Data Prism

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

Pseudonymise per scope

The pseudonym is a function of both the subject and the case id the caller's session carries. Two investigations, with two different case ids, cannot correlate their findings by comparing pseudonyms.

Scope isolation
Fail closed

The server and starter load only the bundled privacy profiles; both set unclassified: FAIL_REQUEST, refusing the whole response for a field nobody classified.

Configuration contract
Verifiable audit trail

Replaying a writer's chain catches an edit to any hashed field anywhere in it, including the last record, and catches a deletion when later records follow it. It cannot detect truncation of the most recent records, deletion of a whole boot's records, or recomputation by someone who already has write access.

What this does and does not prove

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.

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, embedded Hazelcast identity cache and read budget, 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=0.3.1; 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.

Where to go next

Local Compose quickstart — one command, no local JDK or Maven install, a real pseudonymised MCP response.
Protect your own API — a YAML-only walkthrough from a real JSON REST API to a working get_entity_context call.
Connect an agent client — local stdio for development, or authenticated Streamable HTTP against a real deployment.
FAQ and Comparison — direct answers, and how Data Prism relates to Presidio, LLM Guard, NeMo Guardrails and MCP gateways.