Claude MCP: how to connect Claude to an MCP server

What does MCP let Claude do?

MCP lets Claude call tools that live outside the model. Instead of you pasting a spreadsheet into the chat, Claude asks a server for the data at the moment it needs it — and, where the server allows it, asks that server to do something. The protocol handles discovery (what tools exist), invocation (calling one with arguments), and authorization (proving Claude is allowed to).

Two consequences matter more than the mechanics. First, the context is live: a connected server answers with today's numbers rather than whatever you last uploaded. Second, the server, not Claude, defines the boundary. If a server only exposes read tools, no amount of prompting makes Claude write. That is why evaluating an MCP server is mostly a question about its permission model.

The three ways Claude connects to an MCP server

Most confusion about “Claude MCP setup” comes from mixing these up. They are different paths with different trade-offs, and the right one depends on whether the server runs on your machine or on the internet.

One detail worth knowing before you debug a failed connection: for custom connectors, Claude reaches the server from Anthropic's infrastructure rather than from your laptop. A server behind a VPN, a firewall, or a local tunnel will not be reachable that way even though it works fine in your browser. Anthropic publishes the IP ranges to allow. Request-header authentication — the option for servers that use a fixed API key or bearer token instead of OAuth — is a beta feature rolling out gradually, so it may not appear in every account yet.

How to connect Claude to your Zylx Business Brain

Zylx Studio runs one MCP server at https://zylx.studio/api/mcp. Every client — Claude, ChatGPT, or any other MCP-compatible agent — connects to that same endpoint with the same capabilities and the same approval queue behind it.

1. 1

   Sign in to Zylx Studio and open Servers. Choose a connection scope: workspace (one brand only — the recommended default) or account (every workspace you own, with the assistant able to switch between them).

2. 2

   Generate a token. It is shown once. Regenerating replaces it and disconnects anything still using the old value; tokens expire after 90 days.

3. 3

   In Claude, open Customize → Connectors and choose Add custom connector. Enter the server URL above. On Team and Enterprise plans an owner adds it under Organization settings first, and members then click Connect.

4. 4

   Authenticate. Zylx publishes OAuth discovery metadata, so Claude can run the sign-in-and-approve flow itself. If you prefer a fixed credential, supply the generated token as an Authorization request header — Claude sends the value exactly as typed, so it must read Bearer <token>, including the space.

5. 5

   Enable the connector in a conversation from the + menu, then test it: ask Claude to list my Zylx workspaces. If that returns your workspace, the Brain is reachable.

MCP access is included on Pro and above. On the free plan the connection still authenticates and answers identity questions — which account, which workspaces — and returns an upgrade notice for the rest, so a failed tool call tells you why rather than looking like a broken connector.

What business context Claude can read

Claude reads the Business Brain: the persistent business context Zylx assembles from the systems a company already runs. Depth depends on what is connected — an account with nothing connected has little for Claude to read. These are the supported sources, verified 2026-08-09:

• Shopify

  Shopify Admin API — catalog, orders, customers, content

• Stripe

  Stripe API, read-only — payments and revenue

• Google Search Console

  Search performance, index coverage, sitemaps

• Google Analytics 4

  GA4 Data API — traffic, engagement, conversions

• Google Ads

  Campaigns, keywords, search terms, conversions

• Microsoft Clarity

  Behaviour signals — heatmaps, rage clicks, recordings

• Klaviyo

  Email campaigns, flows, lists, attributed revenue

• Ahrefs

  Backlinks, keywords, competitor overlap

• Semrush

  Organic keywords, position changes, keyword gap

• GitHub

  Repositories, pull requests, releases, CI status

• Site crawl

  Zylx's own crawler — pages, metadata, schema, internal links

What is read-only, and what can actually act

“Reads by default” is easy to claim and rarely evidenced. Zylx's MCP server classifies every tool it advertises by what that tool does to the workspace, and enforces the classification: a write-class call made with a read-only token is refused, and the refusal names the missing scope. Here is the full census as the server itself reported it on 2026-08-09 — 923 tools, of which 77% cannot change anything.

The shape matters more than the totals. Most of the surface reads. The next largest group does not act either — it writes proposals into an approval queue where a human decides. The tools that touch a live system are a minority, they need a token that was explicitly granted write scope, and the approval tools refuse a machine identity outright, so a connected assistant cannot approve its own proposal.

What this looks like in a conversation

Limitations worth knowing before you connect

• An assistant can only reach the workspace its token resolves to. A caller-supplied workspace ID that does not match is refused, not silently honoured.
• Write-class tools refuse a read-only token and return the exact missing scope instead of failing quietly.
• Approval tools refuse a machine identity — an assistant cannot approve its own proposal.
• MCP access is a paid capability. On the free plan the server answers session tools (who am I, which workspaces) and returns an upgrade notice for everything else.
• Tokens expire after 90 days. Regenerating one disconnects any assistant still using the old value.
• The server advertises tools only. It does not expose MCP resources or prompts.
• Context quality is bounded by what is connected. With no data sources connected, the Brain has little to answer from.

Claude MCP FAQ