nycdata

A developer-first CLI and MCP server for discovering, inspecting, querying, and profiling live NYC Open Data.

nycdata sits above NYC's Socrata APIs. It does not mirror datasets or replace NYC as the source of truth. It makes the path from project idea to verified API integration faster and less error-prone.

Repository: https://github.com/HT224/nycdata

Status

Early MVP (0.2.0). Implemented CLI commands:

• search

• describe

• query

• profile

The MCP server exposes the same capabilities to coding agents. Planned next: generated types, saved queries, and deterministic client generation.

Related MCP server: ontario-data-mcp

Requirements

• Node.js 20+

• No credentials required for normal public reads

• Optional Socrata application token for higher rate limits

Install locally

npm install
npm run build
npm link
nycdata --help
nycdata-mcp

For development without linking:

npm run dev -- search "restaurant inspections"

Optional token

export SOCRATA_APP_TOKEN="your-token"

The token is sent through the X-App-Token header. It is never added to URLs or generated output.

Discover datasets

nycdata search "restaurant inspections"
nycdata search "motor vehicle crashes" --limit 5
nycdata search "311 complaints" --format json

Inspect schema and metadata

nycdata describe 43nn-pn8j
nycdata describe erm2-nwe9 --format json

Query live data

nycdata query 43nn-pn8j --limit 5

nycdata query 43nn-pn8j \
  --select "camis,dba,boro,grade,grade_date" \
  --where "boro = 'Brooklyn' AND grade = 'A'" \
  --order "grade_date DESC" \
  --limit 10

nycdata query 43nn-pn8j \
  --near "40.6895,-73.9724" \
  --radius 1000 \
  --location-field location \
  --limit 10 \
  --show-url

Formats:

nycdata query h9gi-nx95 --limit 10 --format table
nycdata query h9gi-nx95 --limit 10 --format json
nycdata query h9gi-nx95 --limit 10 --format csv

Profile before building

nycdata profile 43nn-pn8j
nycdata profile erm2-nwe9 --sample-size 250 --format json

Profiles are intentionally bounded. Row counts are queried live; null rates and cardinality are explicitly sample-derived so the CLI does not download an entire dataset by accident.

MCP server

nycdata-mcp runs a local stdio MCP server with four read-only tools:

• search_datasets

• describe_dataset

• query_dataset

• profile_dataset

Each tool uses the same live NYC Open Data client as the CLI. Query results are capped at 1,000 rows per call, profile samples are capped at 1,000 rows, and proximity radii are capped at 50 km.

After building the package, configure an MCP client to run:

{
  "mcpServers": {
    "nycdata": {
      "command": "node",
      "args": ["/absolute/path/to/nycdata/dist/mcp.js"],
      "env": {
        "SOCRATA_APP_TOKEN": ""
      }
    }
  }
}

The token entry is optional. The server writes protocol messages to stdout and diagnostics to stderr.

Agent workflow

An agent can:

1. search for candidate datasets;

2. inspect the chosen dataset's actual fields and row semantics;

3. run a small live query;

4. profile data quality;

5. build application code from verified results instead of guessed schemas.

Architecture

The CLI is a thin presentation layer over reusable TypeScript modules:

• client.ts — catalog, metadata, and SODA requests

• query.ts — dataset validation and SoQL URL construction

• profile.ts — bounded data-quality analysis

• format.ts — table, JSON, and CSV output

• mcp.ts — stdio MCP transport and safe tool schemas

• index.ts — public core exports for future CLI/MCP/scaffold consumers

See plan.md for scope, principles, milestones, and deferred work.

Development

npm run check
npm run test:live-mcp

check runs lint, TypeScript checks, unit/in-memory MCP tests, and the production build. test:live-mcp launches the built stdio server through a real MCP client and calls all four tools against live NYC data.

Data source

• Portal: https://opendata.cityofnewyork.us/

• Catalog: https://data.cityofnewyork.us/

• API documentation: https://dev.socrata.com/

Dataset schemas and update patterns vary by agency. Always inspect and profile a dataset before treating rows as canonical entities.

License

MIT