misata-mcp
Misata MCP server designs, generates, validates, queries, audits, and attests synthetic datasets, and can seed real databases.
List built-in business domains and preview/inspect the tables and schema implied by a natural-language story without generating rows.
Generate reproducible CSV datasets from a sentence or a detailed schema, including keys, foreign keys, constraints, distributions, correlations, time series, state machines, outcome curves, and other declarative guarantees.
Validate
misata.yamlfiles structurally and semantically, and explain what a schema will generate, including planned rows, conflicts, invariants, and warnings.Create an isolated SQLite sandbox, seed it, and run read-only SQL queries against it.
Seed Postgres or SQLite databases safely: plan first by default, then apply/truncate/append with FK integrity verification.
Audit existing CSV datasets for contradictions, check for synthetic-data tells, and validate values against clinical/financial domain rules.
Attest a dataset against its schema: recompute invariants with DuckDB, pin files by SHA-256, and write in-toto/DSSE attestations plus a standalone
verify.py.Solve outcome curves so SQL metrics over generated data hit declared targets within a tolerance.
Misata
Declare what must be true. Get data that makes it true. Prove it.
Misata is a declarative synthetic data generator for Python. You write down the facts a dataset has to satisfy, from a monthly revenue curve to a balanced ledger to a foreign key that must never dangle, and the engine builds multi-table data that satisfies all of them at once. Then DuckDB checks the files on disk and writes a certificate anyone can re-run.
Docs · Use cases · The language · Limitations · Studio · Changelog
Sixty seconds
pip install "misata[duckdb]"# misata.yaml
seed: 7
tables:
customers:
rows: 2000
columns:
customer_id: {type: int, unique: true, min: 1, max: 2000}
name: {type: text, text_type: name}
signed_up: {type: date, start: "2024-01-01", end: "2024-12-31"}
orders:
rows: 40000
columns:
order_id: {type: int, unique: true, min: 1, max: 1000000}
customer_id: {type: foreign_key}
channel: {type: categorical, choices: [web, app, store]}
amount: {type: float, min: 5, max: 900}
ordered_at: {type: date, start: "2025-01-01", end: "2025-12-31"}
relationships:
- customers.customer_id -> orders.customer_id
outcome_curves:
- table: orders
column: amount
time_column: ordered_at
start_date: "2025-01-01"
curve_points:
- {month: 1, target_value: 180000}
- {month: 6, target_value: 240000}
- {month: 12, target_value: 310000}
group_shares:
- table: orders
measure: amount
group_column: channel
shares: {web: 0.5, app: 0.3, store: 0.2}misata generate -c misata.yaml -o data --attest
python data/verify.pyAttestation: data
result: PASSED
files pinned: 2
invariants held: 69 of 69
ok 2 file(s) match their attested sha256
ok 69 of 69 invariant(s) holdThe 69 are derived from the YAML before a row exists: twelve monthly revenue
totals (180,000.00 in January, 240,000.00 in June, 310,000.00 in December,
straight lines between), the channel split inside every one of those months,
the planned row count per month, zero orphaned customer_ids and unique keys.
verify.py needs only the standard library and duckdb, and it shares no
code with the generator.
The YAML above, generated and summed by month and channel. Each bar is a month of generated revenue split by channel; the cap is the total the YAML declares. Drawn by readme_charts.py from a fresh run of this README's own YAML.
Change one cent in one row and it says so:
FAIL orders.csv: sha256 301165f9dc4e does not match the attested 670f89195826
FAIL orders.amount.month:1.outcome_curve_period: expected 180000.00, observed 180000.01
FAIL orders.amount.outcome_curve_total: expected 2945000.00, observed 2945000.01
FAIL orders.amount.web.2025-01.group_share_period: expected 90000.00, observed 90000.01
FAIL orders.amount.web.group_share_total: expected 1472500.00, observed 1472500.01
FAIL 65 of 69 invariant(s) holdRelated MCP server: MCP Data Visualization Server
Why declare instead of generate
Every test fixture has an oracle problem. To check a dbt model, a Spark job, a dashboard tile or a data agent you need input data and an answer you know independently of the code under test. Copying production gives you data without an answer. Hand-written fixtures give you an answer that rots. Random generators give you data whose answer you have to compute with the code you are testing.
Misata turns it around. The specification is the ground truth and the rows are derived from it. "January revenue is 180,000" is true before a single row exists, so your pipeline either returns 180,000.00 or it has a bug.
flowchart LR
A["Declare<br/>curves, shares, ledgers,<br/>keys, lifecycles"] --> B{"Can they<br/>all hold?"}
B -- "no" --> R["Refused, with the<br/>arithmetic and a fix"]
B -- "yes" --> C["Build rows<br/>closed form where it can,<br/>by construction elsewhere"]
C --> D["Write<br/>CSV, Parquet, JSONL,<br/>DuckDB, Postgres"]
D --> E["Attest<br/>DuckDB re-runs every<br/>derived invariant"]
E --> F["Certificate<br/>in-toto Statement<br/>plus verify.py"]Three properties make this different from a generator with options:
A declaration is honoured or refused, never quietly replaced. Declarations that cannot all hold stop generation with an error that names each one, shows the arithmetic and suggests the change.
The certificate is checked by code that did not write the data. A bug in Misata's arithmetic shows up as a failed invariant, not a self-consistent wrong answer.
The same spec, seed and version give the same bytes on Linux, macOS and Windows, so a fixture can live in git with its fingerprint.
Where this is going
Misata's aim is that nobody writes a Python script to make a dataset, of any kind. You say what has to be true; the engine works out the rows, refuses what cannot hold, and proves what it built.
It is not there yet, and it is getting there brick by brick. Each release adds things you can declare instead of programming, and a brick is laid only when it passes three tests: it can be declared in a few lines of YAML, it is refused before generation when it contradicts another declaration, and it is re-checked on the rows afterwards by code that did not write them.
Every brick is something you declare instead of code. Drawn from misata/registry.py, the list a contract test holds the code to, so the wall cannot claim a brick the engine does not have.
