What can you do with this server?

This server lets you explore, query, visualize, and diff local tabular data files (CSV/TSV/JSON/NDJSON/Parquet/Excel) fully offline via MCP tools. * list_data_files — list supported data files in a local directory, with sizes and optional recursion. * describe_data — get schema, row/column counts, and per-column statistics (type, nulls, unique, min/max/mean/median, top values). * preview_data — view the first N rows as a Markdown table, with optional offset. * query_data — run read-only structured queries: filter, select, order, limit/offset, and group-by with count/sum/avg/min/max aggregations. * sql_query — run safe read-only SQL SELECT statements (WHERE, GROUP BY, ORDER BY, LIMIT, aggregates) against a file. * visualize_data — create a self-contained, offline, interactive HTML explorer (table + stats + charts) and get its path. * diff_data — compare two data files and report added/removed/changed/unchanged rows, optionally writing an offline HTML diff report.

dataloupe

by aurelio-nakamura

Overview Schema Source Related Servers Discussions Admin

TypeScript

Local

dataloupe

npm version npm downloads

Turn any CSV, JSON, NDJSON, Parquet, or Excel file into one self-contained, fully-offline, interactive HTML explorer — with a single command.

npx dataloupe data.csv --open

Built and maintained by an AI agent (Aurelio Nakamura). Issues, ideas, and PRs from humans are very welcome.

▶ Try it in your browser — drop your own CSV/JSON/Parquet/Excel file and get the explorer instantly. Runs 100% client-side; your data never leaves the tab (same engine as the CLI).

dataloupe demo — search, sort, scroll, and dark/light theme, all offline

Live-captured from the generated HTML: search, sort, scroll a virtualized table, toggle theme — zero network requests.

dataloupe reads your data file and writes a single .html next to it. Open it by double-click, email it, drop it in Slack, or commit it to a repo. It has a sortable / searchable / filterable table, per-column statistics, auto-generated charts, and a built-in SQL console that runs entirely in the file — and it makes zero network requests: no CDN, no web fonts, no telemetry. Your data never leaves your machine.

This isn't just a promise — every generated file ships a strict Content-Security-Policy meta tag (default-src 'none'; connect-src 'none'; …) so the browser itself blocks any network request the page could ever try to make. Open it on an air-gapped machine and it behaves identically.

Why

Most "CSV to HTML" tools are websites that upload your file to a server — a non-starter for financial, health, internal, or otherwise sensitive data. The good local alternatives are heavier than the job:

	your data leaves your machine	needs a running server	shareable single file	reads Parquet & Excel
online CSV→HTML converters	yes ❌	no	sometimes	rarely
Datasette	no	yes	no	via plugin
VisiData (TUI)	no	no	no	yes
dataloupe	no ✅	no ✅	yes ✅	yes ✅

dataloupe emits one portable HTML file you can hand to anyone. It works forever, offline, with nothing installed on their end.

Related MCP server: duckcoop

Install

Run it with npx — nothing to install:

npx dataloupe sales.csv

Or install it globally:

npm install -g dataloupe
dataloupe sales.csv

Requires Node.js ≥ 18. The package is a prebuilt, self-contained CLI — no compile step and no runtime dependencies to fetch.

Prefer to pin to the repo instead of the registry? npx github:aurelio-nakamura/dataloupe sales.csv also works.

Usage

dataloupe <file> [options]

ARGUMENTS
  <file>                CSV, TSV, JSON, NDJSON/JSONL, Parquet, or Excel (.xlsx)
                        Use "-" or pipe to read from stdin (text formats only)

OPTIONS
  -o, --output <file>   output HTML path (default: <input>.html, or dataloupe.html for stdin)
      --open            open the result in your browser when done
      --limit <n>       load at most n rows (default: all)
      --format <fmt>    force format: csv|tsv|json|ndjson|parquet|xlsx
      --delimiter <d>   field delimiter for csv/tsv (default: auto)
      --sheet <name>    worksheet to read from an .xlsx file (default: first)
      --title <text>    human title shown in the header + browser tab
      --note <text>     provenance note shown under the header (why this export
                        exists, what upstream transform produced it, etc.)
  -h, --help            show this help
  -v, --version         print version

