mcp-statcan

Statistics Canada MCP Server

MCP server and CLI for Statistics Canada's Web Data Service (WDS) and SDMX REST API. Gives any MCP client — Claude, Cursor, VS Code Copilot, Gemini — structured access to Canadian statistical data. Includes a standalone statcan CLI for direct downloads without an LLM.

Hosted on Render — no install required for most users. See Quick Start.

⚠️ LLMs may fabricate data. Always verify important figures against official Statistics Canada sources.

---

Table of Contents

• Quick Start
• Examples
• Setup by Client
• How Claude.ai Uses This Server
• MCP Prompts
• statcan CLI
• Features & Tools
• Project Structure
• Known Issues

---

Quick Start

Pick the option that fits you. You don't need to install anything for Option 1.

Option 1 — Use the hosted server (recommended)

Connect directly to the public server on Render. No uv, no terminal, no local setup.

Claude Desktop / Claude.ai

1. Open Settings → Connectors → Add Custom Connector
2. Name: mcp-statcan
3. URL: https://mcp-statcan.com/mcp
4. Save and restart

Claude Code

claude mcp add statcan --transport http https://mcp-statcan.com/mcp --scope global

The hosted server provides all WDS + SDMX tools. Database tools (SQLite) require local setup (Option 3) — they are intentionally excluded from the shared server.

---

Option 2 — Self-host HTTP (WDS + SDMX, no DB)

Run a local server with the same tools as the hosted version.

Step 1 — Install uv:

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Step 2 — Start the server:

uvx statcan-mcp-server --transport http
# Listening at http://localhost:8000

Step 3 — Connect your client to http://localhost:8000/mcp.

---

Option 3 — Full local setup (WDS + SDMX + SQLite)

Everything from Option 2, plus database tools for storing and querying data with SQL. Runs via stdio.

Step 1 — Install uv (same as above).

Step 2 — Configure your client with the stdio snippets in Setup by Client below.

uvx downloads and runs the server automatically on first use.

---

Option 4 — statcan CLI (no LLM needed)

Download StatCan data directly from the terminal. See statcan CLI.

uvx statcan-mcp-server        # installs the package
statcan search "labour force"
statcan metadata 14-10-0287-01                                    # run first to find dimension IDs
statcan download 14-10-0287-01 --key "1.1.1" --last 12 --output lfs.csv  # run `statcan metadata <pid>` first to find real dimension IDs
statcan describe                                                  # machine-readable command reference for agents

---

Examples

Chat examples

Dataset	Query	Demo	Source
Canada's Greenhouse Gas Emissions	"Create a simple visualization for greenhouse emissions for Canada as a whole over the last 4 years"	Chat	Table 38-10-0097-01
Canada's International Trade in Services	"Create a quick analysis for international trade in services for the last 6 months with a visualization"	Chat	Table 12-10-0144-01
Ontario Building Construction Price Index	"Generate a visualization for Ontario's Building Price index from Q4 2023 to Q4 2024"	Chat	Table 18-10-0289-01
Canadian Unemployment Dashboard	"Create a Canadian Unemployment Dashboard using statcan mcp"	Chat	Table 14-10-0287-01

Dashboard examples

Title	Link	Source
Canada's Critical Minerals Economy	Dashboard	Table 36-10-0708-01
Price of Everything: CPI Dashboard 2015–2026	Dashboard	Table 18-10-0004-01
Canada's Biomedical & Biotech Industries	Dashboard	Table 27-10-0297-01

---

Setup by Client

Hosted server (Option 1)

Claude Desktop — Settings → Connectors → Add Custom Connector

• Name: mcp-statcan
• URL: https://mcp-statcan.com/mcp

Claude Code

claude mcp add statcan --transport http https://mcp-statcan.com/mcp --scope global

Cursor — .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "statcan": {
      "url": "https://mcp-statcan.com/mcp"
    }
  }
}

VS Code (GitHub Copilot) — .vscode/mcp.json:

{
  "servers": {
    "statcan": {
      "type": "http",
      "url": "https://mcp-statcan.com/mcp"
    }
  }
}

---

Self-hosted HTTP (Option 2)

Start uvx statcan-mcp-server --transport http first, then configure your client.

Most clients need mcp-proxy to bridge stdio ↔ HTTP. Claude Code connects natively.

Claude Desktop — Settings → Developer → Edit Config:

{
  "mcpServers": {
    "statcan": {
      "command": "uvx",
      "args": ["mcp-proxy", "--transport", "streamablehttp", "http://localhost:8000/mcp"]
    }
  }
}

Claude Code

claude mcp add statcan --transport http http://localhost:8000/mcp --scope global

Cursor / VS Code / Gemini — same mcp-proxy wrapper, pointing to http://localhost:8000/mcp.

