mcp-data-platform-v1.90.0

Highlights

v1.90.0 makes the memory -> insight -> knowledge lifecycle legible and simple for the agents that use it. An agent, and a human reviewer, can now form a correct mental model from platform_info and tool descriptions alone: enumerate the review queue, cite any entity by a canonical reference, find feedback through search, and understand that a promotion's destination is chosen at apply time rather than frozen at capture.

This release contains no database migration and no breaking changes to existing tool callers. It is a safe upgrade from v1.89.x.

What's new

Enumerable review queue, with authors

• apply_knowledge action bulk_review gains an itemize: true mode that returns the pending insights themselves (paginated by offset/limit), each with its id, captured_by, sink_class, status, and entity_urns. This is how an agent enumerates the review queue; the relevance-ranked search tool could only return a ranked sample, never the full set. The default counts-only bulk_review stays cheap and bounded.
• Correctness fix: under the memory store's status mapping (pending, approved, and applied share one active status), a single status-filtered page could hide pending insights behind applied ones. The itemized path walks the full set, so the review queue is always complete.
• The insight author now also appears on unified search hits.

Canonical entity references everywhere

• Every search hit (assets, prompts, connections, knowledge pages, DataHub) and every list_connections entry now carries a reference string in its canonical mcp:<type>:<key> form (or a urn:li:... for DataHub). An agent can cite an entity from a knowledge page directly, without hand-assembling or guessing the form. References are validated before they are emitted, so an unresolvable reference is never returned.
• File-configured connections are seeded into the connections table at startup so that mcp:connection:(kind,name) references resolve for them. The seeding is insert-only and never overwrites a connection managed through the admin API.
• A reference that crosses the two namespaces (for example urn:li:mcp:connection:(...)) is now rejected with an error that names the correct form, instead of being stored as an unresolvable reference.

Feedback is discoverable through search

• Feedback threads are now a federated source in unified search, completing the search corpus. A caller's own feedback surfaces alongside the catalog, memory, insights, knowledge pages, prompts, endpoints, and connections. The source is per-user and lexical.

Persona-aware orientation

• platform_info now describes the review-and-promote lifecycle, and exposes the knowledge feature flags, only to a persona that can actually reach those tools. The orientation also states the boundary between DataHub references (urn:li:...) and internal references (mcp:...).

Simpler lifecycle: destination decided at apply

• A promotion's destination, a DataHub catalog entity or a knowledge page, is the sink chosen when the insight is applied, suggested by whether the insight is tied to a specific dataset or column. It is no longer frozen by the capture-time class. The capture-time class becomes a non-binding hint, and the validation that used to reject a mismatch is removed, so a reviewer is never blocked by an early classification choice.

Upgrade notes

• No database migration is required. Existing stored records and existing memory_capture callers continue to work unchanged; the capture-time class simply becomes a descriptive hint rather than a binding route.
• The startup connection seeding is insert-only and leaves admin-managed connections untouched.

Changelog

Others

• e235c23: Simplify and make the memory -> insight -> knowledge lifecycle legible (#686) (#687) (@cjimti)

Installation

Homebrew (macOS)

brew install txn2/tap/mcp-data-platform

Claude Code CLI

claude mcp add mcp-data-platform -- mcp-data-platform

Docker

docker pull ghcr.io/txn2/mcp-data-platform:v1.90.0

Verification

All release artifacts are signed with Cosign. Verify with:

cosign verify-blob --bundle mcp-data-platform_1.90.0_linux_amd64.tar.gz.sigstore.json \
  mcp-data-platform_1.90.0_linux_amd64.tar.gz