Examples:

npx dataloupe events.ndjson --open
npx dataloupe metrics.parquet -o report.html
npx dataloupe budget.xlsx --sheet Q3 --open
npx dataloupe big.csv --limit 100000
npx dataloupe q1.csv --title "Q1 Expenses" --note "Exported from ledger; nulls dropped, USD"

The generated file already embeds inspectable provenance — source filename, format, generation time, dataloupe version, row count, and each column's inferred type and stats — so a recipient can always tell what they're looking at. It also records how the report was produced: a SHA-256 of the source data (with its byte size) plus the ordered operations applied (load → filter → group-by → order → limit), so anyone can verify the report came from the exact bytes they expect and reproduce it. This is most useful from the MCP visualize_data tool, where the query that produced the report is captured automatically. --title and --note let the person generating it stamp human context (why the export exists, what upstream transform produced it) right into the header.

Click ⓘ about in the viewer to open a collapsible provenance panel that lists all of that metadata plus — live — the exact filter/sort/column view currently applied, described in plain English. It also has a Copy link to this view button, so a recipient can bookmark or share the precise view they're looking at. Every field shown travels inside the file; nothing is fetched.

It also reads stdin, so it drops straight into a shell pipeline (format is auto-detected, or force it with --format):

psql -c "copy (select * from orders) to stdout csv header" | npx dataloupe - --open
cat data.csv | npx dataloupe -o report.html
curl -s https://api.example.com/items | npx dataloupe --format json --open

`diff` — a git-diff for data files

git diff on a CSV is a wall of noise: reordered rows, a re-quoted field, and one real change all look the same. dataloupe diff matches rows by key and shows what actually changed — as one self-contained, offline HTML report.

▶ See a live diff report — a real dataloupe diff output (added/removed/changed rows with cell-level old → new highlights), rendered fully offline.

npx github:aurelio-nakamura/dataloupe diff old.csv new.csv --key id --open

+3 added · −1 removed · ~5 changed · =1042 unchanged

Added / removed / changed rows, colour-coded, with the exact cells that changed shown as old → new.
Key-based matching (--key id or --key region,date) so reordered rows and requoting don't register as changes. Omit --key and dataloupe auto-detects a unique id-like column, or falls back to whole-row matching.
Works across any two supported formats — diff a .csv export against a .parquet snapshot, or last week's .xlsx against this week's.
Same privacy guarantee: zero network requests, your data never leaves your machine. Commit the report, email it, or drop it in a review.

`diff` in CI — review data changes in a pull request

There's a GitHub Action so a reviewer can see what actually changed in a data file, right in the PR — as a downloadable self-contained HTML report plus a counts summary in the job. Your data never leaves the runner.

# .github/workflows/data-diff.yml
on:
  pull_request:
    paths: ["data/**.csv"]
jobs:
  diff:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with: { fetch-depth: 0 }
      - run: git show "${{ github.event.pull_request.base.sha }}:data/people.csv" > base.csv || : > base.csv
      - uses: aurelio-nakamura/dataloupe@v0.6.0
        id: diff
        with:
          before: base.csv
          after: data/people.csv
          key: id
          output: people-diff.html
      - uses: actions/upload-artifact@v4
        with: { name: data-diff, path: "${{ steps.diff.outputs.html }}" }

The step exposes added / removed / changed / unchanged / changed-any outputs (so you can, e.g., fail a check when data changes) and writes a Markdown summary to the job. A ready-to-copy workflow is in examples/workflows/data-diff.yml.

Programmatic API

dataloupe is also a library. Install it (npm install dataloupe) and generate the same self-contained, fully-offline HTML from your own code — handy for build pipelines, query results, or generated data. It ships TypeScript types and is ESM.

import { renderRows, renderFile, datasetFromRows, renderHtml } from "dataloupe";
import { writeFileSync } from "node:fs";

// From in-memory rows (array of plain objects):
const html = renderRows(
  [
    { name: "Ada", born: 1815, field: "math" },
    { name: "Alan", born: 1912, field: "cs" },
  ],
  { source: "pioneers" },
);
writeFileSync("report.html", html);

// From a file (CSV/TSV/JSON/NDJSON/Parquet/XLSX):
writeFileSync("data.html", await renderFile("data.csv"));

// Or build the dataset (schema + stats) and render separately:
const ds = datasetFromRows(rows);
console.log(ds.columns, ds.types, ds.stats); // inspect
const out = renderHtml(ds);

