Research workbench · MCP● Releasedv2.2.1 · 2026-10-11 · ★ 2,657

Vibe Research

A research workbench for China A-shares, US and Hong Kong stocks. 168 data endpoints, six-stage company research, backtesting and portfolio checks run on your own computer. Through MCP, Claude Code, Codex or Cursor can fetch data, start research and read reports and evidence directly.

For Independent investors who want their own agent to fetch data and run research, and want the numbers in a report to trace back to their sources.

Runs on Windows 11, macOS and Linux with Node.js 22.18+ (24 LTS recommended), Python 3.11+ (3.12 recommended) and Git. Portable zips for Windows x64, Apple silicon and Intel Macs are also available; they are unsigned and have not been validated on physical machines. Remote cloud mode supports macOS and Linux only. Setup and verification scope in the README ↗

Vibe Research dark home: indices, watchlist, holdings and the chat area
From the v2.1.0 repository · demo workspace, 10 October 2026 · watchlist and holdings are example data, not a real portfolio
Vibe Research first-run Connect AI dialog: Codex, Claude, WorkBuddy CLI and other API options
From the v2.1.0 repository · the dialog shows options only and is not evidence that any account passed a connection test
Vibe Research six-stage report: numbers carry evidence and calculation IDs, with the evidence appendix and fetch ledger on the right
From the v2.1.0 repository · a development test report from 8 October 2026 shown unchanged to illustrate the evidence structure; not investment advice

The work

What it keeps together

  • Keep indices, watchlist, holdings, daily review, news and research in one local workbench
  • Six-stage company research keeps sources and calculations: raw data carries its origin, derived numbers carry their formula
  • Give your own agent data, research status, reports and evidence through MCP

One task through the system

  1. 01

    Connect

    Add the MCP server in Claude Code, Codex or Cursor, or choose a subscription or model API under Connect AI. A setting is saved only after a successful test.

  2. 02

    Fetch

    The orchestrator fetches data through controlled scripts, saving raw responses with hashes. Research stages can call only five controlled tools.

  3. 03

    Research

    Six stages: company profile, financials, earnings expectations, valuation, risks and report. Missing data is recorded as a gap, never as zero.

  4. 04

    Deliver

    Numbers in the report carry evidence or calculation IDs, and the evidence appendix traces back to raw data. A number with no traceable source is replaced with an unverified marker and the run is marked incomplete. Records stay on your computer.

The workspace

Eight sections in one local workbench.

01

Home

Market and global indices, watchlist and holdings P&L first, then a conversation with AI.

02

Daily review

Breadth, volume, short-term sentiment, index trend position, margin balances and fund flows. Blocks that fail to load say why instead of showing zero.

03

Industry news

News radar and industry signals: news, filings and event leads, plus supply-chain signals such as GPU rental rates and monthly industry data.

04

Industry research

Twenty sectors, each with step-by-step research prompts. Your AI researches the content; the product does not prefill it.

05

Research tools

Company deep dive (six-stage research for A-shares, Hong Kong and US stocks), bull/bear debate (A-shares only for now), backtesting and a candidate screen.

06

My stocks

Watchlist and holdings, an event calendar and after-close reminders. Holdings come from a trade ledger; there is no order placement.

07

Library

Import your own research documents and ask questions with citations back to the source; research, backtest and debate reports are kept by date.

08

Connect AI

Choose a subscription or model API. Once the test passes, the whole workbench uses it.

MCP

Connect it to your agent.

Run this in the repository folder after setup. For Codex use codex mcp add with the same arguments; Cursor setup is in the README.

claude mcp add vibe-research -- node "$(pwd)/orchestrator/src/mcp.ts"
  • Eight tools: seven read-only (endpoint list, data fetch, research status, report, evidence, research list and knowledge files) and one that starts research, currently A-shares only.
  • Fetching data and reading reports or evidence needs no AI. Starting research uses this computer's default background AI source and consumes that provider's quota.
  • The MCP server runs on your computer after Vibe Research is installed. It is not a cloud API. Results are the same ones you see in the web workbench.

Connect AI

Use an existing subscription or a compatible API.

Claude Code subscription
Install Claude Code locally and sign in with a claude.ai subscription. API key, Bedrock and Vertex sign-ins do not count as a subscription.
Codex subscription
Install the official Codex CLI (0.160.0 or later) and sign in with a ChatGPT account. Codex signed in with an API key is not accepted.
WorkBuddy / CodeBuddy
Use an installed, signed-in WorkBuddy desktop app or Tencent's official CodeBuddy CLI.
Model API
Any service compatible with Chat Completions and tool calling, such as DeepSeek, Qwen, Zhipu GLM, Kimi, MiMo or OpenAI. Built-in templates have not been tested with real keys; your own connection test decides.

Current product

What exists today

  1. 01v2.2.1 was released on 11 October 2026 with source code and Windows / macOS portable builds
  2. 02A local MCP server for Claude Code, Codex and Cursor: eight tools, seven read-only and one that starts research (A-shares for now)
  3. 03A data registry of 168 endpoints in 35 layers covering quotes, financials, consensus estimates, filings, research, fund flows, macro data and more
  4. 04Research stages can call only five controlled tools, with the subscription CLI's built-in tools switched off; numbers in chat without a traceable source are withheld
  5. 05Six-stage research covers A-shares, Hong Kong and US stocks; one full Hong Kong run and one full US run have been completed with a Claude Code subscription

Not done yet

Still open in this version

  • Built-in model API templates have not been tested with real keys and are labelled as untested in the interface
  • Windows has only passed selected automated tests, not validation on a physical machine; portable builds are unsigned
  • Starting research through MCP and bull/bear debate support A-shares only for now
  • The Hong Kong test report still has three numbers without a cited source; other AI sources have not yet passed hard tests on Hong Kong or US research

Boundary

What it will not do

  • No stock picks
  • No price predictions
  • No entry or exit timing
  • No order placement, and no position sizing, target prices or stop levels in the output
Back to products