One fact, two ways. What a script has to do to make a fact true, and what you write instead:
The fact | In a generator script | In Misata |
January revenue is 180,000.00 | draw amounts, sum them, rescale, round, chase the last cent |
|
web is half of every month's revenue | rescale inside each month and channel, then repair the monthly totals you just broke |
|
every order has a real customer | sample ids from the parent table and keep both in step |
|
some churned customers were active first | branch on status, write timestamps in order, null the rest, count by hand |
|
1% of transfers form laundering patterns, in a set mix | build each ring and fan by hand, then prove the background has no cycle of its own |
|
the trial balance is zero | post both legs of every entry and reconcile at the end | a |
What still takes code, and where each one stands:
Still scripted by hand | Where it stands |
The expected output of logic Misata cannot invert: window functions, custom SQL, business rules |
|
Correlations you can only describe by example |
|
Free text with a voice of its own | text is generated from grammars and agrees with its row; prose that reads as written is not declarable yet |
Version histories (SCD2) and inventory trajectories | generated to balance, not yet declared step by step |
Documents, images and audio | not started |
LIMITATIONS.md keeps the full list. When a line leaves this table, it is because the brick that replaces it has passed all three tests.
The evidence
The Gauntlet: 13 tables and 136 SQL assertions run by DuckDB. The Faker script ties wherever a fact is about one column at a time and loses wherever columns, rows or tables have to agree. Source, rerun with python -m benchmarks.gauntlet.
A gradient-boosted classifier tries to tell generated rows from real ones; mean of five seeds. Without seeing a real row, a declared schema beats SDV trained on the real orders. On taxi trips SDV wins: blind fares are a miss. Source, numbers in summary.json.
Claim | Measured | Reproduce |
Valid by construction | 0 violations of keys, CHECK, UNIQUE, NOT NULL and self-reference trees on five DDL schemas. A per-column Faker script: 15,014 to 50,416 per schema. SDV's multi-table model trained on valid data: 3,055 to 5,557, and it refuses two of the five |
|
Conformance | 345 SQL assertions in four suites, all held: the Gauntlet 136, the Warren 110, the Ledger 48, the Borough 51 |
|
Reproducible | identical fingerprints on Linux, macOS and Windows |
|
Scale | 10,000,000 rows across 14 tables in 36.3 s at 0.78 GB peak on an Apple M1, every foreign key checked afterwards |
|
What you can declare
Each declaration has a guarantee stated so you could test it, a list of the columns it owns, a cost and a refusal. LANGUAGE.md is the reference. All 25 declaration kinds are checked for feasibility before a row exists and re-checked on the rows afterwards; the ones marked ✓ are also in the DuckDB certificate.
Declare | What holds | Certificate |
| a measure's total per period, to the cent | ✓ |
| the share of rows flagged per period, to the nearest row | ✓ |
| each group's share of a measure inside every period | ✓ |
| debits equal credits per entry, and the trial balance is zero | ✓ |
| opening plus movements equals closing, per component | ✓ |
relationships, | no orphans, no childless parent, no key crossing a tenant | ✓ |
| no row breaks the rule | ✓ |
| a status implies its whole history: the right timestamps set, in order, and the rest null | ✓ |
| no cycle outside a planted case; each ring, fan and chain at its exact count | ✓ |
| inventory balances, chains period to period, and opens where declared | ✓ |
| a child, or a parent's children together, stay under the parent's bound, without moving a declared total | ✓ |
| a declared clock; a child after its parent existed, and inside its span | ✓ |
| an entity in one place at a time; no resource holding more than it can at any instant | ✓ |
roll-ups, | parents equal their children, versions tile a timeline | |
| dirt at exact counts, with an answer key | |
| cohort curves, joint margins, logs that agree with status |
Underneath any declaration, foreign keys resolve by construction (parents are built before children), text agrees with its row, money is skewed, timestamps have a daily rhythm and a few customers place most of the orders, so the data looks like a business before you declare anything. A sentence picks one of 18 story domains, from SaaS and fintech to healthcare, logistics and streaming, each with its tables, keys and calibrated distributions.
Pick your entry point
You have | Call | Notes |
A sentence |
| fastest start; print the tables or |
YAML or a dict |
| every declaration holds, or it refuses |
SQL DDL |
| NOT NULL, UNIQUE, composite keys, CHECK ranges and enums, self-references |
A live database |
| introspects, inserts in one transaction, never truncates unless you pass |
A dbt project |
| fixtures from |
Prisma or Django |
| |
More rows than memory |
| yields |
A pipeline whose output must hit a number |
| drives inputs through your real dbt or Python pipeline to a declared outcome, within a stated tolerance |
A dataset to share |
| re-verifies, then uploads the data with its spec, certificate and verifier |
Real rows to imitate |
| fitted to real rows; not anonymous |
Pass a seed every time. The same seed is the same data.
Twelve recipes
Each is a folder with the schema, the commands, and the check that proves the
declaration held. tools/stranger.py --clean runs every command on every page
from a freshly built wheel in a clean virtualenv before each release, so the
numbers on the pages are what you will see.
