How to check if an MCP server is running: a guide for founders using Claude Code, Codex or Cursor
Ask the client, not the process list. In Claude Code, run claude mcp list: it connects to every approved server and prints ✔ Connected or ✘ Failed to connect with the reason. In Codex, run codex mcp list or type /mcp. In Cursor, open Customize and the MCP Logs panel. To test your own server, use the MCP Inspector.
This page is for founders who build with a coding agent and have just added an MCP server that the agent does not seem to use. We ran the Claude Code check on our own build machine on November 15, 2026: it health-checked 17 servers, 16 answered ✔ Connected and 1 answered ✘ Failed to connect. The real output is below, with what each status means and what to do about it.
How do you check in each coding agent?
Each client has a command or panel that shows its servers and whether it reached them. Only Claude Code's list command connects to each server as it runs; the others show what is configured, and you read the connection state in the session or the logs.
| Client | Check from the terminal | Check inside a session | Where errors show |
|---|---|---|---|
| Claude Code | claude mcp list (health-checks every approved server); claude mcp get <name> for one | /mcp | The failure detail is appended to the status line in claude mcp list, and shown on the Issue: row of claude mcp get and /mcp |
| Codex | codex mcp list (configured servers) | /mcp in the Codex terminal UI shows active servers; in the IDE extension, gear menu, then MCP servers | The IDE server list shows which servers are enabled and which need OAuth sign-in |
| Cursor | No list command in Cursor's MCP docs | Customize in the sidebar lists each server with an on/off toggle | Output panel (Cmd+Shift+U on Mac, Ctrl+Shift+U on Windows), then "MCP Logs": server initialization, tool calls and errors. A failed tool call also shows an error in chat |
From each vendor's MCP docs, read November 15, 2026 (links under Sources). Codex and Cursor rows are from the docs; the Claude Code row was also run on our machine.
What does a healthy server look like? Our own output
A healthy server shows ✔ Connected. This is what claude mcp list printed on our build machine on November 15, 2026 (Claude Code 2.1.322), cut to three of the 17 lines (the long plugin command shortened with ...):
Checking MCP server health…
playwright: npx -y @playwright/mcp - ✔ Connected
plugin:whyusersleave:whyusersleave: sh -c ... WUL_API_KEY="[REDACTED]" exec npx -y whyusersleave@1.7.1 - ✔ Connected
firecrawl: npx -y firecrawl-mcp - ✘ Failed to connect — CONNECTION_CLOSED: Connection closed
Four things to read in it:
- The status is per server. Claude Code's docs say a failure status means it could not connect to that server, not that the list command failed. One broken server does not stop the others.
- The reason is on the same line.
CONNECTION_CLOSEDon a server started withnpxmeans the process started and then exited before it finished the handshake. Run the same command by hand (npx -y firecrawl-mcp) and read what it prints; a missing API key is a common cause. - Plugin servers get a
plugin:prefix. The third line is the MCP server inside a Claude Code plugin (the UserProof plugin under its old name, WhyUsersLeave). - Keys are hidden. Claude Code printed
[REDACTED]in place of the API key in that line, so the output is safe to paste into a bug report. Check before you paste anyway.
What do the other statuses mean?
Claude Code's docs list more statuses (also Rejected, Disabled for this project and Blocked by enterprise managed policy); these four are the ones you will most likely see in claude mcp list:
| Status | What it means | What to do |
|---|---|---|
| ✔ Connected | Claude Code started or reached the server and finished the handshake | Nothing. If the agent still does not use it, type /mcp and check its tools are listed |
| ✘ Failed to connect | It could not connect. The detail is the HTTP status or error code, plus any text the server returned | Run the server's command by hand, or open the URL's docs for that error |
| ! Needs authentication | A remote server wants you to sign in | claude mcp login <name>, or sign in from /mcp |
| ⏸ Pending approval | A server from a project's .mcp.json that you have not approved yet; Claude Code does not connect to it | Run claude in that folder and approve it |
A slow server can also fail only because it took too long to start. Claude Code reads a start-up timeout from the MCP_TIMEOUT environment variable; its docs give MCP_TIMEOUT=10000 claude for 10 seconds.
