Autario MCP

autario-mcp

autario is data app infrastructure: your tools and thousands of public datasets in one normalized layer, with the data apps that run on top of it.

Connect your platforms, then let your assistant interpret the numbers. Connect your ad accounts, your shop, your web analytics, your social profiles, your CRM and your finance tools on autario once. Through this MCP server your assistant reads those tables, joins them and explains what moved: spend, revenue and ROAS per channel, the audience behind a traffic change, the posts that worked. The public catalogue (World Bank, FRED, Eurostat, OECD, WHO, IMF, SEC) sits next to your own data as context, so a sales dip can be read against inflation or market growth in the same answer.

• Brands and instances. Group connected accounts by brand, and pick which accounts (instances) a report reads, so an agency asks about one brand without another brand's numbers in the answer.
• Numbers the model does not make up. Report tools return what autario computes server-side from your own tables, and every public value cites its primary source, so the model explains figures instead of recomputing or guessing them.
• 57 tools, each marked read or write. The Access column in the Tool Reference below comes from each tool's own annotations, so a client can allow the read tools and ask before a write.
• Remote or local. Point any HTTP MCP client at https://autario.com/mcp with no install, or run it over stdio with npx autario-mcp.

autario.com | Developer docs | Get an API Key

Why autario-mcp

• Your platforms, not just public data. Ads, shop, analytics, social, CRM and finance connectors land in tables your assistant can query, join and compare, and marketing_report, social_360 and the other report tools return the same numbers the app shows.
• No hallucinated numbers. Every value is sourced from a known publisher and cited back to a primary URL. Use verify_value to double-check any claim.
• Cross-dataset joins, no setup. Indicators across different datasets share autario_time + autario_entity shadow columns, so get_entity_data(USA, [gdp, unemployment, life_expectancy]) returns one wide table, joined automatically.
• Statistical primitives built-in. correlate, regression, find_drivers, lag_analysis, seasonality_decomposition and more, with effect sizes, p-values, and plain-language interpretations.
• Charts that persist. publish_chart writes a Plotly spec to autario.com. The result is a permanent, embeddable URL like autario.com/chart/{slug} | the LLM builds the spec, autario pulls real rows for it, no hallucinated data path.
• LLM-agnostic. Works with Claude, GPT, Gemini, local models | anything that speaks MCP.

Quick Demo

Install the server, then ask your assistant questions like these. The model picks the right tools and answers with cited data.

Ask:    "Why did ROAS drop last week for this brand? Compare Meta Ads and Google Ads

against Shopify revenue." Tools: get_my_workspace -> marketing_report Output: Spend, revenue and ROAS per channel against the previous period, from the connected accounts of that brand only.

Ask:    "What drives US inflation? Look at money supply, oil prices, and unemployment."

Tools: list_indicators -> find_drivers Output: Ranked drivers with r, p-value, R squared per candidate.

Ask:    "Compare life expectancy in Germany, USA and Japan from 2000 to 2023, then publish a chart."

Tools: compare_entities -> publish_chart Output: Wide-format table joined on year, plus a permanent autario.com/chart/{slug} URL.

Ask:    "Is consumer confidence a leading indicator of US retail sales?"

Tools: lag_analysis Output: Cross-correlation peak at lag k, with interpretation in months.

Ask:    "Find a dataset on unemployment, build me a small app that charts it for

Germany, publish it as unlisted and give me the link." Tools: search_datasets -> get_dataset_schema -> create_app -> update_app_manifest -> upload_app_bundle -> get_app_preview_url -> publish_app Output: A live URL you own. No review queue, no waiting list, nobody to ask.

Install

Claude Desktop, Cursor, Cline (stdio)

~/.config/claude/claude_desktop_config.json on Mac/Linux, %APPDATA%\Claude\claude_desktop_config.json on Windows.

{

"mcpServers": { "autario": { "command": "npx", "args": ["autario-mcp"] } } }

Restart your client. The server reports tool count to stderr on launch.

Claude Web, OpenAI Custom GPTs, any HTTP MCP client

Point your client at the hosted endpoint. No install needed.

URL:        https://autario.com/mcp

Transport: Streamable HTTP (POST /mcp)

The hosted endpoint also supports MCP prompts: analyze-dataset, create-chart, compare-countries.

Enable write tools (publish charts, create datasets)

Add API credentials to the stdio config or send them as x-api-key / x-api-secret headers to the HTTP endpoint.

{

"mcpServers": { "autario": { "command": "npx", "args": ["autario-mcp"], "env": { "AUTARIO_API_KEY": "your_key", "AUTARIO_API_SECRET": "your_secret" } } } }

Get keys at autario.com/account.

Tool Reference

_57 MCP tools, organized by function._

Discovery & Query

Search the catalog, inspect schemas, pull rows.

Cross-Dataset Joins (Ontology)

The differentiator. Join indicators across datasets via shared time + entity shadow columns. No manual relationship setup.

Statistical Analysis

Run analyses against verified data. Outputs include effect sizes, p-values, and plain-language interpretations.

| pct_change | read | Period-over-period percentage change for an indicator. Use for growth rates (YoY, QoQ, MoM) \| "how fast did German GDP grow", "what is US CPI doing month over month". Runs on any verified autario i... | **entit