Which AI assistants support MCP in October 2026? Only 2 of 6 document it

Most engineers assume that if a model is "agentic", their chat app can plug it into any tool. It cannot. Whether an assistant can talk to a Model Context Protocol server is a property of the app and the plan you pay for, not of the model. GPT-6 Astra has a first-class mcp tool in OpenAI's API, yet a ChatGPT Plus user may only get read access to a custom server. Gemini 4 Argon drove Blender over MCP in Google's launch demo, yet the consumer Gemini app's support for custom servers is not in Google's own release notes.

So which assistants can actually call a custom MCP server today, and can they do anything with the answer? Here is the state of play as of October 14, 2026, with a clear line between what the vendors document and what is only reported.

What MCP is, in one paragraph

MCP is an open protocol that lets an AI client discover and call tools on a server: search this, fetch that, create the other. A "remote" server is reached over HTTP (the current transport is Streamable HTTP); a "local" one runs on your machine. The server declares its tools with a schema, the client shows them to the model, and the model decides when to call them. For hardware work that means a model can ask a parts server for live stock and prices instead of guessing from training data. MCP in five minutes sets one up end to end.

The chat apps

App	Custom MCP servers?	Read or write?	How well documented
Claude (claude.ai, Claude Desktop, Cowork)	Yes. "Custom connectors" to remote MCP servers on Free (one connector), Pro, Max, Team and Enterprise. The server must be publicly reachable. Local servers via claude_desktop_config.json on Desktop only.	Tools run as the server defines them.	Official (support.claude.com)
ChatGPT	Yes, in beta "Developer mode". Remote servers only. Built-in apps and connectors on all tiers.	Full MCP including write actions on Business, Enterprise and Edu. Pro individuals (and Plus, per secondary sources) get read and fetch only on custom servers.	Official page located (help.openai.com), content partly verified
Gemini app	Partly. Google's October 13 release notes describe "Customize" with 40+ connected apps on all tiers and do not mention MCP. Third-party guides say custom apps from any remote MCP URL have existed since about June 2026 under Settings, Connected Apps (web only, AI Pro and Ultra). Gemini Enterprise has an official custom MCP connector, Streamable HTTP only.	Enterprise: documented. Consumer: unknown.	Enterprise official; consumer unverified
Grok app	Reported yes: connectors plus "Bring Your Own MCP" launched May 6, 2026 on web, iOS and Android for paid tiers, public URL required. Grok Build and the xAI API also take MCP servers.	Reported, not documented.	Third-party only (usecarly.com)
Muse app (Meta)	Not mentioned anywhere we found. Meta's launch page does not mention MCP; OpenRouter lists tool calling for the API.	Unknown	Unverified
Le Chat (Mistral)	Mistral's model page is silent on MCP.	Unknown	Not documented

Two of six, Claude and ChatGPT, have official documentation for end users connecting their own server. Gemini documents it for enterprise customers only. Grok is reported by third parties. Meta and Mistral say nothing.

The coding tools

If you work from a terminal or an editor, the picture is simpler. These tools all document custom MCP servers officially: Claude Code, Cursor, Codex CLI, Windsurf, VS Code with Copilot agent mode, Gemini CLI and Zed.

That list matters more than the chat table for hardware work, because a coding agent can see your repo. It can read the KiCad project, the CadQuery scripts and the platformio.ini before it asks a server for parts, which is the difference between "find me a 3.3 V regulator" and "find me a replacement for U3 on this board with the same footprint". Our guide How to use Claude Code for hardware design shows what that looks like in practice.

The model behind the tool is a separate choice. Grok 4.7, for instance, is available in Cursor, GitHub Copilot and Grok Build, all of which support MCP, while the consumer Grok app was still a model generation behind as of late September (x.ai). For a head-to-head of the models themselves on hardware tasks, see Which AI model is best for hardware engineering in 2026.

"Read-only" versus "write", and why it matters for parts

MCP tools can be annotated by the server as read-only (they look something up and change nothing) or not (they create, modify or delete something). Some clients enforce that distinction by plan. ChatGPT is the clearest case: on Business, Enterprise and Edu, Developer mode allows write actions; for individual Pro users, custom servers are limited to read and fetch.

For a parts server the split is intuitive. Scanning a repo for its BOM, searching live stock and pricing, and quoting a build are reads: nothing in the world changes. Creating a checkout link is a write: it creates an order waiting for payment. On a read-only plan the model can tell you what everything costs and where it is in stock, but you would place the order yourself on the website. On a plan with write access the model can prepare the order and hand you the payment link.

What no plan allows, and what you should not want, is the model paying. A well-designed server keeps payment out of the protocol entirely: the agent assembles the order, a human opens a hosted page and confirms. None of the clients above should hold a credit card.

How Next Everything plugs in

Next Everything is one remote MCP server, Streamable HTTP, at https://mcp.nexteverything.io/mcp, authenticated with a bearer token in the Authorization header. You create a free account, get a personal nx_live_… key, and paste one snippet into whichever tool you use. For Claude Code it is:

claude mcp add --transport http nexteverything https://mcp.nexteverything.io/mcp --header "Authorization: Bearer nx_live_your_key"

The docs have the equivalent snippets for Claude, Cursor, Codex, VS Code, Windsurf, Gemini CLI and Zed. Because the server is public and uses Streamable HTTP, it also fits the "custom connector" slot in claude.ai, Developer mode in ChatGPT and the enterprise Gemini connector.

The tools are the same whichever client calls them:

• nx_scan_project reads a repository or folder, extracts a normalized BOM and resolves each line to an orderable part, flagging ambiguous ones. It understands KiCad 7 to 9, atopile, tscircuit, Flux, Altium and EasyEDA BOM exports, Gerber plus pick-and-place, STEP, STL, 3MF, DXF, OpenSCAD, build123d and CadQuery output, and platformio.ini, Arduino and ESP-IDF or Zephyr manifests for picking dev boards.
• nx_search_parts returns live stock and pricing by part number, parametric spec or plain English, with alternates ranked by compatibility and filters for package, temperature range, RoHS and minimum stock.
• nx_quote prices a full build including shipping and duties; nx_quote_fabrication gives an instant quote for one manufactured part or a Gerber package across FDM, SLA, SLS, MJF, CNC milling and turning, laser, sheet metal, PCB and PCBA.
• nx_checkout creates a payment link. The agent cannot pay; you open the link and confirm. nx_track_order then follows every sub-order and tracking number.

Quotes are valid for 24 hours and prices are landed: parts, fabrication, shipping and import duties, delivered to the US, Europe and the UK. If a part is out of stock, you get a clearly marked matching replacement and nothing is swapped without your approval. There is no charge for using the service; you pay only for the parts in an order you approve, at the price shown before you pay.

If your assistant is one of the ones without custom MCP support, you are not locked out: upload a BOM, Gerbers or a STEP file on the site after signing up and get the same quote with no assistant at all.

Sign up free, paste the snippet into the client you already use, and ask it to order the parts for the thing you are building. You approve every order yourself.