siat-mcp

imagePython imageMCP

Demo

imageWatch the demo

MCP server for Uzbekistan official statistics — siat.stat.uz and nsdp.stat.uz.

What is this?

siat-mcp connects AI assistants (Claude, Cursor, etc.) to Uzbekistan's official statistical portals via the Model Context Protocol. The marketplace and desktop plugin packages expose 13 read-only tools, including one SIAT tool that searches and retrieves data in a single request:

• siat.stat.uz — National Statistics Committee's SIAT portal: hundreds of datasets covering population, economy, agriculture, labour, and more. Browse by category, search by keyword, and retrieve full historical tables with regional breakdowns.
• nsdp.stat.uz — National Summary Data Page (IMF DSBB-compliant): 22 internationally standardized macroeconomic indicators (GDP, CPI, exchange rates, BOP, government debt, banking sector, etc.) in SDMX 2.1 XML format.

Features

• Claude and ChatGPT compatibility: Local stdio and remote Streamable HTTP at /mcp, with read-only tool annotations. See connection and marketplace setup for hosting, connection steps, and public directory submission requirements.
• High-Performance Caching: Implements asyncache and cachetools to cache the catalog, dataset files, and API responses for one hour, avoiding redundant multi-megabyte fetches while allowing updates to appear.
• Fast JSON Handling: Uses pysimdjson for SIMD-accelerated inbound JSON parsing when available, with orjson kept for fast output serialization and as a parser fallback.
• Connection Pooling: Maintains an active HTTP keep-alive connection pool via a shared httpx async client to bypass repeated TLS handshake latency.

Available Tools

siat.stat.uz tools

nsdp.stat.uz tools

Detailed SIAT tools (enabled in plugin packages)

Plugin packages enable these ten additional tools automatically. A standalone server still starts with three tools; set SIAT_MCP_LEGACY_TOOLS=1 in its environment to expose all 13. They support catalog inspection, raw tables, bulk downloads, and compatibility with existing clients.

SIAT quick use

Call siat_query directly; catalog browsing is unnecessary for a normal data request:

siat_query(request="Gross domestic product")
siat_query(request="dataset 361 in 2024", classifier="1703", lang="ru")
siat_query(request="1.01.01.0001", period="all", start_year=2020)

request accepts a natural-language indicator description, exact catalog code, or numeric dataset ID. By default the response returns the latest period and at most 15 classifier rows. A year in the request selects that year; phrases such as "trend" or "since 2020" select a bounded time series. Use classifier for a territory, sector, or classifier code. Use period="all", start_year, end_year, max_rows, and max_periods for explicit control. If the name matches several indicators, the tool returns up to five named choices with IDs; call it once more with the selected ID. The response reports when rows or periods were truncated and links to the source file.

Detailed SIAT tool reference

list_categories(lang?)

Returns the top-level statistical domains (e.g. "Population", "Economy", "Agriculture") with a count of how many sub-items each contains.

get_category(category_id, lang?)

Returns the immediate children of a category — either sub-categories (drill deeper) or leaf datasets (ready to fetch). When a category contains multiple leaf datasets, the response includes a bulk_fetch hint with comma-separated ids for get_datasets.

search_catalog(query, lang?)

Case-insensitive substring search through the full catalog tree. Each result includes a path field showing where in the hierarchy it lives (e.g. "Economy > National Accounts > GDP"). Results with has_children: false are leaf datasets you can pass directly to get_dataset.

list_datasets(lang?)

Returns every leaf dataset in the catalog as a flat list. Includes id, multilingual names, code, period, updated_at, and status. Use this for bulk inspection or filtering.

get_dataset_metadata(dataset_id, lang?)

Uses SIAT's dedicated /sdmx/{id}/metadata/ endpoint to fetch descriptor fields without downloading the full data table. Fields include a stable key when the API provides one. The dedicated endpoint can contain fewer publication fields than the metadata embedded in the downloadable JSON; use get_dataset when those additional fields are needed.

get_dataset_info(dataset_id, lang?)