Recipe | For | Proof |
app developers, coding agents | 22 invariants from three | |
analytics engineers | a dbt test that catches the last-day-of-month bug | |
founders, data leads | 291,000 rows of partitioned Parquet, 80 invariants | |
sales engineers, PMs | churned and healthy accounts, a few whales; realism 0.96 | |
AI labs, eval teams | text-to-SQL answers verified on three rotated databases | |
agent builders | the sandbox and the grader agree row for row | |
platform engineers | hot keys and a daily curve, keys checked before replay | |
risk and ML teams | an acyclic background and every ring at its declared count | |
finance engineering, audit | two fiscal years, 52 invariants, trial balance zero | |
hiring teams, course authors | a different answer key per candidate, each verified | |
regulated teams | staging data from production DDL, never production rows | |
DX, API and SDK teams | nested JSONL whose ids resolve across files |
A status that implies its history
Declare the state machine once and every row's timestamps follow from its status. From the demo tenant:
stateDiagram-v2
direction LR
[*] --> trialing
trialing --> active
trialing --> churned
active --> past_due
active --> churned
past_due --> churnedlifecycles:
- name: subscription
table: subscriptions
state_column: status
start_column: started_at
initial: trialing
states:
- {name: trialing}
- {name: active, timestamp: activated_at}
- {name: past_due, timestamp: past_due_at}
- {name: churned, timestamp: churned_at, terminal: true}
transitions:
- [trialing, active]
- [trialing, churned]
- [active, past_due]
- [active, churned]
- [past_due, churned]
weights: {trialing: 0.07, active: 0.6, past_due: 0.07, churned: 0.26}Of 1,200 subscriptions, exactly 720 are active, 312 churned, 84 past due and 84 still trialing: weights are counts, allocated by largest remainder, not probabilities. Every past-due row has an activation date and a past-due date, in that order and after it started; no active row has a churn date. The certificate checks both, row by row. A state reachable several ways splits its rows across every route: of the 312 churned, 156 left during the trial, 78 after being active and 78 after falling past due.
The demo tenant's 1,200 subscriptions, grouped by which timestamps are set. Six routes, and no row outside them. Drawn by readme_charts.py from the recipe's own YAML.
When it says no
$ misata explain bad.yaml
Generation would refuse: 1 conflict(s).
1 declaration cannot be satisfied together:
1. [orders.channel] group_shares on amount
declared shares sum to 1.500 (web=0.6, app=0.6, store=0.3); a share of a whole cannot exceed 1.0
→ Suggestion: scale the shares so they sum to 1.0, or drop a group. Misata will not renormalise them for you, because the result would be a specification you did not write
Nothing was generated. Change the declarations each conflict names; its remedy says how.misata explain prints the planned rows, what each declaration will write,
how many invariants the certificate will carry and every conflict, without
generating anything. Every command takes --json and prints one envelope;
exit code 3 means the declarations cannot all hold and nothing was written.
With a coding agent
The agent designs the schema, Misata does the arithmetic, and DuckDB checks it. An agent cannot certify its own output; this is how it gets an answer it did not write.
claude mcp add misata -- uvx --from "misata[mcp]" misata-mcp # Claude Code
misata mcp install --client all # Cursor, Windsurf, Claude DesktopFifteen tools: generate_from_schema, generate_dataset, explain_schema,
validate_yaml, inspect_schema, preview_story, list_domains,
validate_domain, seed_database, create_sandbox, query_sandbox,
solve_outcome, attest_dataset, audit_dataset and check_realism.
AGENTS.md tells an agent which entry point fits which job and
which claims the evidence supports.
What it is not
Misata imitates nothing unless you ask it to. It does not learn a joint
distribution from your data. If you need a statistical twin of a specific
private table, use a differentially private synthesizer with a stated
epsilon. mimic fits to real rows and is not anonymous.
It is exact for what can be declared (sums, rates, shares, roll-ups,
ledgers, keys, lifecycles) and not for arbitrary SQL. A model built on a
window function or business logic Misata cannot invert still gets coherent,
key-valid inputs, and you write the expected output. misata solve closes
part of that gap by driving inputs through your real pipeline to a declared
outcome within a stated tolerance, and the certificate says which tolerance.
Priors are English-first and US-leaning on money, and rates round to whole rows. LIMITATIONS.md has the full list and wins any argument with the rest of the docs.
Imperative | Inferential | Declarative | |
You give it | code that emits values | real data to fit | facts that must hold |
It guarantees | nothing after the fact | resemblance, approximately | the facts exactly, or a refusal |
Foreign keys | your problem | per table fine, across tables fragile | by construction |
Needs real data | no | yes | no |
Examples | Faker, Mockaroo, LLM seed scripts | SDV, learned synthesizers | Misata |
Cite
If Misata helps your work, cite the software. GitHub's Cite this repository button reads the same metadata from CITATION.cff.
@software{rasin_misata,
title = {Misata: declarative synthetic data},
author = {Rasin, Muhammed},
url = {https://github.com/rasinmuhammed/misata}
}Contribute
git clone https://github.com/rasinmuhammed/misata && cd misata
uv sync --all-extras && uv run pytest -q
python tools/stranger.py --clean # every documented command, from the built wheelA capability earns a place only if it can be declared and verified. The
conformance suites in benchmarks/ are the contract: write the assertion
first, watch it fail, then make it pass. MIT licensed.
Available Tools
15 toolsattest_datasetAttest a dataset against its schemaAIdempotent
Prove a written dataset holds what its schema declared, in a standard format.
Recomputes every invariant the schema implies (outcome-curve totals, rate
anchors, group shares, ledger identities, foreign keys, unique columns)
with DuckDB against the files, and pins each file by SHA-256. Writes
attestation.intoto.json (an in-toto v1 Statement that cosign and
GitHub's actions/attest accept), attestation.dsse.json and
verify.py into dataset_dir. Anyone can rerun python verify.py
with only duckdb installed.
Use it after generate_from_schema with the same schema and seed, and
report result to the user: PASSED means every declared property
holds in the files on disk.
Args:
dataset_dir: Folder with one <table>.csv or <table>.parquet
per table (the output_dir of generate_from_schema).
schema: The same schema dict passed to generate_from_schema.
seed: The seed used to generate, recorded in the Statement.
max_invariants: How many invariants to include in the response.
Returns:
{"ok", "result", "invariants_held", "invariants_total", "failures", "subjects", "files", "invariants"}.
Inputs: schema (required), dataset_dir (required), seed, max_invariants. Returns: result
audit_datasetAudit a dataset for reader-visible contradictionsARead-onlyIdempotent
Score a folder of CSVs for the contradictions a human reader would catch.
This is Misata's coherence audit run on data that already exists — data an agent generated in an earlier step, data a user built by hand, or the output of some other tool. It checks, among other things:
timestamps that run backwards (shipped before ordered, resolved before opened),
derived columns that do not reconcile with their inputs (
total!=quantity * unit_price),geographic fields that disagree (city / state / postcode / country),
near-constant columns (98% one value — a distribution tell),
filler text and out-of-scale numerics.
