Dataville

dataville-mcp

MCP server exposing Dataville's data source API as tools for MCP clients (Claude Desktop, Claude Code, etc.).

Dataville is a unified REST API over ten public datasets — Wikipedia, arXiv, Project Gutenberg, US Census, USDA FoodData, Papers with Code, SEC EDGAR, OpenAlex, PyPI, and Stack Exchange — behind one interface and one API key, with CSV/Parquet export and SQL query support. This package lets an MCP client search any of those sources as a tool call.

Requires a Dataville API key — get one from the Dataville dashboard.

Docs: docs.dataville.com · Connect an AI tool · Data sources. Examples: dataville-wikimedia, dataville-sec-edgar, company-intelligence-agent. Live demos: app.dataville.com/demos.

Tools

All four only read data. query_dataville covers every stored source except census and news, which are fetched live per search. It returns at most 1,000 rows per call and is billed per row returned. It reads Dataville's stored copy: some sources are loaded in bulk, others only hold records fetched before, so for one specific item search_dataville is the reliable route.

What you can ask

Once connected, ask in plain language and the client picks the source:

• "Find an arXiv paper on retrieval-augmented generation and summarize it."
• "What did Apple report in its latest SEC filing?"
• "What's the median household income in Travis County, Texas, per the US Census?"
• "How much protein is in 100 g of cooked lentils, according to USDA FoodData?"
• "Look up the requests package on PyPI — what's the latest version and license?"
• "List ten Project Gutenberg books by Jane Austen, with their subjects."
• "List five USDA FoodData entries with 'lentils' in the name, with their categories."

Setup

This package is a local (stdio) MCP server: the client launches it on your machine via npx. If you don't need it running locally, connecting to Dataville's hosted endpoint instead takes one line and no install.

Hosted, in one line (no install)

Dataville also serves MCP directly over HTTP, so a client can connect without running anything locally — no Node, no config file, no restart:

claude mcp add --transport http dataville https://api.dataville.com/mcp

That works with no credentials at all (anonymous limits). Add --header "Authorization: Bearer dataville_your_key_here" for the full quota. Cursor and VS Code users can use the install buttons at the top of this page. Other clients take the same URL; the app's Integrations page has the exact snippet for each. Use the hosted endpoint unless you specifically want to pin a version or work offline — the rest of this section covers that local setup.

Prerequisites

• Node.js (LTS) installed — this is what runs npx. Without it the server fails to start. Check with node --version.
• A Dataville API key — get one from https://app.dataville.com/api-keys.

The config block is the same everywhere; only *where* you put it differs:

{

"mcpServers": { "dataville": { "command": "npx", "args": ["-y", "@dataville/dataville-mcp"], "env": { "DATAVILLE_API_KEY": "dataville_your_key_here" } } } }

No install step needed — npx fetches and runs the package on demand.

Claude Desktop

1. Open Settings → Developer → Edit Config. This opens claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

1. Add the block above (merge into mcpServers if the file already has one), with your real key.
2. Fully quit and reopen Claude Desktop — quit from the menu bar / system tray, not just closing the window.
3. The dataville tools now appear under the tools icon in the chat box, and Settings → Developer shows dataville running.

Note: the server appears as tools, not in the Connectors directory — that directory only lists remote (hosted) connectors and will not find a local server. Ask naturally ("get Apple's latest revenue from dataville") and the client calls the tool.

Claude Code

claude mcp add dataville -e DATAVILLE_API_KEY=dataville_your_key_here -- npx -y @dataville/dataville-mcp

Restart the session so the tools load. Add -s user to make it available in every project instead of just the current one.

Configuration

DATAVILLE_API_BASE_URL is optional and defaults to https://api.dataville.com; set it to http://localhost:5000 to point at a local backend during development.

Where your API key goes

The server reads DATAVILLE_API_KEY from its environment and sends it only as an Authorization: Bearer header to DATAVILLE_API_BASE_URL, which is https://api.dataville.com unless you change it. It never logs the key or sends it anywhere else. The base URL must use https. Plain http is accepted only for localhost, 127.0.0.1 and [::1], so the key never crosses the network unencrypted.

All four tools are read-only: they search and query Dataville and never create, change or delete anything.

Running from source

git clone https://github.com/datavilleorg/dataville-mcp.git

cd dataville-mcp npm install npm run build

Check that it works

Ask your client one of these. Each answer is checkable on purpose — a model that skipped the tool and answered from memory sounds just as confident, so a reply on its own proves nothing.

Clients show when a tool ran. If you don't see that, say "use dataville" in the prompt to make it explicit, and check the answer against the source.

Development

npm run dev    # tsx watch

npm test # node test runner npm run build # tsc

Releasing

Publishes run from CI. npm uses trusted publishing (OIDC), so no npm token is stored; the MCP Registry step signs in with the MCP_PRIVATE_KEY repo secret. To cut a release: bump the version, update CHANGELOG.md, then publish a GitHub Release for the new tag. The Publish workflow builds, tests, and publishes to npm, then publishes server.json to the MCP Registry (its version is set from package.json, so don't bump it by hand).

The registry name com.dataville/dataville-mcp is authorized by a DNS TXT record on the dataville.com apex (v=MCPv1; k=ed25519; p= ), checked against the Ed25519 private key in the MCP_PRIVATE_KEY repo secret (hex). If the mcp-registry job fails at login, check that record and secret, then re-run just that job.