Uses /sdmx/{id}/ to identify the dataset's classifier type (for example, territories), catalog code, frequency, status, and update timestamp without downloading observations.

get_dataset_table(dataset_id, lang?, classifier_ids?, periods?, start_year?, end_year?, max_rows?)

Uses /sdmx/{id}/table/ to return classifier IDs, labels, root flags, map codes, and period values. Responses are limited to 10 classifier rows by default (maximum 100) and report when additional rows were omitted. periods accepts exact comma-separated labels such as 2025,2026-Q1.

get_portal_counters()

Uses /common/counters/ for current portal totals. These counts describe SIAT coverage and activity, not an indicator's statistical observations.

get_dataset(dataset_id, lang?)

Fetches the complete dataset. Returns three sections:

• summary — indicator name, unit, year and period ranges, and row count
• metadata — all descriptor fields
• data — list of observation rows: code, classifier label (also in the legacy region field), and annual, quarterly, or monthly values ("2024": 1234.5, "2025-Q1": 300.2, "2025-M01": 100.1, ...)

get_datasets(dataset_ids?, dataset_codes?, lang?, include_metadata?, start_year?, end_year?)

Fetches multiple SIAT datasets in one MCP tool call. Pass comma-separated dataset ids such as "544,548,582" or leaf catalog codes such as "1.01.01.0001,1.01.01.0002". By default, metadata is omitted to reduce token usage in bulk calls; set include_metadata=true when descriptor fields are needed. Use start_year and end_year to trim annual, quarterly, and monthly observation columns.

Example: get_datasets(dataset_ids="3093,3094,3095,3096,3097,624", include_metadata=false).

Tool details — nsdp.stat.uz

nsdp_list_datasets()

Returns the full static catalog of 22 NSDP datasets. Each entry includes:

• code — IMF/SDMX identifier (e.g. "NAG", "BOP", "EXR")
• name — human-readable English name
• freq — publication frequency ("A", "M", "Q", or combined like "A/Q")
• xml_url — direct URL to the SDMX 2.1 XML file
• source — hosting domain (stat.uz, cbu.uz, api.mf.uz, uzse.uz)

nsdp_get_dataset(indicator_code)

Fetches and parses an NSDP dataset. indicator_code is case-insensitive. Returns:

• dataset — name, code, prepared timestamp, reporting period, freq, xml_url, source domain
• series — list of time series, each with:
  • indicator — SDMX INDICATOR code (e.g. "NGDP_PA_XDC")
  • freq — frequency code
  • unit_mult — scale exponent (0 = units, 6 = millions, 9 = billions)
  • dimensions — all SDMX dimension attributes
  • observations — [{time_period, value}] (ISO 8601 periods; value is null when missing)

Available NSDP codes: NAG, CPI, GGO, CGO, CGD, DCS, CBS, INR, SPI, BOP, EXD, ILV, MET, IIP, EXR, IND, LMI, PPI, FSI, POP, SDG, DOTS

Language Support

SIAT tools that return names accept an optional lang parameter:

If a name is unavailable in the requested language, the server falls back to English, then Uzbek.

Typical Workflows

Get a SIAT indicator:

siat_query(request="Gross domestic product")

Get a SIAT trend for one classifier:

siat_query(request="361", classifier="Uzbekistan", period="all", start_year=2015)

If the indicator description is ambiguous, use an ID from the returned matches list in a second siat_query call.

Use nsdp.stat.uz:

1. nsdp_list_datasets()           → see all 22 IMF datasets with codes
2. nsdp_get_dataset("EXR")        → fetch monthly exchange rates (UZS/USD)
3. nsdp_get_dataset("NAG")        → fetch GDP national accounts (annual + quarterly)

Install from a marketplace or desktop package

No separately hosted MCP server is needed. The tools run on the user's computer and fetch public statistics directly.

Release status: These installation files are prepared locally. Repository marketplace commands will work for other users after the changes are pushed to GitHub. Desktop downloads must first be attached to a GitHub release. There is no approved public directory listing yet.