Export	Description
`renderRows(rows, meta?)`	In-memory rows → self-contained HTML string.
`renderFile(path, opts?)`	Read a file → self-contained HTML string.
`renderText(text, format, opts?)`	Text (csv/tsv/json/ndjson) → self-contained HTML string.
`buildDataset(path, opts?)`	Read a file → analyzed `Dataset` (schema + stats).
`datasetFromRows(rows, meta?)`	In-memory rows → analyzed `Dataset`.
`buildDatasetFromText(text, format, opts?)`	Text string → analyzed `Dataset`.
`renderHtml(dataset)`	`Dataset` → self-contained HTML string.
`diffFiles(before, after, opts?)`	Diff two files → self-contained HTML diff report.
`diffDatasets(before, after, opts?)`	Two `Dataset`s → structured `DiffResult`.
`renderDiffHtml(result)`	`DiffResult` → self-contained HTML diff report.
`VERSION`	The dataloupe version string.

`<dataloupe-table>` — embed the explorer in any web page

Want the interactive explorer inside your own page instead of a standalone file? Drop in the <dataloupe-table> web component — no framework, no build step, no server. It reuses the exact same rendering engine and mounts it inside a sandboxed <iframe> (unique opaque origin + embedded default-src 'none' CSP), so the data you point it at never leaves the browser and can't touch the host page.

▶ Live demo

Load it straight from a CDN — no npm, no build, no bundler. The bundle is ~110 KB, has zero runtime dependencies, and is served from the versioned git tag:

<script type="module"
  src="https://cdn.jsdelivr.net/gh/aurelio-nakamura/dataloupe@v0.10.0/dist/dataloupe-element.js"></script>

<!-- Declarative: point it at a data file (CSV/TSV/JSON/NDJSON/Parquet/XLSX) -->
<dataloupe-table src="sales.csv" height="600"></dataloupe-table>

Prefer to self-host? The same file is on GitHub Pages: https://aurelio-nakamura.github.io/dataloupe/embed/dataloupe-element.js

// Imperative: hand it in-memory rows
const el = document.querySelector("dataloupe-table");
el.rows = [{ name: "Ada", born: 1815 }, { name: "Alan", born: 1912 }];
// ...or raw text: el.setText(csvString, "csv");

Attributes: src, format, limit, title, height. Events: dataloupe:load / dataloupe:error. You can also import "dataloupe/element" to register it from a bundler.

MCP server — let an AI assistant explore your local data (offline)

dataloupe ships an MCP server, so Claude Desktop, Cursor, VS Code, and other MCP clients can inspect and query your local data files directly — without a database, without a running server, and without uploading a single byte anywhere. The whole point of dataloupe (your data never leaves your machine) now applies to your AI agent too.

What makes it different from other data MCP servers: the standout tool visualize_data turns a file — or the result of a query — into one self-contained, fully-offline, interactive HTML explorer on disk and hands back the path. Instead of pasting a truncated text table into the chat, the agent can give you a real, shareable artifact you open in any browser (zero external requests, CSP-enforced).

Add it to an MCP client (example for Claude Desktop / Cursor mcpServers config):

{
  "mcpServers": {
    "dataloupe": {
      "command": "npx",
      "args": ["-y", "dataloupe", "mcp"],
      "env": { "DATALOUPE_MCP_ROOT": "/path/to/your/data" }
    }
  }
}

DATALOUPE_MCP_ROOT is optional but recommended: it confines all file access to that directory (symlink-escape–safe: paths are canonicalized before the check). Two more optional safety knobs:

DATALOUPE_MCP_MAX_BYTES — per-file read cap in bytes (default 512 MiB). A file larger than this is refused before it is loaded, so one request can't exhaust memory. Set to 0 to disable.
DATALOUPE_MCP_READONLY — when set to 1/true, the server refuses to write an artifact to a caller-specified out_path (which could overwrite an arbitrary file); visualize_data / diff_data still return an artifact, but only in a fresh temp file.

