Datamimic open-sources deterministic synthetic test data for MCP-connected coding agents

A synthetic-data engine built for coding agents

Datamimic, an MIT-licensed synthetic test data generator from rapiddweller, has landed on Hacker News's front page with a pitch aimed directly at AI-assisted development: instead of letting a coding agent fabricate its own fixtures, hand it a governed engine that produces deterministic, realistic datasets. According to the project's GitHub repository, the Community Edition is Python-native, works standalone, and is MCP-ready, with a commercial platform layered on top for enterprise governance.

What the open-source edition does

The Community Edition generates fully synthetic, deterministic datasets from declarative models, with no source data required, and can pseudonymize staging or QA exports in either a seeded, deterministic mode or a non-seeded mode the project describes as privacy-maximized. Pipelines target a wide set of stores and formats, including PostgreSQL, MySQL, Oracle, MS SQL, SQLite, MongoDB, CSV, JSON, XML, XLSX, DbUnit and fixed-width files. Behavior can be modeled with weighted state machines, composite multi-field references, control flow such as while loops and assertions, and a scriptable in-memory store for staged aggregation. For auditability, runs emit append-only execution logs and a per-output content hash so results can be re-executed and checked. Installation is a single pip install (datamimic-ce).

The CLI doubles as an agent contract

The agent integration is the most distinctive part. The repository treats the CLI as the baseline contract for AI agents: a capabilities command returns a machine-readable JSON index of available operations, reference commands let an agent load schema fragments progressively rather than guessing, and a scaffold command bundles compile, lint, bounded dry-run and acceptance checks into one transaction that either verifies or returns structured issues for repair. Agents are told to stop once verification succeeds and not to touch the generated XML. The repo also ships an AGENTS.md guide plus copy-paste prompts for driving the tool from an agent session.

An optional MCP adapter, installed with pip install "datamimic-ce[mcp]", exposes four canonical operations — reference, scaffold, check and bounded-run — as Model Context Protocol tools, so IDE-connected agents can call them natively. The project deliberately keeps domain generation in the Python/CLI layer rather than building a parallel MCP authoring path.

Where the open core ends

The Enterprise Platform carries the features regulated industries actually get audited on. Per the repository, that includes a PII scanner with probability-scored field detection, coordinated multi-system execution across Oracle, MongoDB and Kafka with referential integrity, and deterministic templates for industry message formats such as EDIFACT, SWIFT MT, HL7 v2.x and HL7 FHIR. It adds role-based access, audit trails, approval workflows and scheduling, a performance core with a Rust fastpath and an ML-based engine for complex distributions, and on-premise or air-gapped deployment via podman-compose or Helm. Rapiddweller claims deployments in regulated EU banking environments, with reference customers available under NDA. The Community Edition, by contrast, executes single-system pipelines only.

Why it matters

Test data is a quiet bottleneck in precisely the organizations adopting coding agents fastest. Agents that invent their own inputs write tests against data no production system will ever produce, while copying real customer records into dev environments is a compliance liability in finance and healthcare. A seeded, deterministic generator offers a middle path: reproducible CI runs, referential integrity that survives relational assertions, and provenance logs that satisfy auditors. The MCP adapter matters because it meets agents where they already work, treating test-data authoring as just another tool call. The open question is the split: the free edition covers single-system generation and pseudonymization well, but scanning, multi-system execution and the governance layer sit behind the paid platform, so regulated teams will need to work out for themselves where the Community Edition stops being enough.