---

Full local / stdio (Option 3)

Claude Desktop — Settings → Developer → Edit Config:

{
  "mcpServers": {
    "statcan": {
      "command": "uvx",
      "args": ["statcan-mcp-server", "--db-path", "/Users/<you>/.statcan-mcp/statcan_data.db"]
    }
  }
}

Pass --db-path with an absolute path. Claude Desktop overrides the subprocess HOME env var, which can break default path resolution.

Claude Code

claude mcp add statcan --scope global -- uvx statcan-mcp-server

Cursor / VS Code / Gemini — use uvx statcan-mcp-server as the stdio command.

---

How Claude.ai Uses This Server

Claude.ai (web) has no bash sandbox — it can't run shell commands. Instead, it uses MCP tools for discovery and its Python script tool to fetch data without bloating the context window.

The pattern:

1. MCP tools (small payloads — metadata only):
   search_cubes_by_title("labour force")  → productId
   get_sdmx_structure(productId=...)      → dimension layout + codes
   get_sdmx_key_for_dimension(...)        → OR key for large dimensions

2. Python script (data never enters context):
   url = "https://mcp-statcan.com/files/sdmx/<pid>/<key>?lastNObservations=12"
   → validate URL domain → write to ./statcan_<pid>.csv → print summary only

3. Follow-up script (analysis from local file):
   rows = list(csv.DictReader(open("./statcan_<pid>.csv")))
   → filter / sort / aggregate → print only the result

For questions that need computation (rank, compare, growth), query_sdmx runs the SQL on the server and returns only the answer, so raw rows never enter the context window.

Claude Code (bash sandbox) uses the statcan CLI instead:

statcan search "labour force"
statcan metadata 14-10-0287-01  # run first to find dimension IDs
statcan download 14-10-0287-01 --key "1.1.1" --last 12 --output ./lfs.csv  # run `statcan metadata <pid>` first to find real dimension IDs
awk -F',' 'NR>1 && $1=="Canada"' ./lfs.csv | sort -t',' -rn -k5 | head -10

---

MCP Prompts

The server ships five prompts accessible as slash commands in supported clients. Each has dual instructions — Claude Code (bash) and Claude.ai web (Python script).

Prompt	What it teaches
/statcan-data-lookup	End-to-end: search → structure → build key → fetch to local file → analyze
/sdmx-key-builder	SDMX key syntax: wildcards, OR keys, time parameters, download URL format
/statcan-download	Download a specific table: CLI commands + Python script alternative
/statcan-explore	Sample before committing: 3-period fetch, column layout, size estimate
/statcan-vector-pipeline	Multi-series download and cross-series comparison

Usage in Claude Code:

/statcan-data-lookup topic="consumer price index" analysis_goal="trend last 5 years"
/statcan-download product_id=18100004 last_n=24

---

statcan CLI

A standalone CLI for downloading StatCan data without an LLM. Outputs pipe-friendly CSV/JSON to stdout; progress and errors go to stderr.

Exploring a table (search, then metadata, then download):

Agent mode (--json envelopes, typed exit codes, statcan describe):

Recordings are generated from the scripts in assets/tapes/ with VHS: vhs assets/tapes/cli-agent.tape.

Install:

pip install statcan-mcp-server   # or: uvx statcan-mcp-server (no install)

Commands:

statcan search <term>            Search tables by keyword
statcan metadata <product-id>    Show table structure (dimensions + members)
statcan download <product-id>    Download observations via SDMX
statcan vector <vector-id>...    Download one or more vector series
statcan codeset                  Show StatCan code definitions (UOM, frequency, etc.)
statcan describe                 Commands, options, exit codes and JSON envelopes as JSON

Common usage:

# Find a table
statcan search "consumer price index"
statcan search "labour force" --max-results 10 --format json

# Inspect structure before downloading
statcan metadata 18-10-0004-01
statcan metadata 18100004 --full        # show all dimension members

# Download data (run `statcan metadata <pid>` first to find real dimension IDs)
statcan download 18-10-0004-01 --key "1.1.1" --last 12 --output cpi.csv
statcan download 18-10-0004-01 --key "1.1.1" --start 2020-01 --end 2024-12
statcan download 18-10-0004-01 --key "1.1.1" --last 5 --dry-run   # preview SDMX URL

# Download by vector ID
statcan vector v41690973 --last 24 --output series.csv
statcan vector v41690973 v41690974 --last 12 --output multi.csv

# Decode numeric codes
statcan codeset --type uom
statcan codeset --type frequency --format json

Output formats: csv (default for download/vector), table (default for search/metadata/codeset), json, ndjson

For agents and scripts: put --json before the subcommand (or set STATCAN_OUTPUT=json) to get one JSON envelope on stdout, JSON errors with a hint on stderr, and stable exit codes (0 ok, 2 usage, 3 not found, 4 rejected, 5 retryable, 6 empty). With --output, the envelope summarizes the file (rows, columns, period range, series count) so it never has to be read back. Details: machine-readable mode.