A score of 100 is clean. Below ~85 usually means the schema is missing
realism structure, not that individual rows need patching: add
__correlations__, profiles, time_series, a __state_machine__,
or an __outcome_curves__ declaration and regenerate.
Args:
dataset_dir: Directory containing one CSV per table (e.g. the
output_dir returned by generate_from_schema).
top_findings: Max findings to include in the response (default 20).
Returns:
{"score": 0-100, "clean": bool, "counts": {...}, "findings": [...]}.
Inputs: dataset_dir (required), top_findings. Returns: result
check_realismCheck a dataset for the tells of synthetic dataARead-onlyIdempotent
Find the statistical shapes that make a dataset look generated.
Where audit_dataset finds contradictions, this finds shapes real data
almost never has, with no real data to compare against: money spread
evenly from min to max, every customer with about the same number of
orders, perfectly balanced categories, flat weekday or hour-of-day
profiles, timestamps piled up at midnight, placeholder values
(@example.com, John Doe), emails unrelated to names, small name pools,
templated or repetitive free text, prose with a tiny vocabulary, review
text whose tone ignores its rating, descriptions unrelated to their
titles, tables with no nulls at all. Works on data from any generator,
including a script you wrote.
Fix a finding in the schema, not the rows: lognormal for money, a curve or
weights for time, sampling on foreign keys, null_rate or
__missingness__ for optional fields. Pass skip for tells that are
intended (a randomised A/B split is balanced on purpose).
Args:
dataset_dir: Directory with one CSV per table (the output_dir from
generate_from_schema).
skip: Check names to leave out, e.g. ["too_clean"].
Returns:
{"score": 0-1, "passed": bool, "counts": {...}, "tells": [...]},
each tell with its check, table, column, status and evidence.
Inputs: dataset_dir (required), skip. Returns: result
create_sandboxCreate an instant sandbox databaseAIdempotent
Create and seed a real, isolated SQLite sandbox database for testing.
Perfect for AI coding agents: spins up a queryable database with realistic, referentially intact tables (0 orphan foreign keys) so you can test SQL queries, analytics logic, or application code immediately without touching a real database.
Args: domain: Pre-built domain template ('saas', 'ecommerce', 'fintech', 'healthcare'). story: Plain-English description (e.g. 'A restaurant with tables, bookings, and guests'). schema: Explicit schema dict if you want exact column control. rows: Base row count (default: 100). Transactions scale proportionally. seed: Random seed for 100% reproducible results (default: 42). db_path: Target SQLite file path (default: '.misata/sandbox.db').
Returns: Dictionary with db_url, db_path, table summaries, DDL, and sample verification queries.
Inputs: rows, seed, story, domain, +2 more. Returns: result
explain_schemaExplain what a schema will generateARead-onlyIdempotent
Show what generate_from_schema would do with a schema, without generating.
Call it before generate_from_schema on a schema you designed: it is
cheap, and it catches the mistakes that would otherwise surface after a
long run. Returns each table's planned row count and approximate memory,
the declarations and the columns each one writes, any column two
declarations both write, every outcome curve's period plan and warnings,
and the invariants attest_dataset will check on the output. When
will_generate is false, conflicts names each contradiction with
the arithmetic and a remedy; fix those in the schema and call again.
Args: schema: The schema dict you would pass to generate_from_schema. seed: The seed you would generate with. max_invariants: How many invariants to include in the response.
Returns:
{"ok", "will_generate", "total_rows", "estimated_bytes", "tables", "relationships", "declarations", "writers", "overlaps", "conflicts", "conformance", "invariants", "invariants_total"}.
Inputs: schema (required), seed, max_invariants. Returns: result
generate_datasetGenerate a dataset from a sentenceA
Generate a synthetic dataset from a story and write it to disk as CSV files.
Returns the output directory, file paths, row counts per table, and a small sample of rows for each table so the agent can show the user what was produced without loading every row into context.
Args: story: Plain-English description of the dataset. rows: Default row count for the primary table. seed: Optional random seed (same seed → byte-identical output). output_dir: Where to write CSVs. Defaults to a fresh temp dir. sample_rows: Number of rows from each table to include in the response (max 50).
Inputs: story (required), rows, seed, output_dir, +1 more. Returns: result
generate_from_schemaGenerate a dataset from a schemaAIdempotent
Generate a dataset from a schema you design. This is the primary Misata tool.
DIVISION OF LABOUR You (the agent) design what the data should look like — the tables, columns, business rules, and declared targets. Misata handles the hard guarantees: every FK resolves, every rollup reconciles to the cent, every declared outcome curve is hit exactly, every seed produces byte-identical output. The response includes a per-relationship integrity proof.
SCHEMA FORMAT {table_name: {column_name: spec, ...}, ...}
TABLE-LEVEL KEYS (inside a table dict) "rows": 5000 Per-table row count; overrides the global rows arg. Always set this — one global count rarely fits all tables. "constraints" List of row-level business rules (see CONSTRAINTS below). "correlations" List of pairwise Pearson targets (see CORRELATIONS below). "state_machine" Markov terminal-state assignment (see STATE MACHINE below).
COLUMN TYPES integer, float, decimal Numeric. string Short categorical or free text. text Long free text (descriptions, notes). email, phone, url, uuid Semantic strings; always valid format. date, datetime Temporal; realistic granularity applied automatically. boolean True/False with declared probability. object (json) Nested object: {"type": "object", "fields": {"email": "email", "age": {"type": "integer", "min": 18}, "tags": {"type": "list", "items": "string"}}}. Optional: "optional": [...], "optional_rate". list (array) {"type": "list", "items": , "min_items": 0, "max_items": 5, "unique_items": true, "sorted": true}. Values are JSON text.
