whatshot-mcp
MCP server that connects to a WhatsHot Backend over HTTP to provide hot topic discovery, trend analysis, and historical data tools such as current hotlists, trend series, and event analysis.
README
WhatsHot MCP
whatshot-mcp is the single open-source MCP server for WhatsHot. It connects
to either a local WhatsHot Backend or the hosted WhatsHot Backend through the
same versioned HTTP contract.
This repository contains the Contract v1 models and the MCP implementation. The MCP server only talks to a versioned WhatsHot Backend over HTTP; it never opens DuckDB or PostgreSQL itself.
The project is licensed under the MIT License and owned by alisen39.
Requirements
- Python 3.12 or newer
- One standard installation; no optional dependency extras are currently defined
The supported installation form is python -m pip install whatshot-mcp. No
[all] or other extras are defined.
For local development:
python -m pip install -e .
pytest
python -m build
Run
Copy config.example.toml, then start the primary Streamable HTTP transport:
whatshot-mcp serve --config ./config.toml
Clients connect to http://127.0.0.1:6691/mcp by default. Streamable HTTP is
the only supported transport.
The HTTP process also exposes a minimal public GET /health probe and a
deployment-level GET /ready probe. When static inbound authentication is
enabled, /ready requires the same Bearer token while /health remains public.
Operational commands:
whatshot-mcp config validate --config ./config.toml
whatshot-mcp backend check --config ./config.toml
whatshot-mcp version
backend check validates the capabilities envelope, Contract v1 and
boardKeyVersion. Its exit codes are 0 success, 2 invalid configuration,
3 Backend unavailable, 4 invalid/error Contract response and 5
incompatible board-key version.
The Backend API key may be stored as backend.api_key in the uncommitted local
configuration file, or resolved from backend.api_key_env; the environment
value takes precedence. The key is sent only as an Authorization: Bearer
request header and is never a tool argument.
There are two independent authentication boundaries:
- MCP client → MCP Server: Streamable HTTP may use the configured inbound static Bearer token.
- MCP Server → Backend
/api/v1: Core data endpoints are public, while Cloud data endpoints require the configured Backend Bearer token and enforce their declared scopes. The shared OpenAPI contract marks Bearer as optional and records the normative choice inx-whatshot-deployment-auth.
An inbound MCP token is never reused as a Backend token, and the Backend API key is never exposed as a tool argument. Supported environment overrides include:
WHATSHOT_MCP_SERVER_BIND
WHATSHOT_MCP_SERVER_PORT
WHATSHOT_MCP_SERVER_PATH
WHATSHOT_MCP_SERVER_AUTH_MODE
WHATSHOT_MCP_SERVER_TOKEN_ENV
WHATSHOT_MCP_SERVER_TOKEN
WHATSHOT_MCP_BACKEND_URL
WHATSHOT_MCP_BACKEND_API_KEY
WHATSHOT_MCP_BACKEND_TIMEOUT_SECONDS
WHATSHOT_MCP_BACKEND_CAPABILITIES_TTL_SECONDS
The first fixed Universal Tool Catalog contains:
whatshot_get_capabilitieswhatshot_list_sourceswhatshot_get_source_schemawhatshot_get_currentwhatshot_get_current_batchwhatshot_query_historywhatshot_search_historywhatshot_get_trend_serieswhatshot_get_data_coveragewhatshot_analyze_hot_eventwhatshot_analyze_newsflash_coverage
When the startup capabilities snapshot reports navigation=true, the same
open MCP package additionally registers these Cloud tools:
whatshot_list_navigation— cursor-paged category/site discoverywhatshot_fetch_category_hotlists— bounded current boards for one category
Both tools require the Backend navigation capability and data:read scope.
They are selected from capabilities, not a Backend-name check. A request with
freshness=live is still authorized by the Cloud Backend, which additionally
requires its live:fetch scope.
The analysis tools scan history/search pages through the same Backend
Contract. scanBudget limits evidence examined while evidenceLimit separately
limits evidence returned. Responses always report analysisComplete,
scannedCount, and coverage; lifecycle times are explicitly marked
approximate when the scan budget stops pagination. Analysis does not create or
update a research run.
At startup the MCP validates and freezes one deployment-level capabilities
snapshot. It registers only tools whose required Backend capability is enabled;
changing Backend features requires restarting the MCP. User permissions never
change tools/list: they are enforced by the Backend on each call, so every
user of one deployment sees the same publicly cacheable catalog.
Contract v1
The source of truth is the Pydantic model package at
src/whatshot_mcp/contracts/v1/.
- All public JSON fields serialize as
camelCase. - All models reject undeclared fields.
- All timestamps require an explicit timezone.
- Successful Backend responses use
{ "data": ..., "meta": ... }. - Failed Backend responses use the stable error envelope defined by
ErrorEnvelope. core-readandhistory-readare distinct capability profiles. A Backend may offer current data while history storage is disabled.boardKeyis generated by the documented canonical algorithm and must be identical in local and hosted Backends.
The contracts/ directory describes how OpenAPI, JSON Schema, and fixtures are
published without creating a second hand-maintained contract definition.
Configuration
config.example.toml records the runtime configuration boundary. A real Backend
key may be kept in the ignored local copy or supplied by environment variable.
Unauthenticated Streamable HTTP is restricted to loopback. A non-loopback bind
requires server.auth.mode = "static_token" and a non-empty token resolved from
server.auth.token_env. The incoming token is used only at the HTTP boundary;
it is never a tool argument or log field. oauth is reserved and currently
fails closed.
License
MIT License. See LICENSE.
Release
PyPI publishing uses GitHub Actions Trusted Publishing; the repository does not
store a long-lived PyPI token. After validation and a SemVer version update,
push the matching v<version> tag. The dedicated publish.yml workflow builds
the wheel and source distribution, then publishes them from the protected
pypi environment.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。