Claude Code

Install uv once if it is not already available, then restart Claude Code so it can find it. In Claude Code:

/plugin marketplace add alishersuyunov/siat-mcp
/plugin install siat-mcp@siat-statistics

Where your client has Add marketplace, use https://github.com/alishersuyunov/siat-mcp, then select siat-mcp → Install. The first tool start needs internet access to download its runtime/dependencies. You do not need to edit Claude's JSON configuration.

Codex

With uv available, register and install from a terminal:

codex plugin marketplace add alishersuyunov/siat-mcp
codex plugin add siat-mcp@siat-statistics

Or use your client's marketplace picker when available. Start a new local chat after installation. A cloud chat cannot launch this local process.

Claude Desktop: no Python installation

Obtain the .mcpb release asset matching your operating system and processor. Open it with Claude Desktop and confirm Install. Ask it to list NSDP datasets. The Windows AMD64 package is built locally; other operating systems need native builds before corresponding download links can be published. A .mcpb file is for Claude Desktop, not ChatGPT.

ChatGPT Desktop

The native desktop-plugin.zip is the separate OpenAI package. It includes the same local executable and can be used in supported local plugin workflows. It is not a universally installable web app or an approved public listing. OpenAI currently restricts public local-MCP submissions to approved partners. See local packaging and release instructions.

Prefer manual setup or want to develop the server? Continue below.

Installation: start here

New to MCP? It is the connection that lets an AI assistant use this project's statistics tools. You do not need to write any code. You will download the project, install its dependencies, and tell your assistant where to find it.

The guide below sets up Claude Desktop on your own computer. You do not need an API key or a hosting account for this setup. Keep the downloaded project folder: Claude will use it whenever it starts the tools.

Using ChatGPT or Claude in a web browser? Those need a publicly accessible HTTPS server. Installing the project on your laptop alone does not connect it to the web app. See the separate remote connection guide. This project does not currently provide a shared hosted endpoint.

1. Install Python and Claude Desktop

• Install Python 3.11 or newer from python.org. On Windows, enable Add python.exe to PATH if the installer offers it.
• Install Claude Desktop from Claude's download page.

On Windows, open PowerShell from the Start menu and run:

py --version

On macOS, open Terminal and run:

python3 --version

You should see Python 3.11 or a newer version. If the command is not found, close and reopen the terminal after installing Python.

2. Download this project

1. Open the siat-mcp GitHub page.
2. Click the green Code button, then Download ZIP.
3. Extract the ZIP into a folder you plan to keep, such as your Documents folder. On Windows, right-click the ZIP and choose Extract All.
4. Open the extracted folder containing server.py and pyproject.toml. The folder will usually be called siat-mcp-main.

You do not need Git for this method.

3. Install the project's dependencies

Open a terminal in that extracted folder. On Windows, open the folder in File Explorer, click the address bar, type powershell, and press Enter. On macOS, type cd in Terminal, drag the extracted folder into the window, then press Enter.

Copy and run the following commands one line at a time. Do not copy the surrounding code-box markers.

Windows — PowerShell:

py -m venv .venv
.\.venv\Scripts\python.exe -m pip install .
.\.venv\Scripts\python.exe server.py --help

macOS — Terminal:

python3 -m venv .venv
./.venv/bin/python -m pip install .
./.venv/bin/python server.py --help

The first command creates a .venv folder for this project's Python dependencies. The second installs them and needs an internet connection. The final command should show help text containing --transport, --host, and --port.

If you see that help text, installation worked. You do not need to activate the environment or leave the terminal running.

4. Connect it to Claude Desktop

Claude needs two full paths: the Python program inside .venv, and this project's server.py file.

Print the exact paths for your computer:

Windows — in the same PowerShell window:

(Resolve-Path .\.venv\Scripts\python.exe).Path
(Resolve-Path .\server.py).Path

macOS — in the same Terminal window:

printf '%s\n' "$PWD/.venv/bin/python" "$PWD/server.py"

Next, open Claude's configuration file in a plain-text editor such as Notepad on Windows or TextEdit in plain-text mode on macOS:

On Windows, press Win + R, paste %APPDATA%\Claude, and press Enter to open the folder. On macOS, use Finder's Go → Go to Folder and enter ~/Library/Application Support/Claude.

