Data quality service

Know which data you can trust

DQaaS is a data quality service. Developers, analysts and data teams use DQaaS to profile datasets, check them against rules, score their quality and keep a governed catalogue, from a web console, a REST API, a CLI or an AI assistant.

In the browser, your file is checked on your own device and never leaves it.

Ways to use

Four ways to use it

  • CLI

    A command-line program that can profile and check files on your own machine without an account or server. It uses temporary disk copies; review configured storage and network connections. Optional --publish sends profile and rule metadata.

    dqaas run customers.csv -o report.html
  • API

    REST with a generated OpenAPI document: register sources, run jobs, manage rules and read the catalogue and reports.

    GET /v1/datasets/{id}/report?format=md
  • MCP

    Give an AI assistant the same access: a local MCP server that can also profile a file on your machine, and a hosted MCP endpoint that uses your access token.

    POST /mcp  (Authorization: Bearer dqat_...)
  • In the browser

    Attach a file and profile it on your own device. Nothing is uploaded.

    Open Attach locally in the console

All four share the same checks and report format. Browser results are computed on the device.

Process

How it works

  1. Connect or upload

    Upload a file or register a URL or table in a project.

  2. Profile and check

    DQaaS profiles every column, suggests rules, runs them and scores the dataset.

  3. Catalogue and report

    Add owners, descriptions and classifications, track lineage and export a report.

CLI

export DQAAS_URL=https://your-dqaas-host
dqaas login --token dqat_...
dqaas upload orders.csv --project PROJECT_ID --wait
dqaas analyze DATASET_ID --wait
dqaas rules DATASET_ID --suggest
dqaas report DATASET_ID --format html -o report.html

REST API

POST /v1/datasets/upload
POST /v1/datasets/{id}/analyze
GET  /v1/datasets/{id}/rules/suggestions
GET  /v1/datasets/{id}/report?format=md
GET  /v1/catalogue/search?q=customer
Authorization: Bearer dqat_...

Features

Capabilities

  • Included

    Ingest files and sources

    Start from what you already have.

    Upload CSV, TSV, JSON, JSONL, Parquet, Avro, Arrow, XLSX or embedded database files (for a workbook with several sheets, the largest visible sheet is read unless you pick another), or register an HTTPS URL or a read-only Postgres table. The format is detected from content and name.

  • Included

    Profiling

    See what is really in a column.

    Types, null and distinct counts, value statistics, placeholder detection, and schema and statistical drift between profiles.

  • Included

    Quality rules and scoring

    One score per dataset, with the failing rows counted.

    Suggested rules from the profile, rule packs and your own checks. Each rule reports failed over total rows; scores are weighted by severity and grouped by quality dimension, with passed and failed rule counts per dimension. A pass/fail gate for pipelines is available as an API route, a CLI command and an MCP tool.

  • Included

    Sensitive-column detection

    Find likely personal data early.

    Name and pattern heuristics flag columns such as email, phone or date of birth. Findings are labelled inferred, and sensitive sample values are redacted.

  • Included

    Profile on your own device

    Check a file without sending it anywhere.

    Attach a CSV, TSV, JSON, JSONL, Parquet, Avro, XLSX, Arrow or embedded database file and profile it on your own device; the file never leaves it. Optionally publish only the profile and rule results to the catalogue. Browser results are computed locally and are not verified by the server.

  • Included

    Catalogue and governance

    Everyone finds the right table and who owns it.

    Full-text search, glossary, column descriptions and classifications, policies and violations, stewardship tasks and coverage indicators.

  • Included

    Lineage and impact analysis

    Know what breaks downstream.

    Lineage edges are inferred or verified (verified requires evidence). Column-level lineage can be parsed from SQL you supply; it is inferred, labelled as parsed and never observed, and anything that cannot be resolved is listed, not guessed. Impact analysis runs upstream or downstream, optionally on verified edges only.

  • Included

    Reports

    Share results with people who never log in.

    Dataset and project reports as JSON, Markdown, HTML or PDF.

  • Included

    REST API, OpenAPI, CLI and MCP

    Fit it into pipelines and assistants.

    A bearer-token REST API with a generated OpenAPI 3.1 document, a command-line client, and MCP servers for AI assistants, local and hosted, that sign in with your access token or OAuth. Tool lists are in the docs.

  • Included

    Duplicate, reconciliation and semantic checks

    Catch problems a simple rule misses.

    Fuzzy near-duplicate detection, reconciliation of two datasets (row counts, aggregates, key sets), and inferred semantic and secret types for 25 kinds of value. Fuzzy matches and types are inferred, never proof.

