AI Guides › Workbench
By Nigel Guy · 7 min read
You copy a spreadsheet out of one tab, paste it into Claude, ask a question, copy the answer back, and repeat when the numbers change. Each round trip takes a minute, so it feels fine, but you are acting as a human clipboard: the model sees only the stale snapshot you remembered to give it. MCP exists to cut out that loop.
The rule: if you paste from the same system more than twice a week, connect it instead, and give the connection only the permissions the job needs.
The Model Context Protocol is an open standard for connecting AI applications to outside systems. The official site describes it as "a USB-C port for AI applications": one agreed plug shape, so a tool built once works in Claude, ChatGPT, VS Code, Cursor and other apps that support it.
Three parts matter:
| Part | What it is | Example |
|---|---|---|
| Host / client | The AI app you type into | Claude, Claude Code, ChatGPT |
| Server | A small program that speaks MCP on behalf of one system | A calendar, GitHub, a database |
| Tools | The actions a server offers the model | "search files", "create issue", "run query" |
A server is either remote (a web address, usually with a sign-in) or local (a program on your own computer). That decides which option below you can use.
What it does not replace: your judgement about what the data means, a proper integration for anything that must run unattended, or a backup.
| Option | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Claude directory connectors | One-click, vendor-listed remote servers in claude.ai | Available on all plans, Free included | Mainstream tools (calendars, docs, project trackers) | Only what is listed in the directory |
| Claude custom connector | Paste any remote MCP server URL into claude.ai | Free plan is limited to one custom connector; paid plans more | A vendor that publishes an MCP URL but is not in the directory | You vouch for the server yourself |
| Claude Desktop extensions | Installs local servers as .mcpb packages |
Included with the desktop app | Local files and apps on your own machine | Desktop only; runs code on your computer |
Claude Code claude mcp add |
Command-line set-up of remote or local servers | Needs a Claude plan or API access that includes Claude Code | Developers, repos, databases, team-shared config | Easy to over-connect and flood the context |
| ChatGPT developer mode | Custom MCP apps in ChatGPT | Full read-and-write MCP is Business, Enterprise and Edu only; Pro can connect read/fetch only | Organisations already standardised on ChatGPT | Admin-controlled; web only, not mobile |
Anthropic's pricing page lists Claude Pro in US dollars only ($20 a month, about £15 at time of writing); check the sterling price including VAT at UK checkout.
On Team and Enterprise plans, owners can set each action to Always allow, Needs approval or Blocked. Use it: reads on "Always allow", writes on "Needs approval".
Remote servers (the recommended transport):
claude mcp add --transport http <name> <url>
Local servers. The -- is required; it separates Claude's flags from the server's own command:
claude mcp add --transport stdio <name> -- <command> [args...]
Add --scope user to make a server available in every project, or --scope project to write it to a .mcp.json file your team can commit. The default, local, applies to the current project only. Check the result with claude mcp list, then type /mcp inside a session to see status and sign in to servers that need it. SSE transport still works but is deprecated; use HTTP where a vendor offers it.
Claude Code warns when one tool result passes 10,000 tokens and caps it at 25,000 by default (MAX_MCP_OUTPUT_TOKENS raises it). Narrow the query before raising the cap.
A workspace admin creates the app under Workspace settings > Apps > Create, enters the server endpoint, clicks Scan Tools, then publishes it from Drafts. ChatGPT connects to remote servers only.
1. Vet a server before you connect it. Paste the server's documentation or tool list where indicated.
You are a cautious IT reviewer helping a small UK business decide whether to connect an AI assistant to an external tool through MCP.
Goal: tell me whether this connection is safe enough for my intended use, and with which permissions.
Inputs:
- Server name and publisher: [SERVER_NAME_AND_PUBLISHER]
- Where I got the URL or package: [SOURCE_OF_URL]
- Its tool list or documentation: [PASTE_TOOL_LIST_OR_DOCS]
- What I want to use it for: [INTENDED_TASK]
Steps:
1. Sort every tool into read-only, write/modify, or delete/send.
2. Flag any tool that reaches beyond my intended task.
3. Note anything that fetches outside content (web pages, emails, shared documents), as that is where hidden instructions can arrive.
4. Recommend a permission setting for each tool: allow, ask first, or block.
Output: a table (tool | type | needed for my task? | recommended setting), then a one-line verdict: connect, connect with limits, or do not connect.
Rules: do not assume a publisher is official because of its name. If the source of the URL is unclear or the tool list is missing, stop and ask me for it instead of guessing. Before answering, check that every tool in my list appears in your table.
Fill in: who publishes the server, where the link came from, its tool list and your task.
2. Replace a paste loop with a connected request. Use once the connector is switched on in the chat.
You have access to my [CONNECTED_TOOL] through a connector. Act as a careful assistant who reads before acting.
Task: [WHAT_YOU_USED_TO_PASTE_IN_AND_ASK], using the current data in [LOCATION_EG_FOLDER_SHEET_OR_PROJECT].
Steps:
1. Tell me which tool calls you plan to make and what each will read.
2. Fetch only what this task needs; do not open unrelated items.
3. Answer the question, citing the item names or IDs you used.
4. If an action would change, send or delete anything, describe it and wait for my "yes" before doing it.
Output: a short answer first, then the sources you read as a bullet list.
If the location is ambiguous or the connector returns nothing, say so and ask me, rather than filling gaps from memory. Before replying, confirm every figure you give came from a tool result in this conversation.
Fill in: the connected tool, the question you used to answer by pasting, and exactly where the data lives.