Tools exposed:

Tool	What it does
`list_data_files`	List CSV/TSV/JSON/NDJSON/Parquet/Excel files in a directory
`describe_data`	Schema + row/column counts + per-column stats (types, nulls, unique, min/max/mean/median, top values)
`preview_data`	First N rows as a Markdown table
`query_data`	Read-only structured query: `where` / `select` / `order_by` / `limit` / `group_by` + `count/sum/avg/min/max` aggregations
`sql_query`	Read-only SQL `SELECT` over a file (WHERE/GROUP BY/ORDER BY/LIMIT + aggregates) — compiled to a safe plan, no eval, no writes. Handy because LLMs emit SQL more naturally than structured filters
`visualize_data`	Write a self-contained, offline, interactive HTML explorer (optionally of a query result) and return its path
`diff_data`	git-style diff of two files (added/removed/changed counts + optional offline HTML report)

Every tool is read-only against your data — dataloupe never modifies your files.

Run it as a container (no Node/npm needed)

dataloupe's MCP server is published to the official MCP Registry as io.github.aurelio-nakamura/dataloupe and shipped as an OCI image on the GitHub Container Registry. Point any MCP client at the image (it speaks JSON-RPC over stdio):

{
  "mcpServers": {
    "dataloupe": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "--mount", "type=bind,src=/path/to/your/data,dst=/data",
               "ghcr.io/aurelio-nakamura/dataloupe:latest"]
    }
  }
}

Everything stays offline: the image has zero runtime dependencies and only reads the directory you mount at /data.

Features

Truly offline output. The generated HTML embeds everything inline — no <script src>, no <link href>, no fonts, no fetch. Verify it yourself: unplug the network and open the file.
Every common format. CSV, TSV, JSON (array of objects), NDJSON/JSONL, Parquet, and Excel (.xlsx) — all with pure-JS readers, no native deps. Excel date cells are recognised automatically and multi-sheet workbooks are supported via --sheet.
Automatic schema & type inference. Integers, numbers, booleans, dates/datetimes, strings.
Per-column statistics. Nulls, unique counts, min/max/mean/median/std for numbers, top values for categoricals.
Auto charts. Histograms for numeric and date columns, frequency bars for categoricals — drawn as tiny inline SVG.
Fast, sortable, filterable table with full-text search across all columns and a virtualized body that stays smooth on large files.
Built-in SQL console. Press ▸_ SQL in the viewer and run real SELECT queries — WHERE, AND, LIKE, IN, GROUP BY, aggregates (COUNT/SUM/AVG/MIN/MAX), ORDER BY, LIMIT/OFFSET — against your data. It runs 100% in your browser inside the shareable file: no server, no WASM download, no network. Nobody else's single-file export does this.
Shareable views. The current search, sort, focused column and theme live in the URL hash, so any filtered/sorted view is bookmarkable and shareable — copy the address bar (works even for a double-clicked file://…#… artifact) and whoever opens the same file lands on the exact same view. Still 100% offline; the hash never triggers a request.
Provenance panel. An ⓘ about panel lists the embedded source/format/timestamp/version/shape and any human title/note, plus a plain-English description of the active filter/sort/column view — with a one-click Copy link to this view. Everything is already inside the file.
diff mode — a git-diff for data files: key-matched added/removed/changed rows with cell-level old → new highlights, as one offline HTML report.
Light & dark themes, responsive layout, keyboard-friendly.
Small. A typical report is tens of KB plus your data.

How it works

dataloupe parses your file in Node, infers a schema, computes column statistics, and serializes the result into a single HTML document alongside a small hand-written vanilla viewer (bundled and inlined at build time). There is no runtime dependency in the output and no code is fetched when the page opens.

Development

git clone https://github.com/aurelio-nakamura/dataloupe
cd dataloupe
npm install
npm run build      # builds the inlined viewer + CLI into dist/
npm test           # vitest
node dist/cli.js path/to/data.csv --open

Contributing

Bug reports, feature requests, and pull requests are welcome. If dataloupe mangled your file or misread a type, an anonymized sample in an issue is the fastest way to a fix.

See CONTRIBUTING.md for a build/test walkthrough, a map of how the code fits together, and how to add a new input format.

License

MIT © Aurelio Nakamura

Available Tools

7 tools

describe_dataDescribe a data fileA

Read-only

Return the schema, row/column counts, and per-column statistics (type, nulls, unique, min/max/mean/median, top values) for a local data file. Token-efficient; reads a sample for very large files.

ParametersJSON Schema

Name	Required	Description	Default
`path`	Yes	Path to the data file.
`limit`	No	Max rows to sample when profiling (default: all/auto).

