Skip to content

mcp-data-platform-v1.87.0

Choose a tag to compare

View all tags
@github-actions github-actions released this 22 Jun 07:16
v1.87.0
4ce8c00
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

mcp-data-platform v1.87.0

This release reshapes the agent-facing tool surface around a single discovery entry point and a single knowledge-write verb, adds end-to-end distributed tracing, and lands structural quality gates. The theme is topology-free discovery: an agent should find what exists across every source with one call, instead of having to know the platform's shape before it can ask.

Highlights

Universal search — one topology-free discovery entry point (#645, #632)

search replaces the older, source-specific search tools with a single call that federates every source a persona can access and returns a balanced, grouped-by-source result set:

  • Corpus: the technical catalog (DataHub), the caller's memory, captured insights, saved assets, prompts, API endpoints (aggregated across API-gateway connections), and connections. API endpoints and connections are in the default corpus, not behind an opt-in.
  • Balanced results, not a flat list. Rather than one relevance list that lets the strongest source dominate, results are built from a total budget with a per-source floor (so breadth stays visible), a per-source ceiling (so no source runs away), and redistribution of unused budget to the sources with more relevant hits.
  • Coverage summary. Every response reports per-source matched-vs-shown counts, so the agent learns where the answer space lives even when only the top few of each source are displayed. This is the anti-tunnel signal.
  • Navigational, callable-first. Hits are snippets (title, reference, source); the agent drills in with the scoped tool (trino_query, api_invoke_endpoint, datahub_get_entity). A sources filter narrows the federation; it never widens it past a persona's access.
  • Per-source scope, fail-closed. Each provider enforces its own access scope; memory, insights, and assets are scoped to the caller server-side.

Unified knowledge read and write paths (#632, #633)

  • Read: a single provider router behind search federates memory, insights, the catalog, prompts, and assets onto one ranked path, replacing the separate per-store search tools.
  • Write: memory_capture is the one verb for recording knowledge, routed by sink-class with recall-first deduplication. Reviewed sink-classes (business knowledge, schema/entity, operational rules) create insights for admin review and DataHub write-back.

OpenTelemetry distributed tracing (#428)

End-to-end tracing spans the MCP middleware chain and the toolkit adapters, so a tool call can be followed from request through auth, enrichment, and the downstream Trino/DataHub/S3 call.

Breaking changes / upgrade notes

The tool surface consolidated. The following tools are removed and their capabilities are now served by search (read) and memory_capture (write):

  • datahub_search -> relevance search now lives in search; structured catalog navigation stays in datahub_browse.
  • memory_recall / recall_insight -> recall (relevance, entity-keyed lookup, and DataHub lineage traversal) is now part of search.

If a deployment's server.agent_instructions or a persona's tool allow-list references any of those names, update them to search / memory_capture. search is registered whenever at least one knowledge source is available; the platform-owned, search-first instruction baseline that makes agents call it first is a follow-on and is not part of this release.

Engineering and quality

  • Structural maintainability gates (#594): a package-size budget test and related structural invariants run in CI, forcing decomposition of large packages rather than unbounded growth.
  • pkcestore extraction (#636): in-flight PKCE state moved out of pkg/admin into its own package, shrinking the admin package below the new budget.

Security

  • Dependency hardening (#643): patched vulnerable npm dependencies and pinned dev base images to digests.
  • dompurify bump (#640).

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:v1.87.0

Verification

All release artifacts are signed with Cosign. Verify with:

cosign verify-blob --bundle mcp-data-platform_1.87.0_linux_amd64.tar.gz.sigstore.json \
  mcp-data-platform_1.87.0_linux_amd64.tar.gz

Full changelog (v1.86.4...v1.87.0)

Features

  • feat(search): universal topology-free discovery tool (#645) (#647)
  • feat(knowledge): collapse search into one knowledge_search verb (#632) (#642)
  • feat(knowledge): unified knowledge read path (knowledge_search + provider router) (#632) (#641)
  • feat(memory): unified memory_capture write verb with recall-first dedup (#633) (#644)
  • feat(observability): OpenTelemetry distributed tracing across the middleware chain and toolkit adapters (#428) (#639)
  • feat(verify): structural maintainability gates (#594) (#637)

Bug Fixes

  • fix(security): patch vulnerable npm deps and pin dev base images (#643)

Others

  • build(deps): bump dompurify (#640)
  • refactor(pkcestore): extract PKCE store from pkg/admin; ratchet budget to 11,800 (#636) (#638)

Full diff: v1.86.4...v1.87.0

Assets 25
Loading