COLUMN SPEC KEYS (inside a column dict) primary_key: true Auto-incremented PK; column excluded from CSV output. foreign_key: {table, column} Child FK; referential integrity guaranteed + verified. min / max Numeric or date bounds. decimals: 2 Decimal places for float output. unique: true All values in the column are distinct. nullable: true Allow nulls (default true). enum: [...] Categorical choices. Add probabilities: [...] for weights; omit for realistic Zipf-shaped rank frequencies. probabilities: [...] Weights for enum choices; must sum to 1.0.
DISTRIBUTIONS (float / integer columns) distribution: normal Also: lognormal, uniform, exponential, gamma (shape, scale), beta (a, b), poisson (lambda), binomial (n, p), power_law (alpha). mean / std Normal params. Can be a scalar OR a per-row parent-entity lookup: mean: {formula: "@patients.hba1c_baseline"} The FK is resolved per-row so each child row's distribution is anchored to its parent's value. Use this for longitudinal data where within-entity variation should be modelled separately from between-entity variation. mu / sigma Lognormal params (mu and sigma are of log(x)). min / max Hard clamps applied after sampling. Use lognormal for money, file sizes, session durations — anything right-skewed and strictly positive. Use normal for measurements.
DERIVED COLUMNS formula: "quantity * unit_price" Row-level arithmetic; pandas eval syntax. formula: "hours * @employees.rate" Cross-table via FK: @parent_table.column. rollup: {from_table, fk, agg, column} Parent column that EXACTLY reconciles with child rows under JOIN. agg: sum/count/mean/ max/min. Add where: {col: val} to filter. RULE: use rollup (not formula) for any parent column that summarises child rows. Rollups are closed-form exact; formulas cannot cross the FK boundary correctly.
CODE-STYLE STRINGS pattern: "SKU-\d{5}" Single pattern expanded per row. pattern: ["A/\d{5}", "\d{6}"] List: one shape drawn per row. pattern_weights: [0.7, 0.3] Weights for pattern list (optional). Supported tokens: \d (digit), [A-Z] (uppercase letter), [a-z] (lowercase), literal chars, {n} repeat count. Example: "[A-Z]{2}-\d{4}" → "AB-3721".
TEXT SEMANTICS text_type: person_name Always beats column-name inference. Options: person_name, email, company, city, country, postal_code, phone, url, description, username, product_name, review_text, address, job_title. Dates: appointment times snap to 15-min business-hours grids; signups follow waking-hour rhythms; machine events keep sub-second precision. Names, genders, and emails are generated jointly and always agree.
STRATIFIED DISTRIBUTIONS (profiles) Use when different subgroups need different distributions for the same column. profiles: [ {when: "arm == 'placebo'", distribution: normal, mean: -0.35, std: 0.50}, {when: "arm == 'high_dose'", distribution: normal, mean: -1.25, std: 0.55}, ] Rows that match no profile get the column's top-level distribution. The when expression is a pandas eval string; reference any already-generated column in the same table. Always list profiles after the columns they reference.
INFORMATIVE MISSINGNESS (MAR) null_when: "dropout == False" Null this column when expression is true. missing_if: Missing-At-Random tied to a predictor column. predictor: hba1c_baseline relationship: higher_increases_probability # or lower_increases_probability base_rate: 0.05 # null probability at predictor median max_rate: 0.40 # null probability at predictor extreme Use null_when for status-conditional nulls (dropout_visit is null when not dropped out). Use missing_if when missingness is correlated with an observed variable.
EXACT INCIDENCE CONTROL exact_incidence: Hit the declared count exactly (not approximately). mode: exact rate: 0.22 # exactly floor(n * 0.22) rows become True group_by: arm # optional: apply per group rates: {placebo: 0.15, high_dose: 0.55} # per-group exact rates Use exact_incidence instead of probability on boolean columns when the user states a precise rate that must hold in the data, not just on average.
WITHIN-ENTITY TIME SERIES (longitudinal autocorrelation) time_series: Re-writes a column to have AR1 autocorrelation entity_id: patient_id within each entity group. order_by: visit_number model: AR1 # AR1 | linear_trend | random_walk | mean_reversion phi: 0.72 # autocorrelation coefficient (AR1 only) noise_std: 0.30 anchor_column: hba1c_baseline # starting value (column in the same table) trend: slope_mean: -0.08 # mean drift per step slope_std: 0.02 # per-entity slope variability Required for any longitudinal dataset (clinical visits, IoT sensors, user sessions). Without it every row is independent and the data fails any time-series test.
CONSTRAINTS (table-level constraints list) {"type": "inequality", "column_a": "visit_date", "operator": ">=", "column_b": "enroll_date", "action": "cap"} Enforces column_a OP column_b. action: "cap" (snap column_a to column_b) or "drop" (remove violating rows). Works on dates and numerics. {"type": "col_range", "low_column": "min_price", "column": "price", "high_column": "max_price", "action": "cap"} Keeps low_column <= column <= high_column. {"type": "max_per_group", "group_by": "user_id", "max_count": 3} Limits rows per group value. {"type": "unique_combination", "columns": ["user_id", "product_id"]} No duplicate (col_a, col_b) pairs. Use constraints for any business rule that must hold on every row: visit_date >= enrollment_date, price > cost, resolution_day > onset_day.
CORRELATIONS (table-level correlations list) [{"col_a": "bmi", "col_b": "systolic_bp", "r": 0.41}] Enforced via Iman-Conover (rank reordering): preserves each column's marginal distribution while hitting the declared Pearson r exactly. Declare correlations for any pair of measurements that co-vary in the real domain (bmi/bp, income/spending, tenure/salary). Also supports full matrix syntax: correlations: matrix: columns: [hba1c, glucose, bmi] values: hba1c: [1.00, 0.65, 0.28] glucose: [0.65, 1.00, 0.22] bmi: [0.28, 0.22, 1.00]
ICC CLUSTER EFFECTS (parent table cluster_effect) cluster_effect: affects_table: visits affects_columns: hba1c: icc: 0.18 # intraclass correlation coefficient sd_total: 1.5 # total standard deviation; sd_between = sqrt(icc)*sd_total systolic_bp: sd_between: 8.0 # supply sd_between directly if preferred Applies per-parent-entity random intercepts to the named child columns. Required for multi-site or multi-centre designs — without it all sites look identical and any ICC statistical test will detect the synthetic origin. icc: 0.10-0.30 is typical for clinical measurements across sites.
