GridNews

市場新聞、新聞稿及股票代號情緒，透過REST、串流及MCP提供給AI代理。

文件

@gridnews/mcp

MCP server for GridNews — market news, press releases, and sentiment for AI agents.

Gives any MCP-capable assistant the latest market coverage: story clusters ranked by how many independent outlets corroborated them, article search across thousands of sources, per-ticker news with an aggregate sentiment reading, primary-source wire releases, and trending topic clusters. With a Pro key it can also run Grid AI research: read documents or search results and return verbatim, cited excerpts, with an optional comparison across them.

Install

Requires Node.js 20+ and a GridNews API key. get_topics works without a key; every other tool needs one.

Claude Code

claude mcp add gridnews --env GRIDNEWS_API_KEY=your_key -- npx -y @gridnews/mcp

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "gridnews": {
      "command": "npx",
      "args": ["-y", "@gridnews/mcp"],
      "env": { "GRIDNEWS_API_KEY": "your_key" }
    }
  }
}

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "gridnews": {
      "command": "npx",
      "args": ["-y", "@gridnews/mcp"],
      "env": { "GRIDNEWS_API_KEY": "your_key" }
    }
  }
}

OpenAI Codex

codex mcp add gridnews --env GRIDNEWS_API_KEY=your_key -- npx -y @gridnews/mcp

Gemini CLI

Add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "gridnews": {
      "command": "npx",
      "args": ["-y", "@gridnews/mcp"],
      "env": { "GRIDNEWS_API_KEY": "your_key" }
    }
  }
}

Configuration

Variable	Default	Purpose
`GRIDNEWS_API_KEY`	—	Your API key. Required for everything except `get_topics`.
`GRIDNEWS_BASE_URL`	`https://api.gridnews.io`	Override the API host.
`GRIDNEWS_TIMEOUT_MS`	`30000`	Per-request timeout. Raise it if you use `get_symbol_sentiment` heavily.
`GRIDNEWS_WS_URL`	`wss://stream.gridnews.io`	Override the stream host `research` and `get_research` connect to.

Tools

Tool	Purpose	Minimum tier
`get_top_events`	The biggest stories now, as clusters ranked by independent corroboration	free (filters need basic)
`get_event_detail`	Every outlet that carried one story, grouped by independent voice	free
`search_news`	Search articles and press releases by text, symbol, source, date, sentiment, and quality	free (filters need basic)
`get_symbol_news`	Recent coverage for one ticker, plus an aggregate sentiment reading	free
`get_symbol_sentiment`	Sentiment analysis for a ticker over a timeframe, with its basis	pro
`list_press_releases`	Wire releases filtered by symbol, provider, company, and date	free
`list_sources`	The sources GridNews indexes, with ids for the `sources` filter	free
`get_topics`	Trending topic clusters as keyword groups with article counts	none
`get_usage`	The key's tier, entitlements, and remaining daily quota	free
`research`	Read documents or search results and return cited excerpts; `deep` also compares them	pro
`get_research`	Collect a research request that was still running when `research` returned	pro

Corroboration, not outlet count

get_top_events returns stories rather than documents, and reports two different numbers:

sourcesCount — how many outlets carried the story. This is reach.
independent voices — how many of those were not republishing each other. This is the evidence.

They are usually not the same. Outlets that redistribute each other collapse into one voice, so a story on five outlets all running the same wire copy is one voice, not five. Where an outlet was carrying someone else's copy, the tool marks it inline:

2 independent voices across 4 outlets · 4 filings · impact 1.47
- Wall Street Journal: https://wsj.com/...
- Dow Jones [carrying wall-street-journal]: https://morningstar.com/...
- GuruFocus: https://gurufocus.com/...

A press release is always a single voice however many wires carried it — one issuer announcing itself is not confirmation.

Pass minVoices: 2 for corroborated stories only. There is no default: single-voice clusters are real distribution records and are not hidden, they just rank last.

Grid AI research

research reads up to 8 documents you name (urls), or searches the web or news and reads the results, and returns the passages that bear on the query, quoted verbatim and numbered for citation. depth: "deep" goes further: it compares the sources and answers up to 3 questions, marking which statements are the sources' and which are inference, and naming gaps and disagreements. A deep request needs at least 3 sources, or 2 different URLs; below that it would be one document paraphrased, which is not analysis.

# Grid AI research: Why was the export rule extended?
id sdk-4f1c... · deep · complete · 4 of 4 sources read · 38.2s

## Sources
[1] Commerce extends chip export rule
    reuters.com · 3h ago (2026-09-29T09:12:00Z, page date)
    https://reuters.com/...
    > The department said the extension gives ...

## Analysis: Why was the export rule extended?
Findings:
- Both reports tie the extension to the licensing review. [1][3]
- (inference) The timing suggests ... [2]

A quick request takes up to 90 seconds and a deep one up to 4 minutes. Many MCP clients give up on a tool call after about a minute, so research waits waitSeconds (default 50) and then returns what has arrived with the request id. The request keeps running, and get_research collects it. Results are kept for an hour. Clients that ask for progress get a notification per source, which lets those that reset their timeout on progress wait for the whole request.

Research runs over the GridNews stream WebSocket and uses one of the key's stream connections while it runs. Each key may run 2 requests at once on Pro and 4 on Business, and research counts against the key's Grid AI allowance.

Tier behaviour

GridNews gates features and history depth by tier. Rather than failing opaquely, the tools report what a call needs:

A 403 names the required tier and the key's current one, so the agent can retry without the gated parameter instead of giving up.
A 429 reports remaining quota and reset time, so the agent waits rather than looping.
Results note when a tier's history window excluded older articles, so a thin result isn't mistaken for an absence of coverage.

get_usage explains any of these on demand.

Development

npm install
npm run build
npm test

The test suite drives the built server over a real stdio MCP session against a stubbed API, so it covers protocol handshake, argument validation, HTTP serialization, formatting, and the tier/quota failure paths without needing a live key.

License

MIT