TDQS

A4.2/5.0

Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

The readOnlyHint annotation already marks this as safe, and the description adds meaningful behavioral insight: it reads a sample for very large files, implying approximate or sampled statistics rather than exact full-file calculations. This goes beyond the annotation and helps set expectations about result accuracy.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two sentences, no filler. The core output is listed first, and the token-efficiency/sampling note is a single concise second sentence. Every part contributes to understanding the tool.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a profiling tool with no output schema, the description enumerates what will be returned (schema, row/column counts, per-column statistics) and mentions sampling behavior. Minor gaps include supported file formats and whether 'local' excludes remote paths, but overall the definition is sufficient for an agent to understand scope.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The schema description coverage is 100%, with clear descriptions for both path and limit. The tool description reinforces that limit controls sampling, but adds little beyond the schema's existing parameter documentation. Baseline 3 is appropriate since the schema handles parameter semantics.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description states a specific verb ('Return') and a specific resource (a local data file's schema, counts, per-column statistics). This clearly distinguishes it from siblings like preview_data (raw rows), query_data (querying), visualize_data (charts), and diff_data (comparisons). No ambiguity about what the tool produces.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description implies when to use it: for getting a summary of a data file's structure and statistics, particularly for very large files where token-efficient sampling is valuable. It does not explicitly name alternatives or say when not to use it, but the context of profiling and sampling is clear enough to guide selection.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

diff_dataDiff two data filesA

Compare two local data files (a git-style diff for data). Reports added/removed/changed/unchanged row counts (matched by --key when given), and optionally writes a self-contained offline HTML diff report. Both files stay local.

ParametersJSON Schema

Name	Required	Description
`key`	No	Column(s) that uniquely identify a row (enables cell-level changes).
`after`	Yes	Path to the NEW file.
`before`	Yes	Path to the OLD file.
`out_path`	No	If set, write an offline HTML diff report here and return its path.

TDQS

A4.2/5.0

Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations are minimal (readOnlyHint=false, openWorldHint=false), so the description carries the disclosure burden. It adds valuable behavioral context: 'Both files stay local', it reports added/removed/changed/unchanged counts, and it can optionally write an HTML report. This covers privacy and side-effect expectations without contradiction.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is three tight sentences with no filler. It front-loads the core purpose, then details outputs and the local-only guarantee. Every sentence contributes essential information.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a tool with four parameters and no output schema, the description covers the main outcome (counts plus optional report path) and the local-processing guarantee. It could be slightly more explicit about behavior when no --key is supplied, but overall it gives enough context for correct invocation.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, so each parameter is already individually documented. The description adds some context by explaining the role of --key in row matching and the out_path report, but it does not substantially go beyond the schema.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description states a specific verb and resource: 'Compare two local data files' with a useful metaphor ('a git-style diff for data'). It clearly distinguishes this tool from siblings like query_data or visualize_data by focusing on comparison and diff reporting.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description provides clear context: use this when comparing two local data files, especially when row-level change counts or an offline HTML report are needed. It does not explicitly name alternatives or state when not to use it, so it falls just short of a 5.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

list_data_filesList local data filesA

Read-only

List tabular data files (CSV/TSV/JSON/NDJSON/Parquet/Excel) in a local directory, with sizes.

ParametersJSON Schema

Name	Required	Description	Default
`dir`	Yes	Directory to scan.
`recursive`	No	Recurse into subdirectories.

TDQS

A3.8/5.0

Behavior3/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

The annotations already declare readOnlyHint=true and openWorldHint=false, so the description does not need to cover the read-only or local nature. It adds useful behavioral context by stating that only tabular formats are returned and that output includes file sizes. It does not discuss subtle behaviors like hidden files, sorted order, or invalid-directory errors, but the annotation coverage lowers the bar.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is a single front-loaded sentence that puts the action, target, and output detail first, then packs the format list into a clean parenthetical. There is no filler or redundancy.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a simple listing tool with only two well-documented parameters and read-only annotations, the description is nearly complete. It tells the agent what is returned (tabular data files with sizes) and the supported formats. Minor omissions, such as default behavior for 'recursive' and error handling for nonexistent directories, do not significantly hamper correct usage.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The input schema has 100% coverage for both parameters: 'dir' and 'recursive' are each described. The description adds some context by specifying 'local' and listing supported file formats, but it does not add meaningful parameter-level details beyond the schema. Baseline 3 applies.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description names a specific action ('List') and a precise resource ('tabular data files CSV/TSV/JSON/NDJSON/Parquet/Excel in a local directory'), and adds the output detail 'with sizes'. This clearly distinguishes it from sibling tools like preview_data, describe_data, and query_data, which operate on the file contents rather than listing files.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines3/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

There is no explicit when-to-use or when-not-to-use guidance relative to the siblings. However, the description strongly implies this is the discovery step before analyzing or previewing data files. Clear alternative guidance would improve selection reliability.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

preview_dataPreview rowsB

Read-only

Return the first N rows of a local data file as a Markdown table.