STATE MACHINE (table-level state_machine) state_machine: state_column: patient_status initial_state: enrolled transitions: enrolled: {on_treatment: 0.97, screen_failure: 0.03} on_treatment: {completed: 0.77, dropout: 0.23} Assigns one terminal state to every row by following the Markov chain. States with no outgoing transitions are terminal. Use for any process with defined states: clinical trial statuses, customer lifecycle, order fulfilment stages, support ticket resolution.
SCHEMA-LEVEL DIRECTIVES (top-level keys, siblings of the tables)
outcome_curves Declare aggregate targets the engine hits EXACTLY. [{"table": "orders", "column": "amount", "time_column": "order_date", "time_unit": "month", "value_mode": "absolute", "start_date": "2024-01-01", "avg_transaction_value": 120.0, "curve_points": [ {"month": 1, "target_value": 50000.0}, {"month": 6, "target_value": 110000.0}, {"month": 12, "target_value": 200000.0} ]}] ALWAYS use this when the user states what a number should sum to per period: "revenue grows from $50k to $200k", "Q4 spike", "10x growth". avg_transaction_value drives row count per period; set it to roughly the median row value for that column.
rate_curves Per-period rate targets for boolean/categorical columns. [{"table": "transactions", "column": "is_fraud", "time_column": "transaction_date", "rate_points": [ {"period": "2024-01", "rate": 0.02}, {"period": "2024-Q4", "rate": 0.05} ]}] Use when fraud rate, churn rate, or conversion rate changes over time.
group_shares Exact shares of a measure across a categorical column. [{"table": "orders", "measure": "amount", "group_column": "plan", "shares": {"Starter": 0.2, "Pro": 0.5, "Enterprise": 0.3}}] The measure sums to those proportions per group, exactly. Paired with an outcome_curves on the same table+measure, the split holds inside every declared period. Use for any "A is 40% of revenue, B is 35%..." statement.
waterfalls Movements that reconcile to declared running balances. [{"table": "mrr_movements", "starting_value": 100000, "points": [{"period": "2026-01", "ending_value": 106000}, ...], "inflow_shares": {"new": 0.7, "expansion": 0.3}, "outflow_shares": {"churn": 1.0}}] Use for MRR bridges, cash-flow statements, any "opening + inflows - outflows = closing" ledger that has to tie out.
stock_flows Per-unit inventory identity: closing = opening + received
shipped, enforced for every SKU across every period. Use for warehouse / inventory data where stock levels must be internally consistent. Add "opening": {"flour": 120, ...} to fix each SKU's first opening level, e.g. to continue from a previous run's closing stock.
no_overlap An entity is in one place at a time: its [start, end) intervals never overlap. [{"table": "visits", "key": "resident_id", "start": "arrived_at", "end": "left_at"}]
capacity A resource never holds more concurrent intervals than its capacity (a constant "capacity", or "capacity_column" on its table). [{"table": "visits", "resource": "place_id", "start": "arrived_at", "end": "left_at", "capacity_column": "seats"}] A relationship's "parent_end" keeps a child inside its parent's span: {"parent": "visits.visit_id", "child": "orders.visit_id", "parent_time": "arrived_at", "parent_end": "left_at", "child_time": "ordered_at"}
lifecycles A state machine with legal transitions (stricter than a table-level state_machine: illegal jumps are impossible, not just unlikely). [{"name": "order_flow", "table": "orders", "state_column": "status", "start_column": "placed_at", "initial": "placed", "states": [{"name": "placed"}, {"name": "paid"}, {"name": "shipped"}, {"name": "delivered"}, {"name": "refunded", "terminal": true}], "transitions": [["placed","paid"],["paid","shipped"], ["shipped","delivered"],["delivered","refunded"]]}]
missingness Why a value is missing, conditionally (schema-level MNAR). [{"table": "contacts", "column": "notes", "rate": 0.75, "else_rate": 0.05, "when_column": "is_active", "when_op": "==", "when_value": false}] The null rate is exact per branch. Use when missingness itself carries signal a cleaning step should be tested against.
processes How each case moves through steps over time, written out as an event-log table (one row per event: case key, step, activity, timestamp). Use for support tickets, claims, onboarding funnels, order fulfilment, anything where the user wants an event log, a process-mining dataset, rework loops or realistic step durations. [{"name": "ticket_flow", "cases_table": "tickets", "case_key": "ticket_id", "start_column": "opened_at", "initial": "opened", "transitions": {"opened": {"triaged": 1.0}, "triaged": {"resolved": 0.85, "escalated": 0.15}, "escalated": {"resolved": 1.0}, "resolved": {"closed": 0.9, "reopened": 0.1}, "reopened": {"triaged": 1.0}}, "dwell": {"opened": {"distribution": "lognormal", "mu": 0, "sigma": 0.8, "unit": "hours"}}, "final_state_column": "status"}] A state with no transitions row is terminal. Each row sums to 1. The event table is named "_events". Structure is exact and audited; path shares and durations are drawn from the declaration.
DIRTY DATA ON PURPOSE (exact counts, so a test has a known number to find) duplicates [{"table": "contacts", "count": 60}] typos [{"table": "contacts", "column": "city", "count": 120}] outliers [{"table": "orders", "column": "amount", "count": 40}] Each injects exactly that many defects, leaving primary/unique/foreign keys intact. Use when the user is building or testing a data-quality or cleaning pipeline.
preset Use-case defaults: "demo" (dates end today, clean), "test" (<=200 rows, fixed seed), "load" (x10 rows), "ml" (declared nulls/outliers/typos/duplicates, keys protected), "eval" (current dates, modest dirt). Declarations always win.
domain Domain hint for post-generation validation.