Pipe patterns:

# Top 10 by value (run `statcan metadata <pid>` first to find real dimension IDs)
statcan download 14-10-0287-01 --key "1.1.1" --last 1 --format csv \
  | awk -F',' 'NR>1' | sort -t',' -k5 -rn | head -10

# Extract unique geographies (run `statcan metadata <pid>` first to find real dimension IDs)
statcan download 14-10-0287-01 --key "1.1.1" --last 1 --format csv \
  | awk -F',' 'NR>1 {print $1}' | sort -u

# Chain search → download (run `statcan metadata <pid>` first to find real dimension IDs)
PID=$(statcan search "CPI" --format json | python3 -c "import sys,json; print(json.load(sys.stdin)[0]['Product ID'])")
statcan download $PID --key "1.1.1" --last 12 --output cpi.csv

For the complete CLI reference see src/cli/README.md.

---

Features & Tools

SDMX Tools — server-side filtered data fetch

Only the slice you request is returned. No downloading full tables.

Tool	Description
get_sdmx_structure	Dimension codelists + key syntax for a table. Call before get_sdmx_data.
get_sdmx_data	Filtered observations by productId + key, returned inline (capped at 500 rows).
query_sdmx	Fetch observations and run one read-only SQL query server-side in a temporary in-memory table — returns only the result. Works on the hosted server; no row cap on the fetch.
get_sdmx_vector_data	Observations for a single vectorId via SDMX.
get_sdmx_key_for_dimension	All leaf member IDs for a dimension as a ready-to-paste OR key — excludes totals and subtotals, so sums and rankings don't double-count (e.g. NOC, CMAs).

Key syntax (passed to get_sdmx_data):

• "1.2.1" — Geography=1, Gender=2, Age=1
• ".2.1" — all geographies (wildcard), Gender=2, Age=1
• "1+2.2.1" — Geography 1 or 2, Gender=2, Age=1

Note: Wildcard (.) returns every member of the dimension, including totals and parent aggregates. For sums or rankings use leaf codes from get_sdmx_key_for_dimension. Large wildcards can exceed get_sdmx_data's 500-row cap; use query_sdmx for those.

WDS Discovery & Metadata

Tool	Description
search_cubes_by_title	Full-text search across all StatCan tables. AND logic, capped at 25 results.
get_all_cubes_list / _lite	Paginated table inventory (offset/limit, default 100/page).
get_cube_metadata	Dimension info, member lists, date ranges. summary=True caps members at 10 per dimension.
get_code_sets	Decode StatCan numeric codes (frequency, UOM, scalar factor, status).

WDS Series Resolution & Change Detection

Tool	Description
get_series_info	Resolve {productId, coordinate} pairs to vectorId + metadata.
get_series_info_from_vector	Resolve a vectorId to productId, coordinate, titles, frequency.
get_changed_cube_list	Tables updated on a specific date.
get_changed_series_list	Series updated on a specific date.
get_changed_series_data_from_cube_pid_coord	Data points that changed for a coordinate.
get_changed_series_data_from_vector	Data points that changed for a vectorId.
get_bulk_vector_data_by_range	Multiple vectors filtered by release date range.

Composite & Database Tools (local/stdio mode only)

These tools are not available on the hosted Render server — SQLite is per-process and not shared across users.

Tool	Description
fetch_sdmx_to_database	Fetch SDMX observations (productId + key) into the observation store — no row cap, rows stay out of context. table_name becomes a view.
fetch_vectors_to_database	Fetch vectors by reference period range into the same store. table_name becomes a view.
store_cube_metadata	Fetch full cube metadata into SQLite — browse all members and vectorIds with SQL.
query_database	Read-only SQL against the local SQLite database.
create_table_from_data / insert_data	Create or append to a table of your own data.
list_tables / get_table_schema / drop_table	Database utilities (tables and views; list_tables also reports store size).