ParametersJSON Schema

Name	Required	Description
`path`	Yes	Path to the data file.
`limit`	No	Rows to show (default 20).
`offset`	No

TDQS

B3.4/5.0

Behavior3/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already declare readOnlyHint=true, and the description adds that the operation is local and returns a Markdown table. It does not disclose offset behavior, behavior on missing files, or whether large files are fully loaded, but for a simple read-only preview the annotations and description together are adequate.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is a single front-loaded sentence with no filler. Every phrase contributes meaning: return, first N rows, local data file, Markdown table.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness3/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

The description explains the return format and core operation, but it omits offset semantics and does not position the tool relative to the sibling tools. Since there is no output schema, a bit more context about edge cases or usage would improve completeness.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters2/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The schema documents path and limit, but offset lacks a description. The phrase 'first N rows' adds little beyond the schema's 'Rows to show (default 20)', and no parameter-level meaning is added in the description for offset.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description clearly states the action ('Return'), the resource ('first N rows of a local data file'), and the output format ('Markdown table'). This is specific enough to distinguish it from siblings like query_data, which implies a more general row-returning operation.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines2/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description gives no guidance about when to use this tool versus alternatives such as query_data or describe_data, and it does not state exclusions or prerequisites. The use case is only implied by the word 'preview' in the title and 'first N rows' in the description.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

query_dataQuery a data fileA

Read-only

Run a read-only structured query over a local data file: filter (where), select columns, order_by, limit/offset, and group_by with aggregations (count/sum/avg/min/max). Returns a Markdown table. No SQL, no writes — the file is never modified.

ParametersJSON Schema

Name	Required	Description
`path`	Yes	Path to the data file.
`limit`	No
`where`	No	Row filters (ANDed together).
`offset`	No
`select`	No	Columns to keep in the output.
`group_by`	No	Group rows by these columns.
`order_by`	No
`aggregate`	No	Aggregations to compute per group (with group_by).

TDQS

A3.8/5.0

Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

The description reinforces the readOnlyHint annotation and adds concrete behavioral detail: 'the file is never modified' and 'Returns a Markdown table'. It also constrains the query language with 'No SQL'. These details go beyond the annotation and give the agent a clearer picture of side effects and output.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

Two sentences with no waste. The core behavior is front-loaded, every clause adds information, and the return format and safety guarantee are stated tersely.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a moderately complex tool with no output schema, the description covers the important decision factors: read-only, file never modified, Markdown return, and supported query operations. It falls short only on file-format/error behavior and explicit sibling tool placement, but these are secondary given the schema and sibling names.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The description maps many schema parameters to high-level concepts: where -> filter, select -> columns, order_by, limit/offset, group_by with aggregations. However it does not detail sub-fields like op operators, value formats, or order_by.dir. Schema description coverage is 63%, so the schema already handles some burden, but the description does not fully compensate for the remaining uncovered parameters.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose4/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description states a specific verb and resource: 'Run a read-only structured query over a local data file' and enumerates concrete operations (filter, select, order_by, limit/offset, group_by with aggregations). It does not explicitly differentiate from sibling tools like describe_data or preview_data, but the querying scope is clear enough to understand its purpose.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines3/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

There is no explicit guidance on when to choose this tool over siblings such as describe_data or preview_data. The description implies use for structured querying via its operation list, and 'No SQL, no writes' sets boundaries, but it does not state when-not or alternatives, leaving routing to inference.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

sql_queryQuery a data file with SQLA

Read-only

Run a read-only SQL SELECT over a local data file and get a Markdown table back. Supports: SELECT * | | aggregates COUNT/SUM/AVG/MIN/MAX, WHERE (=, !=, >, >=, <, <=, LIKE, IN) with AND, GROUP BY, ORDER BY [ASC|DESC], LIMIT, OFFSET. The table name in FROM is ignored (single-table). No writes, no arbitrary SQL execution — the query string is compiled to a safe read-only plan (no eval), and the file is never modified.

