lava-mcp

# lava-mcp

Gives AI agents and quant tools read-only data on how much news attention China A-share stocks and concept sectors get, through a remote MCP server or REST. For stock researchers and quant developers.

Made by dustinliang666-blip on GitHub. These notes are Mindket's, written from its README; the project's own files and license are in its repository.

## About

What it is: lava-mcp connects an AI agent or a quant tool to Lava, a service that tracks how often institutions and influential commentators mention China A-share stocks and concept sectors. The data is structured and read-only, and it arrives through a remote MCP server or a REST API. News hits come without titles or body text.

Who it suits: It suits people who research the A-share market with AI agents, and developers of quantitative trading tools that need hourly or daily measures of news attention.

```table
Industry | Who uses it | How they use it | Example
Quantitative trading | Quant developers | Pull daily heat series and coverage counts for backtests and exports | Read a stock's daily mentions, institution counts and weighted heat over a period of up to 92 days
Investment research | Analysts who work with AI agents | Ask an agent which stocks and concepts are drawing more attention | Rank stocks by how fast their heat is rising over a 24-hour window
Market monitoring | Traders who watch for new themes | Look for stocks mentioned for the first time in a while | List stocks with no mention in the previous 7 days that appear in the current window
Agent building | Developers of stock research agents | Connect an agent to a hosted endpoint instead of building a data pipeline | An agent calls a tool and gets counts, key points and anonymous institution IDs
```

People who need news titles or article text, or who do not follow China A-shares, can skip it.

Where to use it: Any client that can add a custom remote MCP server can use it, and plain REST calls work too.

- Clients that support custom remote MCP servers over Streamable HTTP
- WorkBuddy China edition 5.9.3, tested with browser OAuth and with the get_account tool, so no key typing
- REST calls from your own quant code
- Lava's own site, which has dashboards over the same statistics

```table
Use case | Who | What they get back
Rank hot stocks | A quant developer | Stocks ordered by mentions and people, over windows from 1 hour to 7 days
Spot rising attention | A trader | Stocks ranked by 1-day change, 3-day change or z20
Check institution coverage | A researcher | Counts of covering institutions over time, with no ID lists
Test a new connection | Anyone setting up | Account balance and usage, which is not billed
```

How to set it up: You need a Lava account and a client that supports custom remote MCP servers.

- A Lava account, created with a phone number or Google
- An API key from the AI Data Service page, unless your client signs in with browser OAuth
- A client that supports custom remote MCP servers

1. Sign in to Lava. In the left menu, open AI Data Service, then API Key. The first visit gives you a voucher and creates a key named default.
2. In your client, add a custom remote MCP server. Use the endpoint listed in the project's README, with the Streamable HTTP transport.
3. Send your key in the request header shown below. The README also allows X-API-Key as an alternative.
4. Enable the connector and look at the tool list. Make get_account your first call, because it is not billed. Then query data.

```text
Authorization: Bearer <API Key>
```

How to use it: Ask your agent about a stock or a concept in plain words, and it calls the tools below. The agent can also use get_stock_heat_windows for hourly windows, get_news_stock_hits for news hits, and four agency tools for institution scores.

```table
Tool | What it does
get_account | Shows balance and usage, and is not billed
get_stock_news_heat | Shows a stock's news heat with a list of structured points
search_news_flow | Searches the structured news records by keyword
get_concept_stock_heat | Ranks the stocks mentioned under a concept
get_stock_concept_heat | Ranks the concepts mentioned for a stock
get_market_hot_stocks | Ranks hot stocks across the market, with 1-day change, 3-day change and z20
get_market_hot_concepts | Ranks hot concepts across the market, with windows up to 3d
get_weighted_hot_stocks | Ranks stocks with weight given to high-scoring institutions, with a default minimum score of 1050
get_first_mention_stocks | Lists A-shares mentioned for the first time after 7 days without a mention
get_accelerating_hot_stocks | Ranks stocks by how fast their heat is changing
get_stock_heat_panel | Gives a daily history of mentions, agencies and weighted heat, for backtests and exports
get_agency_coverage | Gives counts of covering institutions over time, and only counts
```

For get_accelerating_hot_stocks, the window can be 1h, 4h, 18h, 24h, 3d or 7d, with 24h as the default. The order can be chg1d, chg3d or z20. The default minimum is 3 mentions, and the limit defaults to 50 with a maximum of 200.

Example: You ask, "Which A-share stocks have the fastest rising news heat over the last 24 hours?" The agent calls get_accelerating_hot_stocks with its defaults. It returns up to 50 stocks, ranked by 1-day change, with plain statistical fields.

For newcomers and experts: Lava counts how often institutions and commentators mention each stock, then lets your agent read those counts. It gives attention data and short key points, not advice. A good first try is to connect, call get_account, then ask for one stock's heat.

Under the hood, heat is built from mention counts and institution coverage, and some tools weight it by institution score. Institutions appear only as anonymous agency_id values. Scores run from about 450 to 1500, with 1000 near the middle and 1050 near the top 10%. Momentum fields such as heat_chg_1d, heat_chg_3d and heat_z20 are plain statistics.

```flow
Your client calls the hosted endpoint with your key
Lava reads its structured news heat data
A tool returns counts, key points and agency IDs
News hits carry no titles or body text
```

Good to know:

- All data is read-only, and news hits give only the news ID, the media, and the stocks hit
- Signing up gives a 100 yuan voucher, and about 1000 calls cost 5 yuan
- get_stock_heat_panel covers at most 92 days and 20 targets in one call
- Sector trends are live on Lava's site, but their subscription API is still planned
- Sentiment data is planned, and it will give aggregate statistics only
- The repository holds no local proxy for forwarding calls, so use the hosted endpoint only
- Anyone who already ran such a proxy should replace the API key from the Lava data console

## On GitHub, 7 November 2026

- Stars: 65
- Forks: 0
- Open issues and pull requests: 1
- License: MIT
- Topics: a-share, agent-skills, claude-skills, dsh-plugin, gemini-cli-extension, mcp, mcp-server, quant

## Links

- Code: https://github.com/dustinliang666-blip/lava-mcp
- Website: https://lava.ai-dolphin.com/
- Releases: https://github.com/dustinliang666-blip/lava-mcp/releases
- Issues: https://github.com/dustinliang666-blip/lava-mcp/issues
- Star history: https://www.star-history.com/#dustinliang666-blip/lava-mcp&Date

## Get the code

```
git clone https://github.com/dustinliang666-blip/lava-mcp.git
```

Then follow the setup in its README. Nothing in this kit runs by itself.

README.md  7 KB