Data Prism

Data Prism

Fail-closed privacy layer that pseudonymises enterprise API data for LLM agents and MCP clients.

Data Prism is an open-source privacy layer for Java/Spring teams putting LLM agents or MCP clients in front of internal APIs holding customer data. It pseudonymises personal data per privacy scope, refuses anything unclassified, and can keep a hash-chained audit trail.

What it is. A privacy enforcement layer for enterprise AI. It gives LLM agents useful, correlated business context — one consistent synthetic identity per entity within a privacy scope — while controlling which personal data crosses the AI boundary, and in what form.

Who it's for. Java/Spring platform and backend teams putting LLM agents or MCP clients in front of internal APIs that hold customer data. If nothing you run exposes personal data to a model, you don't need this.

Status: the walking skeleton and every slice through S9a are built, with 18 Maven submodules (19 Maven projects in the reactor counting the root pom-packaged aggregator itself) and a passing test suite. The privacy engine, correlation and consistency findings, parallel mTLS connectors, an embedded Hazelcast identity cache and read budget (shared only across members that have joined one cluster; see multiple instances), an OAuth2 resource server with session-derived PrivacyContext, audit and metrics are all real and exercised end to end. The standalone server is the primary deployment surface; the Spring Boot starter is the embedded option. A one-command local Compose quickstart also exists: see "Try it" below. Two MCP tools ship today, get_entity_context and compare_entity_sources — the other two named in the design review, search_entity_data and describe_entity_model, are not yet built (docs/tools.md "Not yet built"). A durable, append-only, hash-chained audit sink and an offline AuditChainVerifier ship as of 0.3.0, opt-in via dataprism.audit.sink: hash-chained; the verifier catches an edit or deletion inside a writer's chain, but cannot detect truncation of a writer's most recent records or the deletion of a whole process boot's records, and the trail does not resist an operator, or anyone else, who already has write access to the file (docs/audit.md "What this does and does not prove"). Not built: the re-identification operator surface (deferred past V1 by decision, see docs/architecture.md#decisions-worth-knowing) and the Elasticsearch connector and its search tools. See docs/plan/PLAN.md for what is open.

The problem

An organisation wants an LLM to investigate live business data spread across several systems. Giving the model direct API access is not acceptable: those APIs carry personal and confidential data, each system represents the same entity differently, and raw identifiers let anything downstream correlate across sessions.

The obvious fix — redact everything sensitive — destroys the investigation. Once three systems' names for one person are all [REDACTED], the model cannot tell whether it is looking at one person or three.

What Data Prism does

It sits between the two and does two things that are easy to confuse:

It makes identity consistent. One subject gets one synthetic identity across every source, derived deterministically from (scope, subject, namespace, algorithm version, key) — never random, never stored in plaintext, and reproducible without the cache. The same person in three systems reads as one person to the model.

It leaves the data inconsistent, and says so. If those three systems disagree about a name, the answer carries a finding that says they disagree. The platform never makes enterprise data look cleaner than it is. That distinction is the point of the project:

Identity representation becomes consistent. Underlying data inconsistencies become more visible, not less.

Pseudonyms are scoped. The same person in two different investigations gets two different synthetic identities, so nothing correlates across cases by accident.

What it is not

Not an API gateway, not an ETL platform, not a master-data system, not an identity provider, and not an entity-resolution engine — correlation requires a key the sources already share, behind a documented SPI. It carries no business domain: no Customer, Taxpayer or Employee type exists outside the example application.

It is not anonymisation. Under GDPR Art. 4(5), pseudonymised data is still personal data. Sending Data Prism output to a third-party model is still processing, and still needs a lawful basis, a DPIA, and a transfer mechanism where the provider is outside the EU. The platform reduces exposure; it does not remove the obligation.

Try it

The fastest way to see a real MCP call answered by the real privacy engine — no local JDK, no Maven install, one command:

docker compose up

pulls the published ghcr.io/aindriub/data-prism-quickstart-<name> images (pin one with QUICKSTART_IMAGE_TAG=0.7.0; run docker compose -f compose.yaml -f compose.build.yaml up --build instead to build every image from source) and brings up the standalone server, a synthetic fixture API and a local HTTPS JWT issuer, proving an agent-compatible get_entity_context call returns a pseudonymised response. Walk through it in docs/quickstart.md; connect your own agent client to either that stack or a real deployment via docs/agents/.

