What Lava (熔岩) — A-share news-flow heat MCP does
Lava exposes structured read-only data from lava.ai-dolphin.com for China A-share stocks and concept sectors. Its data describes how often stocks, concepts, institutions, and large-account sources are mentioned, along with summarized key points and source identifiers. Institution and KOL records use anonymous agency_id values rather than names.
The service is suited to agent-based market research and quantitative workflows that need measurable news-flow activity. It does not provide trading mutations. News-hit results contain identifiers, timestamps, media, and matched stocks, but exclude article titles and full text.
How it works
The recommended deployment uses the remote Streamable HTTP MCP endpoint:
https://lava.ai-dolphin.com/subscribe/v1/mcp
Requests can authenticate with Authorization: Bearer <API Key> or X-API-Key. OAuth 2.1 with PKCE S256 and dynamic client registration is also supported. The same service is available through REST interfaces documented alongside its OpenAPI specification.
For clients that need a local endpoint or cannot attach request headers directly, the repository includes a Python 3.9+ proxy using only the standard library. The proxy forwards MCP Streamable HTTP and JSON-RPC requests to the official endpoint; it does not implement the market-data queries itself. By default, it listens on 127.0.0.1:8787/mcp and forwards results from the remote Lava service.
Setup and configuration
Create an API key through Lava’s AI Data Service, then configure the remote MCP URL in a client that supports custom HTTP MCP connections. For the local proxy, run:
export LAVA_API_KEY=lvk_your_key
python3 proxy/server.py
The proxy accepts these settings:
LAVA_API_KEY: API key added to forwarded requests.LAVA_PROXY_PORT: listening port; defaults to 8787.LAVA_PROXY_HOST: bind address; defaults to 127.0.0.1.LAVA_PROXY_ALLOW_REMOTE: required when binding beyond the local machine.LAVA_MCP_UPSTREAM: upstream MCP URL; defaults to the official endpoint.LAVA_PROXY_TIMEOUT: upstream timeout in seconds; defaults to 180, while 0 disables the limit.
Connect a client to http://127.0.0.1:8787/mcp using Streamable HTTP. The proxy’s /health endpoint checks only the local process, not upstream availability. Local tests can be run with python3 -m unittest proxy.test_server.
Tools and capabilities
Available MCP tools include:
get_stock_news_heat for stock heat and structured viewpoint lists.search_news_flow for keyword searches across structured news-flow records.get_market_hot_stocks and get_market_hot_concepts for market-wide rankings.get_weighted_hot_stocks, get_accelerating_hot_stocks, and get_first_mention_stocks for source-weighted, momentum-oriented, and first-mention views.get_stock_heat_windows for multiple time windows, hourly data, and source breakdowns.get_concept_stock_heat and get_stock_concept_heat for concept-to-stock and stock-to-concept rankings.get_stock_heat_panel for daily historical mentions, agency counts, weighted heat, and first-mention-day flags.get_agency_coverage for rolling coverage counts and new-coverage statistics.get_agency_ranking, get_agency_pit, get_agency_history, and list_agency_pit_dates for current, point-in-time, and historical agency scoring.get_news_stock_hits for news identifiers, publication details, and matched stocks.get_account for non-billable balance and usage information.
The historical panel supports stock or concept data for up to 92 days and up to 20 symbols per request. Agency coverage returns counts rather than agency ID lists.
Limitations and notes
All query data is read-only. The service focuses on structured A-share news-flow statistics, not article retrieval: news-hit responses omit titles and bodies. Agency and KOL identities are represented by agency_id. Momentum fields such as one-day change, three-day change, and z20 are statistical measures, not investment recommendations.
Concept and stock heat, rankings, historical panels, coverage, and agency scoring are available through the subscription interfaces. Sector-trend data is available on the Lava website but its subscription interface is still planned, while aggregated sentiment features are also planned. The repository’s runnable server-side component is the forwarding proxy, so market data remains dependent on the official Lava endpoint and an active subscription/API key.