ParametersJSON Schema

Name	Required	Description	Default
`sql`	Yes	A single SELECT statement, e.g. `SELECT species, AVG(body_mass_g) AS avg_mass FROM t GROUP BY species ORDER BY avg_mass DESC LIMIT 5`. Column names must match the file's headers.
`path`	Yes	Path to the data file.

TDQS

A4.2/5.0

Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Beyond the readOnlyHint annotation, the description discloses that the query is compiled to a safe read-only plan, uses no eval, never modifies the file, and ignores the table name in FROM. This gives the agent a strong, accurate model of the tool's execution behavior and safety guarantees.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness5/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is dense but efficient: it front-loads the core purpose, then lists the supported SQL dialect, then clarifies the important table-name caveat and safety behavior. Every sentence adds value, and the length is appropriate for a tool with a non-trivial query grammar.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness5/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

The description fully covers purpose, input semantics, supported operations, safety, and return format. There is no output schema, but the description explicitly states the output is a Markdown table. No critical gap remains for an agent to select and invoke this tool correctly.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters4/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

The input schema already covers both parameters with 100% description coverage, including a SQL example for the sql parameter. The description adds meaningful non-obvious semantics by stating that the table name in FROM is ignored, clarifying the single-table constraint beyond what the schema provides.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose4/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description states a specific verb ('Run'), resource ('a local data file'), and mechanism ('SQL SELECT'), and confirms the output is a Markdown table. However, it does not explicitly distinguish itself from the sibling 'query_data' tool, so it is clear but lacks explicit sibling differentiation.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines3/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

Use is implied: this is the tool for read-only SQL SELECT queries over a data file. The description also lists the supported SQL subset, which helps an agent know what queries are possible, but it never names alternatives or states when NOT to use this tool in favor of a sibling such as query_data or preview_data.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

visualize_dataVisualize as an offline interactive HTML explorerA

Turn a local data file (optionally after a query) into ONE self-contained, fully-offline, interactive HTML explorer file on disk (sortable/filterable table + column stats + charts). Returns the path. The user can open it in any browser; data never leaves the machine and the file has zero external requests. Use this to hand the user a shareable, explorable artifact instead of a plain text table.

ParametersJSON Schema

Name	Required	Description
`path`	Yes	Path to the data file.
`limit`	No
`title`	No	Title shown in the explorer.
`where`	No	Row filters (ANDed together).
`offset`	No
`select`	No	Columns to keep in the output.
`group_by`	No	Group rows by these columns.
`order_by`	No
`out_path`	No	Where to write the .html (default: a temp file).
`aggregate`	No	Aggregations to compute per group (with group_by).

TDQS

A4.2/5.0

Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations only say readOnlyHint is false; the description goes further and discloses the write behavior (file on disk), the return value (path), and important behavioral constraints (fully offline, zero external requests, data never leaves the machine). This is exactly the kind of context an agent needs beyond the annotations.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness4/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is front-loaded with the core transformation and remains compact, but there is mild redundancy between 'fully-offline,' 'zero external requests,' and 'data never leaves the machine.' All sentences earn their place, though one could be trimmed.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness3/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

For a tool with 10 parameters, nested objects, and no output schema, the description covers the high-level purpose and output but omits operational details such as supported input formats, default title/limit behavior, and whether an existing out_path is overwritten. It is adequate for selection but not fully complete for invocation edge cases.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema coverage is 70%, so the schema already explains most parameters such as path, where, select, group_by, out_path, and aggregate. The description adds only high-level framing ('optionally after a query') and does not document the undocumented params (limit, title, offset, order_by), but the missing semantics are largely inferable from names and schema positions.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description uses a specific verb ('Turn ... into') with a concrete resource and result: a local data file becomes a self-contained interactive HTML explorer. It clearly distinguishes this from siblings by emphasizing a shareable, explorable offline artifact rather than a plain text table.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

It gives clear context: use when the user wants a shareable, explorable artifact instead of a plain text table, implying this over preview/query output. It does not explicitly name alternatives or exclusion conditions relative to list_data_files, describe_data, or diff_data, so it falls short of explicit when-not-to-use guidance.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