Once you have seen the demo, protect your own API: docs/quickstart.md ends with a "What next" section pointing at docs/protect-your-own-api.md, a YAML-only walkthrough from a real JSON REST API to a working get_entity_context call.

If you found this on the MCP registry

The ghcr.io/aindriub/data-prism-server image listed there is published as a multi-architecture manifest list covering linux/amd64 and linux/arm64, each built and verified natively — docker run on Apple Silicon or any other arm64 host pulls the arm64 image directly, no emulation required.

It is not a one-command install, on either architecture. docker run alone yields a server that refuses to start: DataPrismContractValidator demands a reviewed DataSourceAdapter bean for every configured source, and DataPrismProperties.validate() demands a full deployment configuration (JWT issuer/audience/JWKS, caller-claim mappings, security policy, HMAC key reference, audit sink, metrics sink, Hazelcast topology). Neither ships in the image. Two things an operator must supply themselves before it serves anything:

• A reviewed DataSourceAdapter (and IdentityResolver) jar for each API you are protecting, mounted onto the image's loader path.
• A deployment configuration satisfying the dataprism.* vocabulary.

docs/configuration.md is the authoritative, complete contract for both. The "Try it" section above is a local Compose fixture for evaluation, not this image or that configuration.

Documentation

The full set of user docs is also published, rendered and searchable, at https://aindriub.github.io/data-prism/.

User docs

Each row links the site page and the repo file it is built from.

Internal / project working docs

Not published on the site.

docs/plan/PLAN.md is the working queue. GitHub Issues is the front door for anything coming from outside — file there, not in PLAN.md.

Stack

Java 21 bytecode, Spring Boot 4.1.x, Maven multi-module, Hazelcast, Model Context Protocol via the official MCP Java SDK. The library modules require Java 21 or newer and Spring Boot 4.1.x; the published images run Java 25. Boot 3 consumers stay on 0.4.x.

Artifacts publish under group io.github.aindriub as data-prism-<module>, with package root io.github.aindriub.dataprism.

Building and running

Building needs JDK 21 or newer (the build compiles with --release 21, so a newer local JDK is fine; CI builds on 21 and 25) and Maven >= 3.6.3 (pom.xml:343-345 enforces this) if you use your own mvn. The repository ships the Maven Wrapper, which downloads the pinned Maven 3.9.16 (checksum-verified) on first use, so you need no local Maven install.

./mvnw -B --no-transfer-progress verify     # Windows: mvnw.cmd -B --no-transfer-progress verify

This is the same command CI runs (.github/workflows/build.yml). It builds all 18 submodules plus the root aggregator, runs the full test suite, the ArchUnit boundary rules, and the enforcer rule that keeps the classpath on Jackson 3, with Jackson 2 banned apart from jackson-annotations.

data-prism-server is the primary executable distribution. Its /health liveness probe is public and carries no deployment detail; its configured MCP path (normally /mcp) requires a verified bearer JWT. It deliberately contains no source schema, fixture adapter, or key. Supply configuration described in docs/configuration.md, plus a reviewed adapter for each configured source. Two ways to get one: a Java adapter extension with an annotated response model (the general case — nested objects, any transport), or, when the source's response is one flat JSON object, the published data-prism-connectors-rest artefact — loaded via -Dloader.path, configured entirely in YAML, no Java required. See docs/extending.md for both paths and exactly where the configuration-driven one's coverage ends (it never descends into a nested object).

Adapter extensions are ordinary jars containing Spring Boot auto-configuration that declares the required DataSourceAdapter beans and an explicit reviewed IdentityResolver (use PassThroughIdentityResolver only when every source genuinely shares the same identifier). Register that configuration in META-INF/spring/org.springframework.boot.autoconfigure.AutoConfiguration.imports, then load reviewed extension jars without rebuilding the server:

LOADER_PATH=/opt/data-prism/extensions \
  java -jar data-prism-server/target/data-prism-server-0.7.0.jar \
  --spring.config.additional-location=file:/etc/data-prism/application.yaml

The process refuses startup if configuration, secrets, operational bindings, or the exact configured adapter set is missing. data-prism-integration-tests is the reactor's cross-module integration test suite, not a fixture-only demo, and is never packaged into a deployable artefact. Container packaging and Compose orchestration for a real, locally runnable instance of this exist too — see "Try it" above and docs/quickstart.md.

Contributing

See CONTRIBUTING.md. The short version: read docs/conventions.md before opening a pull request, and expect the privacy rules in it to be enforced literally.

Licence

Apache License 2.0 — see LICENSE and NOTICE.