  • Included

    Data contracts and lineage standards

    Work with the rest of your stack.

    Import and export Open Data Contract Standard (ODCS v3.2.0) rules, and export or ingest OpenLineage events. Only the rule parts of a contract are imported, and anything that cannot be mapped is listed, not guessed.

  • Included

    Drift and mergeable profiles

    See change, and combine partitions.

    Changes in value distributions between profiles, and estimated distinct counts and quantiles merged across partition profiles without re-reading the data.

  • Included

    Metrics and telemetry

    Watch quality in the tools you already use.

    A metrics endpoint with per-dataset quality figures for your monitoring tools, and an optional export of job metrics for self-hosting. Setup is in the docs.

  • Included

    Roles, tokens and audit log

    Control and trace access.

    Viewer, editor, admin and owner roles, scoped and revocable access tokens, and a tamper-evident audit log you can check.

  • Included

    Quotas and usage metering

    See what you use.

    Monthly quotas with usage shown in the console and API.

  • Included

    Interval and cron schedules

    Re-check data on a rhythm.

    Recurring jobs on a fixed interval (5 minutes to 1 year) or a 5-field cron expression in UTC, at most every 5 minutes.

Audience

Who it is for

  • Developers and data engineers

    Add checks to pipelines through the CLI or REST API.

  • Analysts

    See which tables are trustworthy before building on them.

  • Data teams and stewards

    Track ownership, classification and open stewardship tasks.

  • AI-assistant users

    Query quality and catalogue data through MCP.

Data handling

Data handling principles

  • Working copy deleted after each job. Rows are loaded into a working copy for the job, which is deleted when it ends. An uploaded file is kept only while Keep file is ticked.
  • Hosted in the European Union. Account data and results live on one EU-hosted server; this site is served from a global network and holds no customer data.
  • Retention you can set. Profiles, reports and the audit log have default periods an administrator can change.
  • Source rows are transient: they are processed for the job and deleted at its end. Uploads are kept for later scans only while Keep file is ticked, which is the default in the console.
  • Metadata is persisted: profiles, rule results, scores, catalogue entries and an audit log.
  • Sensitive values are redacted: sensitive columns store no top values, no min/max and no failing examples.
  • Strict privacy mode stores no sample values at all.
  • Remote sources are limited to secure public addresses. The security page lists the checks.
  • Findings are labelled inferred or verified. Sensitive-column detection and relationships are inferred; lineage is verified only with evidence.
  • No accuracy claim without reference data: DQaaS reports completeness, validity, uniqueness and similar checks, not accuracy.
  • In-browser mode: files are profiled on your own device and never leave it.

Read how your data is handled, or the security page for controls and their limits.

Cost

Pricing

Free during launch. All accounts are free, with these limits per month:

  • 1 GB of data scanned
  • 10 minutes of compute
  • 2 scheduled jobs
  • 50 MB of stored metadata
  • 1 GB of kept upload files (in total, not per month)

Pricing will be announced before any charge.

Status

Build status

As of 2026-11-19. We list what is released, what is still being built and what has not been confirmed outside our own tests.