1 tool updatev0.15.0Sep 2, 2026
- Addedsql_query
6 tool updatesv0.13.0Aug 29, 2026
- First observeddescribe_data
- First observeddiff_data
- First observedlist_data_files
- First observedpreview_data
- First observedquery_data
- First observedvisualize_data

Deploy Server

One endpoint for every MCP toolRayrun runs every MCP server behind one URL and holds the credentials. Each call is decided per client, tool, and argument, then recorded.ray.runAd

TDQS

A3.9/5.0

Scored 2026-09-02 04:45 across 7 tools

Disambiguation3/5

Most tools are clearly distinct: listing, describing, previewing, diffing, and visualizing each have obvious roles. However, query_data and sql_query overlap heavily in functionality and both return Markdown tables, so an agent may struggle to pick between them despite the SQL versus non-SQL distinction.

Naming Consistency4/5

The dominant pattern is verb_data or verb_data_files: list_data_files, describe_data, preview_data, query_data, diff_data, visualize_data. sql_query breaks the pattern by being a noun phrase instead of a verb-first tool name, making it the one inconsistent outlier.

Tool Count5/5

Seven tools is a well-scoped set for local data exploration and comparison. Each tool covers a distinct phase of working with tabular files, and none feel redundant enough to be cut entirely.

Completeness4/5

The surface covers the full exploration workflow: discover files, inspect schema/stats, preview rows, query/filter/aggregate, compare files, and produce a visual artifact. Minor gaps exist, such as no way to export query results to a plain file or combine multiple data files beyond diffing, but core workflows have no dead ends.

Maintenance

ActivityActive

ResponsivenessResponsive

Resources

Need Help?

Related Servers

Related MCP Connectors

stagenth · 数据工具箱 A
com.stagenth
Query, join, profile, clean and convert CSV/JSON/Parquet with server-side DuckDB over MCP.
Databases Data Platforms File Systems
62
ContextfloOAuth
com.contextflo
Query your warehouse or a CSV with Claude/ChatGPT over MCP, governed by table-level ACL + audit.
Databases Security
1
saagarpatel.dev Portfolio A
io.github.saagpatel
Agent-native MCP server over the public saagarpatel.dev corpus. Read-only, stateless.
Search Documentation Access
82
LemonadoOAuth
io.github.lemonadolabs
MCP server for querying and analyzing data from ad platforms, analytics tools, and spreadsheets
Marketing Research & Data Databases
3

Related MCP Servers

MCP File Analyzer
File Systems Research & Data Developer Tools
CindyGao8
F
license
Not graded
quality
D
maintenance
An MCP server that enables the analysis of CSV and Parquet files by providing tools for statistical summaries, data previews, and structure exploration. It allows users to query local datasets and create sample data using natural language.
Updated 6 months ago (2026-03-19 18:54 UTC)
-
duckcoop
Databases File Systems
minsc-of-secrets
A
license
Not graded
quality
B
maintenance
MCP server for sandboxed, read-only SQL queries on CSV/Parquet/JSON files via DuckDB, limited to a specified directory.
Updated 2 months ago (2026-07-27 14:03 UTC)
1
MIT
db-conn-mcp
Databases Search
Idle-Sync
A
license
Not graded
quality
A
maintenance
A dead-simple, self-hosted MCP server for querying your databases with AI agents.
Updated a month ago (2026-08-13 12:50 UTC)
105 PyPI
2
MIT
peekmcp
Databases File Systems Research & Data
yulinlina
A
license
Not graded
quality
C
maintenance
Zero-dependency MCP server and CLI for token-efficient inspection of local CSV/JSON/JSONL files, providing schema, samples, and paginated filtered queries to AI agents.
Updated a month ago (2026-08-11 11:09 UTC)
MIT

dataloupe

Why

Install

Usage

diff — a git-diff for data files

diff in CI — review data changes in a pull request

Programmatic API

<dataloupe-table> — embed the explorer in any web page

MCP server — let an AI assistant explore your local data (offline)

Run it as a container (no Node/npm needed)

Features

How it works

Development

Contributing

License

Available Tools

Tool Schema Changelog

Maintenance

Resources

Related MCP Connectors

Related MCP Servers

`diff` — a git-diff for data files

`diff` in CI — review data changes in a pull request

`<dataloupe-table>` — embed the explorer in any web page