socrata-mcp
An MCP server that gives LLM agents typed, cached access to civic open-data portals via Socrata (SODA 2.1 + Discovery API), enabling search, query, profiling, sampling, and CSV export of datasets.
README
socrata-mcp
An MCP server that gives LLM agents typed, cached access to civic open-data portals. Speaks Socrata (SODA 2.1 + Discovery API) today; the provider layer is a thin interface so CKAN can be added later without touching the tool surface.
Highlights:
- Server-side profiling — null rates, distinct counts, min/max, top values computed via aggregate SoQL; the dataset is never downloaded.
- Hard row caps with honest truncation — every query result carries a
truncatedflag; paging uses a stable:idorder. - Disk cache under
~/.socrata-mcp/cachekeyed by query hash, with short TTLs for metadata and configurable TTLs for query results. - Polite by default — request throttling, retries with backoff that honor
Retry-After, optionalSOCRATA_APP_TOKENsent asX-App-Token. - Loud failures — the portal's actual error message is surfaced to the agent, never swallowed.
- Tableau-ready CSV export — streamed, paged download designed to chain
into vizforge's
csv_to_dashboard.
Install
git clone <this repo> && cd socrata-mcp
python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"
Register with your MCP client (see .mcp.json.example):
{
"mcpServers": {
"socrata": {
"command": "/absolute/path/to/socrata-mcp/.venv/bin/python",
"args": ["-m", "socrata_mcp"],
"env": { "SOCRATA_APP_TOKEN": "optional-app-token" }
}
}
}
Tools
| Tool | What it does |
|---|---|
search_datasets(query, domain?, category?, limit?, offset?) |
Full-text catalog search via the Socrata Discovery API. |
get_dataset(domain, dataset_id) |
Columns with types, row count, update cadence, license, attribution. |
query(domain, dataset_id, …) |
Structured SoQL (select/where/group/order/limit/offset, within_circle, within_box) or raw soql. Validated, paged, row-capped, truncated flag. |
profile_dataset(domain, dataset_id) |
Per-column null rates, distinct counts, min/max for dates/numbers, top values for categoricals — all portal-side. |
sample(domain, dataset_id, n?) |
First n rows (capped at 100) to see real values. |
export_csv(domain, dataset_id, out_path, …) |
Streamed, paged CSV export of any query. |
Example agent flow:
search_datasets("crime", domain="data.seattle.gov")
get_dataset("data.seattle.gov", "tazs-3rd5")
profile_dataset("data.seattle.gov", "tazs-3rd5")
query("data.seattle.gov", "tazs-3rd5",
where="offense_date > '2026-06-10T00:00:00'",
order="offense_date DESC", limit=100)
export_csv("data.seattle.gov", "tazs-3rd5", "out/spd_30d.csv",
where="offense_date > '2026-06-10T00:00:00'")
# → vizforge: csv_to_dashboard("out/spd_30d.csv", ...)
Configuration
All optional, via environment variables:
| Variable | Default | Meaning |
|---|---|---|
SOCRATA_APP_TOKEN |
unset | Sent as X-App-Token; raises portal rate limits. |
SOCRATA_MCP_CACHE_DIR |
~/.socrata-mcp/cache |
Disk cache root. |
SOCRATA_MCP_METADATA_TTL |
300 |
Seconds to cache dataset metadata. |
SOCRATA_MCP_CATALOG_TTL |
300 |
Seconds to cache catalog searches. |
SOCRATA_MCP_QUERY_TTL |
3600 |
Seconds to cache query/profile results (0 disables). |
SOCRATA_MCP_DEFAULT_LIMIT |
100 |
Rows returned when a query gives no limit. |
SOCRATA_MCP_MAX_ROWS |
5000 |
Hard row cap for inline query results. |
SOCRATA_MCP_MAX_EXPORT_ROWS |
1000000 |
Hard row cap for CSV exports. |
SOCRATA_MCP_PAGE_SIZE |
1000 |
Rows fetched per HTTP request. |
SOCRATA_MCP_THROTTLE_INTERVAL |
0.2 |
Minimum seconds between portal requests. |
SOCRATA_MCP_TIMEOUT |
30 |
Per-request timeout in seconds. |
Cache layout: cache/<kind>/<sha256>.json (kind ∈ catalog, metadata, query,
profile), each file {"cached_at": <epoch>, "data": …}. Deleting the directory
is always safe.
Notes:
- Discovery searches use the US endpoint (
api.us.socrata.com); EU-hosted portals are still directly queryable viaget_dataset/queryon their domain. - Raw
soqlexports run as a single request, so give them an explicitLIMIT(default cap 50,000); structured exports page automatically.
Development
.venv/bin/pytest # unit tests (all HTTP mocked)
.venv/bin/pytest -m network # live smoke tests against data.seattle.gov
Architecture: deterministic core (soql.py, cache.py, http_client.py,
profile.py, export.py) with the MCP layer (socrata_mcp/mcp/) as thin
wrappers over a Provider interface (providers/base.py). To add CKAN,
implement Provider in providers/ckan.py — the tool surface stays unchanged.
License
MIT — see LICENSE.
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。