← Blog

AI Extensions, October 28: diagnose the protocol, load the right skill

October 28, 2026 · AI Extensions

In brief

Vercel's MCP patch improves version-discovery errors without silent fallback. Tech Leads Club separates installed skills from on-demand retrieval, with a license boundary worth reading.

A failed handshake needs a readable explanation

Originally published on Jack Gross.

AI-generated illustration
AI-generated illustration

Analysis and proposed test, not a product test

A failed handshake needs a readable explanation

An MCP connection fails before any tool runs. Retrying with different settings can hide the cause. Today's digest pairs a small compatibility patch with an on-demand skill catalog. I checked their primary documentation, release records and license files on October 28. I did not install or execute either package; these are source-based reviews, not a quality ranking.

Vercel AI SDK MCP 2.0.86

The October 27 release of @ai-sdk/mcp 2.0.86 improves unsupported-version discovery errors when a server advertises a compatible legacy version. It includes requested and advertised versions and explains the explicit protocolVersionDiscovery: false option. The release preserves original error details and avoids automatic legacy fallback. This is diagnostic behavior, not a promise that every server becomes compatible.

Current AI SDK docs distinguish stateless MCP 2026-07-28 from legacy initialization. They say recognized modern protocol errors do not trigger automatic fallback. For an endpoint known to require legacy initialization, the explicit false setting starts that handshake instead. I would keep the error and endpoint documentation together before choosing it, rather than disabling discovery everywhere.

The documented client import is createMCPClient from @ai-sdk/mcp. HTTP is recommended for production; built-in stdio is for local development. The small example below shows only the compatibility option, not authentication or production hardening. The repository LICENSE is Apache-2.0. The main-branch commit checked today is dated October 28, 04:52:49 UTC; the specific patch release is October 27, 19:35:56 UTC.

Illustrative code only; not executed against the product.

import { createMCPClient } from '@ai-sdk/mcp';
// Only for an endpoint documented to require legacy initialization.
const client = await createMCPClient({
  transport: { type: 'http', url: 'https://example.invalid/mcp' },
  protocolVersionDiscovery: false,
});
await client.close();

Tech Leads Club: installed skills or selective retrieval

The project offers an interactive CLI, npx @tech-leads-club/agent-skills, for choosing skills, target agents, installation scope and copy versus symlink. Its separate @tech-leads-club/agent-skills-mcp server consults the same catalog during a session. The documented path is search_skills, then read_skill, then only needed reference files. prepare_skill_files stages files on disk; staging is different from executing them.

The license file explicitly splits MIT software from maintainer-authored SKILL.md content under CC-BY-4.0 unless individually stated otherwise. Third-party skills retain their own licenses. A repository-level badge is not enough to clear a selected skill. Attribute Tech Leads Club where its content requires it and inspect the chosen skill's own notices.

The October 26 skills-catalog-v0.17.15 release records preserving approvals and screen verdicts after compaction. The main-branch commit checked today is also dated October 26, 03:11:37 UTC. These are maintenance signals, not a security audit. I do not repeat the README's vulnerability percentages or absolute safety claims as verified facts.

A bounded adoption check

For the MCP client, record the requested and advertised protocol versions and test one known read-only call against a disposable endpoint. For the catalog, inspect one selected skill, its license and referenced scripts before enabling it. Compare the actual installed or staged files to the reviewed revision. Neither an integrity hash nor a search score proves that instructions are safe or suitable.

Both approaches widen the agent's input surface. A remote server can return untrusted content; a skill can introduce executable scripts. Keep credentials and private working files out of a trial, use narrow service permissions and retain a rollback path. Popularity is not quality; no star or download counts are used here.

Sources checked October 28, 2026

AI-generated editorial illustration
AI-generated illustration