Living Notes: The MCP Skills Over MCP Working Group, Meeting by Meeting

This is a living article. It is updated after every new meeting of the Skills Over MCP Working Group. Newest meeting first. Source of truth: the “Meeting Notes - Skills Over MCP WG” category in modelcontextprotocol/discussions. These are my summaries, not official notes, so check the linked originals before quoting anyone.

Last updated: 9 Nov 2026. Covers meetings through 3 Nov 2026 (discussion #3502).

Where things stand

SEP-2640 (Skills Extension) is Final (merged 13 Sep 2026). Under the extension lifecycle SEP (SEP-3392, Final 27 Oct, merged 29 Oct) it is now classified Beta, so the spec now lives in specification/beta/ of ext-skills.
The idea in one line: serve skill files as MCP resources, like a virtual file system, and let hosts verify the bytes with a SHA-256 digest and size. Skill format itself is delegated to the Agent Skills specification.
SDKs (as reported on 3 Nov): TypeScript ships it (v2.4.0, ext/skills); Python and C# are merged but unreleased; Go, Rust, Java, PHP and Ruby have open PRs; Kotlin has a tracking issue. FastMCP has an opt-in experimental SkillsProvider.
Open questions: how this relates to the Progressive Disclosure WG’s proposed groups primitive, what a skill “version” should mean, and how to change a finalized extension without breaking implementers.
Next spec release: late March 2027 (SEP-3398), with a soft freeze about 2.5 months before.

Timeline at a glance

Meeting log (newest first)

3 Nov 2026: Beta status, groups, versions, plugins

Original notes #3502

Beta, not Stable. The lifecycle SEP gives extensions four stages (Experimental, Beta, Stable, Core). Skills is Beta: breaking changes are allowed but must be recorded in the repo’s decision log, and a MAINTAINERS.md is now required.
Groups vs. skills. The Progressive Disclosure WG proposes a groups primitive where a group lists tools, prompts, resources and nested groups. A positioning doc weighs three options: competing, complementary, or groups replacing Skills over MCP. Most participants of that WG lean complementary; nothing is decided. Skills over MCP is deliberately simple, and the sequencing worry is that a more general mechanism arriving later would leave implementers unsure which to use.
Integrity. Skill files carry a SHA-256 digest and size. Groups has no equivalent today. One view in the room: a digest may not transfer to tools, because clients can truncate tool descriptions.
Tracking the Agent Skills spec. That spec moved to the Agentic AI Foundation on 22 Oct. The extension already defers format to it, and on a backwards-incompatible change clients must honor any compatibility mechanism it provides and should keep supporting the prior version.
SDK consistency. The conformance suite checks wire-level compliance only. To Nate’s knowledge there is no global SDK design document; API shape is left to each SDK’s maintainers.
What should a skill version mean? Semantic versioning was argued to fit poorly for skills that steer a non-deterministic system. Calendar versions or a content hash were discussed; a hash says “changed” but cannot say “newer”.
Plugins over MCP? The extension stays a skills extension. Hooks and subagents belong to the plugin world (the Agent Plugins specification); distributing plugins over MCP is seen as a distribution question for later.
Open actions: post implementations to the implementations list; review the groups positioning PR; raise integrity checking for groups; Swift and Java SDK help to the SDK WG.

22 Sep 2026: SEP is Final, docs site, conformance

Original notes #3388

SEP-2640 merged as Final on 13 Sep; the extension now has a published docs site.
Use case: whether an internal CLI tool directory should adopt the extension. Guidance: it depends on access-control needs and how skills are integrated.
Conformance: a proposal to add a runnable collection of real calls (Postman or Newman style) next to the automated suite.
Updating a Final extension without semantic versioning: the process is still open. Proposed: separate CODEOWNERS for the spec versus references and examples, and the core protocol’s release-candidate pattern (freeze, publish an RC, give implementers a few weeks).
Reserving the io.modelcontextprotocol/ reverse-DNS key in skill frontmatter metadata with the Agent Skills spec.
Composing skills with tasks, events and apps, with a normative versus non-normative guidance split borrowed from the fine-grained authorization group.

