4 MCP servers worth connecting this week: data pipelines, video, deliveries, and the new official reference build
This week's roundup is deliberately practical. One entry is the official reference build every local agent setup should now be measured against, and the other three are hosted servers that remove a whole integration job: checking a data pipeline, making a marketing video, and tracking a delivery, all from inside the chat tool you already use. Every pick below was checked against its provider's own page on Oct 27, 2026, and anything we could not confirm that way was left out.
1. Official MCP reference servers, v2.0.0 (Model Context Protocol project)
What it does: The reference collection that defines how local MCP servers should behave: filesystem, memory, sequential thinking, and everything (TypeScript), plus fetch, git, and time (Python). Version 2.0.0 refactors all of them to support the modern 2026-07-28 spec while still answering older clients.
Agent use case: Give a coding agent safe local utilities: read and write files in directories you allow, keep working memory, think step by step, fetch a public URL, work with a git checkout, and handle time and time zones.
Hosted or local: Local (stdio).
Auth and permissions: No account sign-in. Permissions are local: the filesystem server only touches directories you pass to it, and the fetch server blocks private, loopback, and metadata addresses by default.
Compatible clients: Any MCP client that supports local stdio servers, including Claude Code, Cursor, and Codex.
Setup friction: Medium, because this is a breaking upgrade. The TypeScript servers now require Node 24, the Python servers require mcp 2.3 or later and below 3, and an unknown tool now returns JSON-RPC error -32602 instead of an isError result, so clients that relied on the old shape need updating.
Where to get it: https://github.com/modelcontextprotocol/servers/releases/tag/v2.0.0 (packages: @modelcontextprotocol/server-filesystem, server-memory, server-sequential-thinking, and server-everything at 2.0.0 on npm; mcp-server-fetch, mcp-server-git, and mcp-server-time at 2026.10.19 on PyPI).
Source and checked date: GitHub release page for modelcontextprotocol/servers, checked Oct 27, 2026.
2. Airbyte MCP (Airbyte)
What it does: Connects an agent to Airbyte Cloud so it can build, monitor, and troubleshoot data pipelines, and read data directly from sources and destinations through Airbyte's context layer.
Agent use case: Ask which syncs failed, check a connection's status, diagnose a broken sync, and create or change a connection, without opening the Airbyte dashboard.
Hosted or local: Hosted. Connect to https://mcp.airbyte.com/mcp; nothing to install.
Auth and permissions: OAuth 2.0 sign-in to Airbyte Cloud. The server acts with your Airbyte role, so it can only see and change the workspaces, connectors, and connections that role allows. When the token expires, your client asks you to sign in again. Automated agents that cannot open a browser can authenticate with an Airbyte application instead.
Permission caution: Direct reads use the credentials saved in that connector in that workspace, not your own credentials for the third-party system. Airbyte itself warns that this can let an Airbyte user see records they are not ordinarily permitted to see in that system, so keep write and destructive actions confirmed by a human.
Compatible clients: Claude, ChatGPT, Cursor, and VS Code Copilot, per Airbyte's docs.
Where to get it: https://docs.airbyte.com/platform/airbyte-mcp
Source and checked date: Airbyte Docs, Airbyte MCP, checked Oct 27, 2026.
3. Runway MCP (Runway)
What it does: Lets an agent generate polished images and videos, including product images, marketing videos, and hero assets, from a text prompt, a reference image, or a product URL, with results returned in the same conversation.
Agent use case: Send an agent a product page and get back a launch-ready marketing video, or drop in a product image and generate a range of new visual assets alongside website copy and code.
Hosted or local: Hosted. Do not install the older local API server; Runway now points users to the hosted MCP.
Auth and permissions: Sign in with your Runway account. No separate API key is required. Generations are tied to your existing Runway plan and billed against it.
Compatible clients: Claude, ChatGPT, Cursor, Replit, and any MCP-compatible agent, per Runway's announcement.
Where to get it: https://runway.com/news/company-news/mcp and connect via runway.com/mcp (server at mcp.runwayml.com). Models named by Runway include Gen-4.5, Seedance, GPT Image 2, Kling 3.0, and Nano Banana Pro.
Source and checked date: Runway, Introducing Runway MCP, checked Oct 27, 2026.
4. Dispatch MCP, beta (Dispatch)
What it does: A hosted server that lets teams get a delivery estimate, create an order, look up order details, and track deliveries in real time, by asking an AI assistant in plain language.
Agent use case: A dispatcher checks where an order is, or a sales rep gets a delivery quote before a customer call ends, without opening the Dispatch dashboard.
Hosted or local: Hosted. Nothing to install or host; Dispatch maintains the server.
Auth and permissions: Sign in with your Dispatch credentials. Connections run through Dispatch's identity and access controls, so the assistant only gets the tools and data your Dispatch account allows. This is a beta release, and Dispatch asks interested teams to reach out to their Dispatch account team to get connected; no public endpoint URL is published on the provider page, so do not guess one.
Compatible clients: Claude, ChatGPT, and other AI platforms that support MCP connections.
Where to get it: https://www.dispatchit.com/blog/dispatch-launches-mcp-server-connecting-ai-assistants-directly-to-deliveries
Source and checked date: Dispatch provider page, checked Oct 27, 2026.
Tested and working
Pass here means the canonical source opened, the server is listed as live or published by its provider, and the claims in the entry match the provider page. No server was installed or executed to fill this roundup, per the verification rule, so the install column below says so plainly.
- Official MCP reference servers v2.0.0: https://github.com/modelcontextprotocol/servers/releases/tag/v2.0.0, Not installed (by rule), None (local), PASS
- Airbyte MCP: https://docs.airbyte.com/platform/airbyte-mcp, Not installed (by rule), OAuth 2.0, PASS
- Runway MCP: https://runway.com/news/company-news/mcp, Not installed (by rule), Runway account sign-in, PASS
- Dispatch MCP (beta): https://www.dispatchit.com/blog/dispatch-launches-mcp-server-connecting-ai-assistants-directly-to-deliveries, Not installed (by rule), Dispatch account sign-in, PASS
- AIOZ AI MCP Server: Announcement only; no provider endpoint confirmed, No, Unconfirmed, WITHHELD
How we verify
Every server in this roundup was resolved to its canonical repository, package, provider page, or hosted endpoint, then re-checked on the checked date above. Cross-listed servers were deduplicated. Items with dead links, unclear provenance, or unverified claims were withheld, and nothing was installed or executed merely to fill the roundup. Each entry's source URL and checked date are on file in the Content ledger.
Past issues: Issue 1, "427 of 750 MCP servers actually work" (Oct 19, 2026).