  • Released
    • Version 0.4.0 of the command-line program and MCP server for Linux x86_64, Linux arm64, macOS on Apple silicon and Windows x86_64, with checksums and the licence files in each archive.
    • Hosted API, MCP endpoint and console are live, invite-only.
    • Automated tests pass on Linux, and each release build is run on its own platform before it is published; recent changes have been through independent security review with the findings fixed.
    • Stored credentials and backups are encrypted.
  • Built but not yet verified in the real world
    • Profiling in the browser is tested automatically, but not yet by a person in a real browser.
    • Hosted MCP sign-in has been completed with two real clients (Claude Code and Claude in the browser); other clients have not been tried.
    • Released on 2026-11-14 and tested automatically, not yet used with real accounts: two-person approval for accepting a risk, setting a classification, and deleting, switching off or lowering the severity of a rule (an organisation setting, off by default); failing-rule tasks that close only when a later scan passes; a failed or warning badge on the quality score; retry, re-upload and delete for a dataset whose first scan failed; a sheet picker for workbooks profiled in the browser; suggested glossary terms; a limit on how much one scan may write; unreadable lines in a delimited text file are skipped and reported by line number without their content; a hint when a text column is mostly numbers; an API route that tests a Postgres or MySQL connection, lists its tables and registers one dataset per chosen table.
    • Released on 2026-11-16, not yet used with real accounts: a Connect a database form on the upload page for one Postgres or MySQL table, with a connection test that lists the tables. It was tested end to end in a headless browser against real Postgres and MySQL test servers; certificate checking was not tried against a server that uses TLS.
    • Released on 2026-11-15 and tested automatically, not yet used with real accounts: a delimited text file whose layout is in doubt (lines skipped, a title line above the table, or two separators that both fit) is profiled with counts and types only, with generic column names where the header cannot be trusted; no value from any column is stored and the profile says so. A file that cannot be read reliably is refused. Profiling in the browser applies simpler versions of the two-separator and title-line checks (tested automatically, not yet seen in a real browser) and does not refuse such files.
    • Cron schedules, column-level lineage and metrics export are tested but not yet used with real accounts or a production monitoring stack.
    • GitHub sign-in is tested against a simulated provider and in the first live sign-ins only.
    • Account administration (suspending or removing members, exporting and deleting an organisation) is built and awaiting use by outside users.
    • Newer profiling and rule features are checked against reference queries but have not been used on real customer data: extra column statistics, aggregate and identifier rules, dataset comparison, table-level checks, tidy CSV export, configurable profiles, a quality scorecard, a comparison of declared against observed column metadata for database sources, a starter rule set applied to each new dataset so it gets a score after its first scan, and data hygiene advice (numbers or dates stored as text, stray spaces, placeholder values, probable typos and inconsistent column names).
    • Invitations can be sent by email; delivery to invitees has not yet been confirmed.
    • The Terms of Use, Acceptable Use Policy and Privacy Notice are published. They are not legal advice, and no compliance or certification is claimed.
    • Billing is built but switched off and has never been used with a real payment account.
    • The Windows command-line program is built and tested automatically on Windows, but has not yet been used on a person's own PC.
    • Publishing a local command-line result to the console without uploading the data file (counts and column structure only, marked as not verified by the server) is tested automatically but not yet used with real accounts.
  • Released on 2026-11-19
    • Live on the hosted service since 2026-11-19. Tested automatically, and where noted in a real browser, against test data; not yet used with real accounts.
    • More sources: zstd-compressed CSV, TSV, JSON and JSONL files; a zip archive that holds exactly one data file (not in browser mode); Dropbox links, GitHub file pages and Google Sheets share links are fetched as the file itself; a public open-data resource link is loaded with a row limit instead of being cut at 1,000 rows. A sheet that is not shared is refused with a clear message.
    • REST APIs as a source: give a URL, header values (stored encrypted and never shown back), a path to the records and a paging style (offset, cursor or link). The Connections page has a form for it and the command-line program has an add-rest command. Paging that does not finish within the page limit is an error, never a silent cut. The form was tested only in a simulated page, not in a real browser, and no real third-party API has been connected yet.
    • Object storage: a form on the Connections page for AWS S3, Cloudflare R2 and other S3-compatible services. It was tested only in a simulated page.
    • Account: a header menu with your profile, settings, organisation switcher and sign out; a Profile page to edit your display name and notification address and to delete your account by typing your login. The menu opens, closes and fits a phone-width screen in a real browser; the GitHub picture and the organisation switcher were not tried there.
    • Console: separate Files and Connections pages. Files takes up to 25 files at once, each scanned as its own dataset, and for an embedded database file you can tick several tables, one dataset per table (the table with the most rows is ticked first). Connections lists a Postgres or MySQL database's tables with a search box and scans each ticked table as its own dataset; row counts there are the database's own estimates.
    • More files read correctly: decimal-comma numbers, a first row that is data rather than a header, stray quote marks in a CSV, gzip files without a .gz name (with a cap on decoded size), and Arrow stream files. The dataset page names the table read and the tables skipped.
    • Data contracts: import is strict about duplicate keys, nesting and size on the API, MCP and in the browser; rules imported from a contract are labelled unverified, and an inferred-provenance marker survives an export and re-import; rule settings ODCS cannot express are kept verbatim on export rather than dropped.
    • Safety limits: more reserved network address ranges are refused for remote sources; database connection tests are limited per organisation and pinned to the address checked; files kept for later scans are deleted on a retry queue that survives a failed deletion after an erasure.
    • Numbers: column statistics fall back statistic by statistic when a value overflows, and drift between two profiles whose columns are single constant values is exact, in the API and in the browser.
    • Local command-line and MCP runs: input is confined to the chosen folder, oversized input and MCP messages are refused, and standalone runs send no webhooks.
    • Incidents: a failing critical or high rule is one incident per rule, with the measured pass rate against its minimum, the rows that failed, the last scan in which it passed, and how many datasets sit downstream in lineage you declared (labelled inferred and incomplete). You can acknowledge it or mute its webhook notices for up to 30 days; a scan closes it only when that rule passes. With a webhook set, one extra signed message per scan lists what opened and what recovered. Tested automatically, and the buttons in a real browser, with test data only.
    • Suggested lineage: from the profiles you already have, the product proposes links between your own datasets (files and database tables) where one dataset's key columns appear in another, or where they share most column names, each labelled inferred with its measured evidence and a guessed direction. Nothing is created until you accept it, an accepted link stays inferred, and a dismissed one is not proposed again. Matches by chance are expected and ranked low. Available in the console, API, CLI and MCP; tested automatically and in a real browser, with generated data only.
  • Not built
    • Paid plans: not offered yet. All accounts are free during launch.
    • Any certification or compliance attestation.