"domain": "clinical_trial" # or "clinical", "financial", "fintech"
When set, generate_from_schema runs domain validation automatically and
returns it as domain_validation in the response — no second call
needed. Built-in ranges: HbA1c 4-14 %, BMI 10-80, systolic BP 60-260,
age 0-130, glucose 2-40, cholesterol 1-20, hemoglobin 3-25 for clinical;
price >= 0, discount 0-1, rate -1 to 100 for financial.
There are ~24 schema-level declarations in total. The ones above cover the common cases; for the rest (retention cohorts, DAG edges, closure tables, graph motifs, time grids, bitemporal history) fetch https://misata.studio/docs/reference/declarations.md and check the exact key before inventing one.
DESIGN RULES — follow these to get the best result in one pass
Always set rows per table. A fintech schema with customers=2000, accounts=4000, transactions=50000 is far better than 1000 everywhere.
Every child table needs a FK column pointing to its parent PK. Without it orphan rows are generated and the integrity proof will fail.
Use lognormal for money, file sizes, response times (right-skewed, strictly positive). Use normal for measurements (height, score, temp).
Declare correlations for any pair that co-varies in the real domain. Generated data with an identity correlation matrix is the clearest synthetic-data tell there is.
Use exact_incidence instead of probability when the user states a precise rate. "3% fraud" with probability: 0.03 gives ~3% on average; exact_incidence gives exactly 3%.
Use rollup (not formula) for any parent column that must reconcile with child rows. customers.total_spent generated independently of orders will never match; a rollup makes it exact.
Use outcome_curves any time the user mentions a revenue shape, a growth trajectory, a seasonal pattern, or a specific period total. It is the single feature most likely to be forgotten and most visible when absent.
Use profiles when two groups need different distributions. A clinical trial where all arms share one HbA1c distribution is statistically wrong and the difference will be caught by any summary table.
For longitudinal data (visits, sessions, sensor readings), add time_series to the key measurement columns. Independent rows fail every autocorrelation test and are visually obvious when plotted.
Add state_machine to any entity that moves through a process. An order table with no status progression is not realistic order data. When the user wants the steps themselves (an event log, durations, loops), declare processes instead.
Typos are errors. An unknown distribution name or a misspelled parameter (lamda, mena) is rejected with a "did you mean" suggestion. Read the error and fix the key rather than dropping the parameter.
After generating, call check_realism on the output_dir and fix what it flags: uniform money, every customer with the same number of orders, flat hour or weekday profiles, emails unrelated to names, no nulls.
Args: schema: Dict of table defs plus optional schema-level directives. rows: Default row count for tables without rows. seed: Random seed (same seed → byte-identical output on any machine). output_dir: Where to write CSVs. Omit to use a fresh temp dir. sample_rows: Rows per table to include in the JSON response (max 50).
Inputs: schema (required), rows, seed, output_dir, +1 more. Returns: result
inspect_schemaInspect the schema behind a storyARead-only
Return the full schema (tables, columns, relationships) for a story without generating data.
Heavier than preview_story — includes every column with its type and
distribution params. Use when the user wants to see the structure they'll
get, or to author a misata.yaml file from a natural-language seed.
Args: story: Plain-English description of the dataset. rows: Default row count for the primary table.
Inputs: story (required), rows. Returns: result
list_domainsList built-in domainsARead-onlyIdempotent
List the 18 built-in business domains Misata can generate from natural language.
Each domain has trigger keywords and a sample story you can pass to
preview_story or generate_dataset. Use this when the user asks
"what kinds of data can you generate?" or to suggest a story format.
Inputs: no named parameters. Returns: result
preview_storyPreview how a story is interpretedARead-only
Inspect what Misata would generate from a story — without generating any rows.
Returns the detected domain, confidence, near-misses, locale, scale, and a preview of the tables that would be produced. Use this to confirm interpretation before committing to a (potentially large) generation.
Args: story: Plain-English description of the dataset. rows: Default row count for the primary table (affects preview only).
Inputs: story (required), rows. Returns: result
query_sandboxQuery the sandbox databaseARead-onlyIdempotent
Execute a read-only SQL query against the sandbox database.
Allows AI coding agents to verify SQL queries, test aggregations, or inspect data in the sandbox directly without opening external terminals.
Args: query: SQL SELECT query to execute. db_path: Optional path to the database (defaults to '.misata/sandbox.db'). limit: Max rows to return (default: 100).
Returns: Dict with columns, rows (as list of dicts), row_count, and execution status.
Inputs: query (required), limit, db_path. Returns: result
seed_databaseSeed a real databaseADestructive
Fill a live Postgres or SQLite database with realistic, connected data, read from the database's own schema.
Reads the tables, columns, and foreign keys directly from the target database, generates data that respects them, inserts parents before children, then queries the database back to confirm every foreign key resolves. No schema file and no ORM are needed: a connection string is enough.
SAFETY — this is the only Misata tool that writes to a user's database:
It plans by default. With
apply=False(the default) nothing is written; you get the table list, insert order, existing row counts, and what would be inserted. Show that plan to the user.Only call again with
apply=Trueafter the user has seen the plan and agreed. Never passapply=Trueon a first call.If any target table already has rows, the write is refused unless the user chooses
truncate=True(wipe and reseed) orappend=True(keep existing rows, seed only empty tables, and draw foreign keys from the rows already there). Never guess between these.truncate=TrueDESTROYS existing data. Only use it on a throwaway development database and only when the user explicitly asks.
Args:
db_url: Connection string, e.g. postgresql://localhost/myapp_dev
or sqlite:///dev.db.
rows: Base row count; reference and transaction tables scale from it.
apply: False (default) plans only. True performs the write.
truncate: Wipe target tables (children first) before seeding.
append: Keep populated tables and seed only the empty ones.
tables: Optional allow-list of table names to seed.
skip_tables: Tables to leave untouched (migrations, auth, etc.).
seed: Random seed; the same seed reproduces the same data.
Returns:
A plan (applied: false) or a result with per-table row counts and
a per-relationship integrity proof (integrity.verified).