Observation store. Both fetch tools upsert into two shared tables, so data accumulates instead of being copied into a new table per call:

Table	One row per	Key columns
_statcan_series	series	series_id (<productId>:<SDMX key>, e.g. 18100004:2.2), product_id, vector_id, dims_json (labels), uom
_statcan_observations	series × period	series_id, ref_date (period start), period (2026, 2026-Q2, 2026-08), value, status, source, fetched_at, release_time

• Re-fetching updates values in place; each fetch reports how many observations were new, revised or unchanged.
• An SDMX and a WDS fetch of the same series share rows (WDS coordinates map to the same SDMX key).
• table_name is a view over the series and periods that call fetched, with the same columns as before (dimension labels, period, value, VECTOR_ID, UOM, ...).
• Cross-table analysis needs no schema discovery:

-- Seasonal component: CPI not seasonally adjusted (18100004) minus seasonally adjusted (18100006)
SELECT nsa.period, ROUND(nsa.value - sa.value, 2) AS seasonal
FROM _statcan_observations nsa JOIN _statcan_observations sa USING (ref_date)
WHERE nsa.series_id = '18100004:2.2' AND sa.series_id = '18100006:1.1'
ORDER BY ref_date DESC;

Typical workflow

Claude.ai web (hosted server):

1. search_cubes_by_title("unemployment rate")
   → productId e.g. 14100287

2. get_sdmx_structure(productId=14100287)
   → dimension positions + sample codes

3. get_sdmx_key_for_dimension(productId=14100287, dimension_position=3)
   → or_key for large dimensions

4. get_sdmx_data(productId=14100287, key=".2.1", lastNObservations=24)
   → returns download_csv URL

5. Python script: validate URL domain → write to ./statcan_14100287.csv → analyze → print summary

Claude Code (bash sandbox):

statcan search "unemployment rate"
statcan metadata 14100287
statcan download 14-10-0287-01 --key "1.1.1" --last 24 --output ./lfs.csv  # run `statcan metadata <pid>` first to find real dimension IDs
awk -F',' 'NR>1 && $1=="Canada"' ./lfs.csv | sort -t',' -rn -k5 | head -10

---

Project Structure

src/
├── api/
│   ├── cube/
│   │   ├── discovery.py         # search_cubes_by_title, get_all_cubes_list
│   │   ├── metadata.py          # get_cube_metadata
│   │   └── series.py            # get_series_info, change detection
│   ├── vector/
│   │   └── vector_tools.py      # vector series, bulk range fetch
│   ├── sdmx/
│   │   └── sdmx_tools.py        # get_sdmx_structure, get_sdmx_data, query_sdmx, get_sdmx_key_for_dimension
│   ├── composite_tools.py       # fetch_sdmx_to_database, fetch_vectors_to_database, store_cube_metadata (stdio only)
│   └── metadata_tools.py        # get_code_sets
├── cli/
│   ├── main.py                  # statcan CLI entry point (Typer app)
│   ├── output.py                # write_output, format helpers
│   └── commands/
│       ├── search.py            # statcan search
│       ├── metadata.py          # statcan metadata
│       ├── download.py          # statcan download
│       ├── vector.py            # statcan vector
│       └── codeset.py           # statcan codeset
├── db/                          # SQLite: observation store (store.py), per-request SQL for query_sdmx (ephemeral.py), schema, queries
├── models/                      # Pydantic input models
├── util/
│   ├── registry.py              # ToolRegistry — @decorator → MCP Tool schema
│   ├── truncation.py            # Response truncation + pagination guidance
│   ├── sdmx_xml.py              # SDMX-ML (structure-specific XML) → labelled tabular rows
│   └── cache.py                 # 1-hour TTL cache for cube list
├── config.py                    # BASE_URL, SDMX_BASE_URL, RENDER_BASE_URL, TRANSPORT, PORT
└── server.py                    # create_server(), MCP Prompts, resources, HTTP transport, CLI

---

Known Issues

Issue	Status	Workaround
"Unable to open database file" on Claude Desktop	Active	Pass --db-path /Users/<you>/.statcan-mcp/statcan_data.db in your config
SSL verification on broken trust stores	Configurable	Verification is on by default; set STATCAN_VERIFY_SSL=false only if your system can't validate StatCan's certificate
lastNObservations + startPeriod/endPeriod → 406	Active	Use one or the other, not both
Wildcard/OR rows missing or mislabelled; wrong VECTOR_ID/UOM	Fixed in v0.12.0	Caused by StatCan's SDMX-JSON index encoding; data is now read as structure-specific XML (#90)
Context overflow may cause data fabrication	Mitigated	Use query_sdmx (hosted) or fetch_sdmx_to_database (local) so raw rows never enter context; get_sdmx_data caps inline rows at 500

---