Next

Roadmap

  • Roadmap

    Paid plans through Paddle, after testing with a real Paddle account

  • Roadmap

    Single sign-on beyond GitHub

  • Roadmap

    OpenTelemetry traces and logs (metrics export exists)

FAQ

Frequently asked questions

What is DQaaS?

DQaaS is a data quality service. It profiles datasets, checks them against rules, scores their quality and keeps a governed catalogue with lineage, reports and an audit log. You can use it from a web console, a REST API, a command-line program, an MCP server for AI assistants, or in the browser on your own device.

Can I try it without an account or a server?

Yes. The command-line program can profile and check a file on your own machine and write a Markdown, HTML, PDF or JSON report. It uses temporary disk copies, which may remain if cleanup fails or is interrupted. Review configured storage and network connections; --publish sends profile and rule metadata. Run dqaas run customers.csv -o report.html.

Which platforms does the command-line program support?

Linux x86_64, Linux arm64, macOS on Apple silicon and Windows x86_64. The Windows build is tested automatically on Windows but has not yet been used on a person's own PC.

Does it work with monitoring tools, OpenLineage and data contracts?

Yes, within limits. It serves per-dataset quality figures to monitoring tools (setup is in the docs), exports and ingests OpenLineage events, and imports and exports the rule parts of Open Data Contract Standard (ODCS v3.2.0) contracts. Anything that cannot be mapped is listed, not guessed.

Does DQaaS store my data?

Not your rows, unless you ask. On upload, source rows are processed for the job and deleted when it ends; an upload is kept for later scans only while Keep file is ticked, which is the default in the console. Metadata is persisted: profiles, rule results, scores, catalogue entries and audit records. Sensitive values are redacted, and strict privacy mode stores no sample values at all.

Can I check a file without uploading it?

Yes. Attach a CSV, TSV, JSON, JSONL, Parquet, Avro, XLSX, Arrow or embedded database file and it is profiled on your own device; the file never leaves it. Those results are computed locally and are not verified by the server. You can publish just the profile and rule results to the catalogue.

Which formats are supported?

Upload: CSV, TSV, JSON, JSONL, Parquet, Avro, Arrow, XLSX and embedded database files, plus HTTPS URLs and read-only Postgres tables. On your own device: CSV, TSV, JSON, JSONL, Parquet, Avro, XLSX, Arrow and embedded database files (pick which tables).

How is quality scored?

Each rule counts failed rows against total rows, giving a score from 0 to 1 that passes at its threshold. Rule scores are weighted by severity and grouped by quality dimension into an overall dataset score.

Is accuracy measured?

No. Accuracy needs trusted reference data, which DQaaS does not have. It reports completeness, validity, uniqueness and similar checks, and makes no accuracy claim without reference data.

How reliable is sensitive-data detection?

It is heuristic, based on column names and value patterns, so it can miss or misflag columns. Findings are labelled inferred. Use strict privacy mode if a miss is unacceptable.

What does it cost?

Nothing during launch. All accounts are free with monthly limits (1 GB scanned, 10 minutes of compute, 2 scheduled jobs, 50 MB of stored metadata), and up to 1 GB of kept upload files. Pricing will be announced before any charge.

How do I sign in?

Sign in with GitHub on the sign-in page. Access can also be by an access token for automation.

How do I connect an AI assistant?

The command-line program includes a local MCP server, and there is a hosted MCP endpoint that uses your token or OAuth sign-in. The docs list the tools. They cover dataset listing, profiling, rule suggestion and acceptance, drift, reports, contracts, glossary, stewardship tasks, schedules, lineage and impact analysis.

Is DQaaS certified or compliant with a standard?

No certification or compliance attestation is claimed. See the security page for the controls that exist and their limits.