AI Guides › Workbench
By Nigel Guy · 7 min read
You connect an AI tool to your calendar, your repo or your database, ask Claude to use it, and get a confident answer. Whether that answer came from the tool or from the model's general knowledge is something most people never check. A server that has quietly dropped, or never signed in, fails in an odd way: Claude carries on without it, and the output looks plausible.
The rule: before you rely on a connected tool, look at its status, and read the status word, not just the fact that a list appeared.
This guide covers Claude Code, where the /mcp command lives. MCP (Model Context Protocol) is the open standard that lets AI applications connect to outside data sources and tools. /mcp is Claude Code's panel for seeing and managing those connections.
| Check | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
/mcp |
Opens an interactive list of your servers, with a status for each, and lets you authenticate, enable or disable | Built into Claude Code; no extra charge for the command. Check the current plan price in £ at checkout | The first look, and fixing sign-ins | Shows connection state, not whether a tool does what you want |
/mcp reconnect all |
Retries every server that failed or needs authentication | Same | Recovering after a network blip or a sleep/wake | Needs Claude Code v2.1.284 or later, interactive terminal |
claude mcp list |
Prints all configured servers with health status, from the shell | Same | Checking before you start a session | Tells you the state outside any conversation |
claude mcp get <name> |
Shows one server's configuration and status | Same | Finding out why one server misbehaves | Read it for warnings, not just the status |
/context |
Shows what is filling the context window, with suggestions for context-heavy tools | Same | Spotting a server whose tool list is costing you space | A usage view, not a health view |
Individual MCP servers may have their own prices or plan requirements. Those are set by the server's vendor, so check the vendor's page.
/mcp and press Enter. You get an interactive list.Without opening the panel, you can run claude mcp list in a normal shell. In non-interactive mode (-p), /mcp with no argument prints a text summary of server status instead of opening the list; the docs say this needs Claude Code v2.1.205 or later.
These are the states as described in the Claude Code docs today. The exact wording may change between versions, so treat the meaning as the durable part.
| What you see | What it means | What to do |
|---|---|---|
| Connected | The server is connected and ready | Nothing. This is the only state that means "go ahead" |
| Needs authentication | The server wants an OAuth sign-in | Select it in /mcp and complete the browser login. Or run claude mcp login <name> |
| Failed to connect | It could not connect; the detail includes an HTTP status or error code | Read the code (below), then retry from /mcp |
| Pending approval | A project server from .mcp.json you have not approved in this workspace |
Run claude interactively and approve it |
| Rejected | A .mcp.json server blocked by your disabledMcpjsonServers setting |
Change the setting if the block was a mistake |
| Disabled for this project | You switched it off in /mcp |
Re-enable it in /mcp |
| Cached, connects on first use | A remote server whose tool list is cached; Claude Code connects when it is first needed (v2.1.221 onward) | Not a fault, but not proof of a live connection either |
| Not configured | A remote server with an empty url |
Set the URL in the detail view or the configuration |
The failure codes carry the clue. According to the docs, 401 or 403 means the server needs authentication, so sign in via /mcp. A 5xx or a plain connection error is usually transient, so reconnect. A configuration problem shows up in claude mcp get <name>.
The "cached" row matters most in practice. A cached server lists its tools, so it looks alive, but the connection is deferred. If your task depends on that server, trigger one harmless call and check again.
claude mcp add, or change scope. Servers can be local (default), project (shared via .mcp.json) or user scope, and a server added in one scope will not appear in another project./mcp reconnect all, then look again./context too, as each connected server adds tool definitions./mcp tells you the connection is up. This prompt tells you whether Claude can actually reach the tool in the task at hand. Paste it into a session after the check passes.
You are helping me verify that a connected tool works before I rely on it.
Tool to test: [SERVER_OR_TOOL_NAME]
Harmless read-only check I want you to perform: [A_SAFE_READ_REQUEST, e.g. "list the first 3 items"]
Steps:
1. Tell me which tools from this server are available to you right now. If you cannot see any, say so plainly and stop.
2. Run the single read-only check above using that tool. Do not write, delete, send or change anything.
3. Report the raw result, then say in one line whether it came from the tool or from your own general knowledge.
Rules:
- If the tool call fails or is unavailable, show me the exact error. Do not substitute an answer from memory.
- If my check request is missing or could change data, ask me for a safer one instead of guessing.
Before you answer, check that your reply states which tool you used, or states clearly that none was used.
Fill in the server name and a safe read request.
.mcp.json. Read it before approving./mcp, then the detail view./mcp reconnect all.claude mcp list.claude mcp get <name>./context./doctor, which the docs describe as a setup checkup that also flags unused MCP servers against their context cost, and reports findings before changing anything./mcp; this keeps the configuration without connecting it./help or check the current docs if your screen differs.