Inputs: db_url (required), rows, seed, apply, +4 more. Returns: result
solve_outcomeSolve outcome curves to hit a pipeline's targetsARead-onlyIdempotent
Find outcome-curve values that make SQL metrics over the data hit targets.
Use it when the number that matters is computed from the generated
tables rather than declared on them: net revenue after refunds, revenue
after a fee, a ratio of two tables. Each query runs in DuckDB over the
generated tables and returns two columns, the period and the value, one
row per period. Each target names a metric, the table.column outcome
curve that moves it (the lever, which must have exact per-period
values), and the value per period. Set direction: "inverse" when a
larger lever lowers the metric, as refunds lower net revenue.
Nothing is written. On success, replace each lever curve's
curve_points in your schema with levers[lever], then generate as
usual; attest_dataset will confirm the curves hold.
Args:
schema: The schema dict, with an outcome curve per lever.
queries: Metric name to DuckDB query, e.g. {"net": "SELECT date_trunc('month', ordered_at), SUM(amount) * 0.97 FROM orders GROUP BY 1"}.
targets: [{"metric", "lever", "values": {"2025-01": 70000}, "direction"?}].
tolerance: Relative tolerance per period, e.g. "0.5%".
max_iterations: Generate-and-query rounds allowed, at most 40.
seed: Seed used for every round.
Returns:
{"ok", "converged", "reason", "message", "iterations", "best_iteration", "levers", "steps"}. reason is one of
converged, max_iterations, max_seconds, diverging, unresponsive,
infeasible, pipeline_error; message says what to change when it
is not converged.
Inputs: schema (required), queries (required), targets (required), seed, +2 more. Returns: result
validate_domainValidate a dataset against domain rulesARead-onlyIdempotent
Flag values that are physiologically or financially impossible for a domain.
Where audit_dataset checks internal consistency, this checks values
against what the outside world allows. Built-in ranges include, for
clinical / clinical_trial: HbA1c 4-14%, BMI 10-80, systolic BP
60-260, age 0-130, glucose 2-40, cholesterol 1-20, hemoglobin 3-25; for
financial / fintech: price >= 0, discount 0-1, rate -1 to 100.
Use after generating a dataset in a regulated or measurement-heavy domain, or when the user asks "is this data plausible for a real clinic / bank?".
Args:
dataset_dir: Directory containing one CSV per table.
domain: One of clinical_trial, clinical, financial,
fintech.
Returns:
{"passed": bool, "errors": [...], "warnings": [...]}. passed
is True when there are no ERROR-level findings.
Inputs: domain (required), dataset_dir (required). Returns: result
validate_yamlValidate a misata.yamlARead-onlyIdempotent
Validate a misata.yaml document at two levels.
Runs both checks in sequence:
Structural — the published JSON Schema (correct field types, required fields, enum values). Catches typos and shape errors.
Semantic —
misata.validate_schema(probabilities sum to 1.0, every foreign_key has a matching Relationship, no cycles, outcome curves reference real columns, etc.). These are the rules that would crash generation; the error messages include suggested fixes.
Use this when an agent has authored or edited a misata.yaml on the
user's behalf and wants to confirm it parses and will actually
generate before invoking generate_dataset.
Args: yaml_text: The full contents of a misata.yaml file as a string.
Returns:
{"valid": true} if both checks pass; otherwise
{"valid": false, "errors": [...], "stage": "structural"|"semantic"}
with the layer that failed first.
Inputs: yaml_text (required). Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Most tools are clearly separated and the descriptions even cross-reference each other (audit_dataset vs check_realism vs validate_domain, preview_story vs inspect_schema, generate_dataset vs generate_from_schema). The remaining soft spots are preview_story/inspect_schema/explain_schema, which all 'show you what you'd get' for slightly different inputs, and create_sandbox vs seed_database, which both write to a database but differ in source of truth.
Every name is snake_case with a verb-first pattern (validate_yaml, audit_dataset, attest_dataset, check_realism, seed_database). Minor deviation: generate_from_schema is verb_preposition_noun while its sibling is generate_dataset, a near-pair that differs only by a preposition.
15 tools sits at the top of the well-scoped band and each one owns a distinct stage: authoring, previewing, validating, generating, attesting, DB seeding, sandbox CRUD, and outcome solving. Nothing looks like filler or a redundant wrapper.
The lifecycle is fully covered end to end: domain discovery and story preview, schema dry-run and YAML validation, two generation paths, outcome solving, three complementary quality checks, cryptographic attestation, plus live DB seeding and a queryable sandbox. The only gaps are trivial (no sandbox delete/cleanup, no patch-in-place of an existing dataset), both workaroundable outside the tool surface.
Maintenance
Related MCP Connectors
Generate realistic, FK-consistent synthetic test data for your databases from your AI assistant.
Generate realistic relational test data — 156 field types, 22 locales, JSON/CSV/SQL, free previews.
Interactive data visualization and executive PDF dashboards for ChatGPT, Claude, and Cursor.
Privacy-safe synthetic financial data for LatAm fintech, AI agents, testing and ML.
Related MCP Servers
- AlicenseBqualityDmaintenanceGenerates realistic mock data using Faker.js for database seeding, API testing, and development environments. Supports person/company data, custom patterns, multi-locale generation, and structured datasets with referential integrity.462 npm7MIT
- AlicenseNot gradedqualityDmaintenanceEnables creating interactive data visualizations from natural language queries using DuckDB for local databases or Databricks for enterprise data warehouses. Supports multiple chart types, CSV imports, SQL queries, and automatic statistical analysis through Claude Desktop.19MIT
- FlicenseNot gradedqualityDmaintenanceGenerates realistic, context-aware synthetic data for AI agents to populate databases, mock APIs, and create test scenarios without exposing real PII.7 npm3-
- AlicenseAqualityDmaintenanceGenerates realistic, referentially-coherent test data (SQL INSERTs, JSON, or CSV) from your database schema, resolving foreign keys and respecting constraints. Paste CREATE TABLE DDL or a JSON schema and get ready-to-run seed data with valid relationships.2MIT