Skip to content

v0.6 MCP-First Skills

Scope and Release Contract

v0.6 makes the official MCP Skills extension the maintained integration direction. A conforming client must discover, verify, and progressively read skills without SDK-specific tools. Existing MCP users retain their legacy tools and resources. All ten distributions remain in the lockstep release, including the deprecated native integrations. No package or API is removed in this release.

Status: Included in v0.6.0. Publication is recorded on the GitHub release page.

The sections below follow the roadmap order.

The release keeps all ten packages and existing APIs. Native framework integrations are maintenance-only, not removed. Native Skills requires MCP SDK 2.2+ in the 2.x line, while legacy mode retains MCP 1.x compatibility. The MCP [fs] and [http] extras install their providers. Native CLI requirements exclude older providers and MCP packages that lack the new APIs.

1. Native Adapter Deprecation and Migration

Framework-specific lifecycle hooks duplicated integration work and limited MCP interoperability. Native LangChain, native Agent Framework, and the Agent Framework MCP bridge now emit caller-attributed deprecation warnings without warning on import.

  • [x] Preserve existing APIs during a full minor-release migration window
  • [x] Provide model-free examples using framework-owned MCP clients
  • [x] Document retrieval, injection, pruning, caching, fast-path, and image gaps
  • [x] Keep format-import adapters maintained and published wheels available

Evidence: PR 161 and the migration guide. Removal is no earlier than v0.7 and requires documented alternatives for unresolved gaps.

2. Official Skills Extension and Protocol Baseline

Legacy SDK tools are not native Skills discovery. Native mode uses MCP SDK 2.2+ in the 2.x line, protocol 2026-07-28, and io.modelcontextprotocol/skills.

  • [x] Advertise Resources and the Skills extension using the supported SDK
  • [x] Implement paginated listing, direct lookup, result/cache fields, and errors
  • [x] Allow direct lookup independently of empty or partial listings
  • [x] Leave optional directory reads unadvertised

Evidence: PR 158 and PR 159.

3. Lossless Skill Resources and Manifests

Body-only APIs cannot preserve canonical files. Optional provider file access and bounded immutable snapshots retain raw bytes, full frontmatter, and complete manifests.

  • [x] Preserve instructions, references, binary files, and arbitrary directories
  • [x] Verify sizes and SHA-256 digests and reject ordinary drift during capture
  • [x] Enforce traversal, file-count, byte-size, and nested-publication boundaries
  • [x] Preserve JSON-compatible unknown frontmatter and canonical alias paths

Evidence: PR 155, PR 157, and PR 159. Hashes provide consistency, not publisher trust. Capture is not an atomic filesystem transaction or protection against a malicious provider.

4. Legacy MCP Compatibility and Host Boundaries

Existing clients need a migration path while native Skills support remains uneven.

  • [x] Preserve eight legacy tools and three static catalog/usage resources
  • [x] Test both MCP 1.29-to-2.2 and 2.2-to-1.29 client/server combinations
  • [x] Keep native discovery free of client-side file prefetch and tool injection
  • [x] Leave approval, activation, selection, context, and execution host-owned

Evidence: PR 160 and the tested client matrix. An older client can read canonical resources without supporting native discovery.

5. MCP Conformance and Client Compatibility Matrix

Transport connectivity alone cannot establish native Skills compatibility.

  • [x] Exercise official SDK clients over stdio and real loopback HTTP
  • [x] Cover pagination, exact bytes, binary data, traversal, nested skills, and errors
  • [x] Verify pinned mcpc 0.7.0 against native MCP 2.2 over stdio
  • [x] Run mcpc direct lookup before listing, two-page discovery, and verified reads in CI

The mcpc test uses isolated session state, bounded subprocesses, and cleanup. It independently checks instruction, reference, and binary bytes against source files and manifests. It does not test model activation or deliberate client-side integrity-failure handling. Server snapshot drift rejection is covered separately.

6. Skill Discovery and Delivery Benchmarks

Native manifests improve verification but can cost more than legacy catalogs. Measurements must report that trade-off rather than assume a universal speedup.

  • [x] Compare native and retained v0.5-style APIs for 1, 10, and 100 skills
  • [x] Report synthetic precision/recall and no-match results separately from delivery
  • [x] Count actual requests, source reads, client bytes, and estimated tokens
  • [x] Verify zero discovery prefetch and preserve measurements as a CI artifact

Evidence: PR 163 and benchmark methodology and observations. The baseline is the retained API in the same checkout, not an installed v0.5 artifact. No model calls, production accuracy, remote latency, or universal performance claims apply.

7. MCP Inspection and Deployment Diagnostics

Operators need actionable publication checks without confusing them with deployment certification. Native inspection and both CLI preflight paths validate publication.

  • [x] Report canonical URIs, manifests, sizes, digests, and protocol requirements
  • [x] Validate provider readiness and limits before opening a listener
  • [x] Close owned provider clients and omit configuration secrets from reports
  • [x] Document local and remote transport boundaries and fallback guidance

Evidence: PR 162.

Release Verification and Deferred Work

The final release commit must pass CI, coverage floors, strict documentation, CodeQL, locked and latest-permitted dependency tests, and version consistency checks. Build all ten distributions before tagging. Publication uses the existing tag-triggered Trusted Publishing workflow and PyPI environment approval.

Production TLS, authentication, reverse-proxy, and gateway deployment validation was explicitly deferred beyond v0.6 by the maintainer on 2026-10-06. It is neither passed nor certified. No production deployment was provisioned for this milestone. Keep the documented secure-deployment requirements when deploying remotely.

Retirement, provenance, authorization, and operational controls continue under v0.7. This release does not certify any host's consent or execution policy.