25 Aug 2026: Changes after core-maintainer feedback

Original notes #3307

Walkthrough of the SEP changes made in response to core-maintainer feedback, with a call for blockers before the vote. The SEP was put up for vote the same day.
Hosts must not prefetch a skill’s files on connection, listing or approval. That is normative text, though not actively policed. Supporting files are fetched when needed; on-demand loading latency and whether size limits are configurable were discussed.
Interim search pattern: a server exposes a tool such as search_skill that returns a skill URI. An official mechanism is expected from the progressive discovery effort, along with tagging.
Identity is the pair of server identity and URI, not the skill name. Clients such as Claude Code already namespace by server name plus URI.
Open: whether the WG becomes an interest group after V1; provenance and signatures continue in the Agent Skills repo with no spec-level mandate planned.

11 Aug 2026: Waiting on the vote

Original notes #3230

OpenAI had implemented the extension in its plugins as an install-a-snapshot approach, a signal that the current state becomes V1.
The SEP goes to core maintainers the next day. The rationale moves to the extension repo so the spec PR can render.
Advisory limits coming: file size and count limits plus advisory size metadata, so clients converge instead of each inventing caps.
Dynamic resources should be marked explicitly rather than by omitting a digest.
Skills-spec conversations move to the Agent Skills repo’s GitHub Discussions.

28 Jul 2026: Day of the spec release

Original notes #3187

Catching everyone up on the scoped-down revision awaiting core-maintainer feedback.
SKILL.md is the entry point; supporting files are sibling resources, each with its own digest and no aggregate digest. skills/get is the cheap single-skill refresh.
Archives stay out, recorded as a deferred feature with the objections that removed them.
The threat model recommends install (verify up front, store immutably) over load-on-demand.

30 Jun 2026: Responding to “not accepted as written”

Original notes #2994

On 24 Jun the core maintainers said SEP-2640 would not be accepted in its current form, citing archive support, local code execution, skill://index.json versus skills/list, and governance patterns.
Archives were reopened and the objections recorded: zip-bomb and path-traversal risk shifted onto clients, optionality creating a capability matrix, and a general preference for fewer things.
Decision: scope down to a minimal SEP (a skills file plus tool handling), defer archives, get core-maintainer alignment on serve-versus-distribute, and settle the small items (index.json versus a list, name conflicts).

16 Jun 2026: SDK strategy and very large catalogs

Original notes #2941

Where experimental SDK helpers live; reserving the io.modelcontextprotocol namespace in frontmatter metadata; skill-usage observability.
Discovery at scale: hierarchical indexes, a recommender-style lookup, and a discovery tool. One sizing data point: about 20,000 skills at roughly 500-character descriptions is an index of around 11 MB.
Go SDK reference implementation demo. No blockers raised to landing the SEP.

2 Jun 2026: Bloomberg’s platform and the resource-listing gap

Original notes #2859

Context from Bloomberg’s MCP platform; where compatibility and frontmatter should live; resources lacking filesystem-style directory listing, search and multi-resource reads.
Archive versus SKILL.md distribution, rate limiting and the cost of N reads, reinstating the digest field, whether resource templates stay, and one extension versus several.

19 May, 5 May and April sessions

#2766, #2701, #2628, #2585, #2539

Topics, from the notes’ headings: an experimental GitHub MCP server using skills as resources instead of server instructions; skills versus always-on instructions; discovery via the Agent Skills well-known path; inferred versus declared dependencies; script execution lifecycle; index format and metadata duplication; resource templates; a fallback plan if clients do not adopt the SEP; and the extension timeline.

Feb and March 2026: The interest group

#2248 (its title says 2025, apparently a typo), #2316, #2421, #2460

Topics, from the notes’ headings: problem space and scope, skills as instructors versus helpers, remote skills and portability, security, skill-format fragmentation, the documentation-as-MCP-server pattern, dynamic updates, and a proposed universal MCP skill loader.

Why this matters

Skills are becoming the unit of agent know-how, and MCP is where agents already reach tools. The group’s journey is a case study in how a protocol absorbs a new primitive: start with a broad idea, get pushback on scope, cut to the part everyone can agree on (serve files, verify bytes), ship, then argue about versioning and relation to adjacent proposals while real SDKs and clients adopt it.

For the hands-on side, see the Cui.Guru tutorial on writing your first agent skill and the companion deep dive, progressive disclosure in Agent Skills.

Update log: 9 Nov 2026, first version, 17 meeting-note discussions read (Feb to 3 Nov).