your agent can go back and check every step.

Every call, visible

The trail one research task leaves in the system. This is an illustration with example data.

TOOL SCANControlled tools
● ACTIVE
Scope · what this task needs
Call log · example
CHKREFACTIONSTATUSTIME
R-01 Read three years of filings for two companies VERIFIED 09:41:02
R-02 Search related announcements and inquiry letters VERIFIED 09:41:05
R-03 Align operating cash flow definitions ◐ RUNNING 09:41:09
A-01 Yesterday's limit-up cohort · two-board filter DONE 09:41:12
A-02 Check review evidence ◐ RUNNING 09:41:15
VIBE RESEARCH · six-stage research
6/6STAGE
● COMPLETE
Current task · example VIBE ASTOCK
Generate today's A-share review Confirm the date and sample, fetch quotes and compute metrics, then check the evidence and write the report.
Done100%

OPEN SOURCE

Fully open source: six public repositories

All four products and related projects are public on GitHub. Read the code, open an issue or deploy it yourself.

● GITHUBPUBLIC
Total stars★ 19,983
  • a-stock-data★ 10,784

    A-share quotes, filings, research and news data.

  • An A-share adaptation of TauricResearch's TradingAgents multi-agent framework.

  • Vibe-Research★ 2,657

    Research workbench for China, US and Hong Kong stocks, with a local MCP server.

  • A data skill for US markets, with Hong Kong coverage.

  • Short-term review and tracking workbench for China A-shares, with a local MCP server.

  • Turn public market information into research leads that can be followed over time.

Data as of 2026-10-11

SECURITY

Boundaries built into the system

The products run on your computer, with clear limits on tool access and on what the output says.

Full security boundaries →

Runs locally

All four products run on your computer. The research systems' local services listen only on your own machine by default, and the data skills fetch from each source directly, never through our servers.

Scoped tools

Research stages can use only the product's controlled tools. Each research system's MCP server has eight tools: seven read-only and one that starts research.

No overreach

No stock picks, no price predictions, no trade timing.

GET STARTED

A few commands to run it locally

Research systems: download the source, set up the environment, then add the MCP server to Claude Code or Codex. You can also start the web workbench and connect your subscription or model API under Connect AI. Data skills: put the skill file in your skills folder and install the dependencies. The terminal shows macOS / Linux commands; see each repository's README for Windows.

Codex subscriptionClaude CodeWorkBuddy / CodeBuddyModel API
git clone https://github.com/simonlin1212/Vibe-Research.git vibe-research-agent
cd vibe-research-agent
scripts/setup
# Add it to Claude Code (for Codex, run codex mcp add with the same arguments)
claude mcp add vibe-research -- node "$(pwd)/orchestrator/src/mcp.ts"
→ Ask your agent: use vibe-research to fetch a year of daily bars and key financials for Tencent (00700)
# For the web workbench: scripts/start. Portable Windows / macOS builds skip the setup

Connect your agent and start with one question.

The products are open source and free to use. Model costs are charged by the AI service you choose, under its own terms.