If the file does not exist, create it with the exact name claude_desktop_config.json. On Windows, make sure it is not accidentally saved as claude_desktop_config.json.txt.

For a new or empty configuration file, paste this example and replace both example paths with your own:

Windows example:

{
  "mcpServers": {
    "siat-stat-uz": {
      "command": "C:/Users/YourName/Documents/siat-mcp-main/.venv/Scripts/python.exe",
      "args": ["C:/Users/YourName/Documents/siat-mcp-main/server.py"]
    }
  }
}

Use forward slashes (/) in the Windows JSON paths, as shown above. If you paste paths from PowerShell, change each backslash (\) to /. Keep the quotation marks, even if your path contains spaces.

macOS example:

{
  "mcpServers": {
    "siat-stat-uz": {
      "command": "/Users/YourName/Documents/siat-mcp-main/.venv/bin/python",
      "args": ["/Users/YourName/Documents/siat-mcp-main/server.py"]
    }
  }
}

Already have other MCP tools configured? Keep them. Add only the siat-stat-uz entry inside your existing mcpServers object, with a comma between entries. Do not replace your entire configuration or add a second mcpServers key.

Save the file, then fully quit and reopen Claude Desktop. On Windows, quit it from the system tray if it remains running there; on macOS, use Cmd + Q. Claude starts the server automatically.

5. Try your first request

Start a new Claude conversation and ask:

Use siat-stat-uz to list the available NSDP datasets.

If Claude asks for permission to use the tool, approve the request. You should receive a list containing codes such as NAG (GDP), CPI (consumer prices), and EXR (exchange rates). This first check uses a built-in list and does not download statistics.

Then try a request that retrieves live data:

Use siat-stat-uz to fetch the EXR exchange-rate dataset and include its source link.

Troubleshooting

Need help? Open a GitHub issue with your operating system, the step that failed, and the error message. Remove any private information from configuration files before sharing them.

Other installation options

Already use Git and uv?

git clone https://github.com/alishersuyunov/siat-mcp.git
cd siat-mcp
uv venv
uv pip install -e .

Then follow steps 4–5 above, using the paths inside your cloned folder.

Using Claude Code? After completing step 3, register the server using your actual full paths:

claude mcp add siat-stat-uz -- "/absolute/path/to/.venv/bin/python" "/absolute/path/to/server.py"

On Windows, use the full path to .venv/Scripts/python.exe for the first path. Keep quotation marks around paths containing spaces.

Plugin packages expose all 13 tools automatically. Manual standalone setups retain the three-tool interface unless you add "env": {"SIAT_MCP_LEGACY_TOOLS": "1"} to the server entry. The plugin still shows one connector: all 13 tools belong to that connection.

Example Prompts

Once configured, you can ask your AI assistant things like:

siat.stat.uz:

• "Estimate the Palma ratio over time and plot the Lorenz curve"
• "Which sectors have driven GDP growth in 2025?"
• "Show wages across sectors in 2025"
• "What datasets are available under the 'Labour Market' category?"
• "Get me the metadata for dataset 42 — what does it measure and in what units?"

nsdp.stat.uz:

• "Show me Uzbekistan's exchange rate history against the USD"
• "What happened to inflation (CPI) after 2017?"
• "Fetch Balance of Payments data and summarize the current account trend"
• "Get the latest GDP figures and compare annual vs quarterly growth"
• "What is the current level of external debt?"

Data Sources

This MCP server fetches public data on demand, caches responses in memory, and returns them to the connected client. SIAT cache entries expire after one hour; NSDP XML entries expire after six hours. It does not modify the upstream sources.

Development checks

Run python -m unittest discover -s tests -v after installing the project. The compatibility test starts a temporary local HTTP server; it needs loopback network access. Build distribution archives with uv build. Bulk retrieval accepts at most 20 dataset IDs/codes per call.