Why can't you just check the process?
Because most local MCP servers are not running until a client starts them. A stdio server, the kind added with a command like npx -y @playwright/mcp, is started by the coding agent for its session and talks over that process's input and output. Before you open a session, ps shows nothing, and that is normal.
A remote server is a URL that stays up, but a plain request to it proves little. The MCP specification says initialization "MUST be the first interaction between client and server", so a server can be up and still reject anything that skips the handshake. Once connected, either side can send a ping, a JSON-RPC request with no parameters ({"jsonrpc": "2.0", "id": "123", "method": "ping"}), and the other side must answer. That is what a client's health check does for you.
How do you test your own MCP server?
Use the MCP Inspector, the testing tool from the MCP project. One package, @modelcontextprotocol/inspector, gives three ways in:
npx @modelcontextprotocol/inspectoropens a full inspector in the browser: connect, list tools, call one and watch the messages.npx @modelcontextprotocol/inspector --cliis the scriptable version, for CI, shell pipelines and coding agents.npx @modelcontextprotocol/inspector --tuiis a terminal UI for when you have no browser.
If the server works in the Inspector but not in your coding agent, the problem is the agent's config (a path, an environment variable, an unapproved project server), not the server.
How do you check which tools a server gives the agent?
In Claude Code, type /mcp and open the server: it lists the tools the agent can call. Codex's /mcp shows active servers; its docs do not say it lists each server's tools. In Cursor, Customize shows each server, and the toggle hides a server you do not want; Cursor's docs say disabled servers will not load or appear in chat. To compare which session replay and analytics tools offer an MCP server at all, see which session replay tools have an MCP server, and for a ranked list across jobs, the best MCP servers for coding agents.
How do you check the UserProof MCP server?
In Claude Code, UserProof comes as a plugin, so check two things: claude plugin list shows userproof@userproof, and claude mcp list shows its server with a plugin: prefix and ✔ Connected. Then ask the agent to run the UserProof setup; it reads your recorded visits through that server and hands back one fix with the file and line. Step-by-step install: set up UserProof in Claude Code, or Codex or Cursor. UserProof is free for 3 sites and 5,000 recorded visits a month; Pro is $29 a month for 50,000 and Scale $99 for 250,000 (pricing).
Sources, read November 15, 2026
- Our run of
claude mcp listandclaude mcp --help, Claude Code 2.1.322, November 15, 2026 - Claude Code, Connect Claude Code to tools via MCP (statuses,
claude mcp get,MCP_TIMEOUT) - OpenAI, Codex: Model Context Protocol (
codex mcp list,/mcp, IDE server list) - Cursor, Model Context Protocol (Customize, toggles, MCP Logs)
- Model Context Protocol, Lifecycle and Ping (specification 2025-06-18)
- Model Context Protocol, MCP Inspector
- UserProof plans: getuserproof.com/pricing
FAQ
How do I check if an MCP server is running?
Ask your coding agent's client. In Claude Code, claude mcp list connects to every approved server and prints Connected or Failed to connect. In Codex, run codex mcp list or type /mcp. In Cursor, open Customize and the MCP Logs output panel. For a server you are building, use the MCP Inspector.
Why does my MCP server say "Failed to connect"?
The client started or called the server and did not finish the handshake. The reason is on the same line: CONNECTION_CLOSED usually means the server process exited, often from a missing API key or a bad path. Run the server's command by hand and read its error.
How do I check which MCP tools are available?
Type /mcp in Claude Code and open the server to see its tools. Codex's /mcp shows active servers. To list tools outside an agent, connect with the MCP Inspector, in the browser or with its --cli option.
Why doesn't ps show my MCP server?
A stdio MCP server is started by the coding agent when a session opens and stops when it ends, so before a session there is no process to find. Remote servers stay up, but you check them with an MCP client, because the protocol expects a handshake first.
Where are MCP logs in Cursor?
Open the Output panel with Cmd+Shift+U on Mac or Ctrl+Shift+U on Windows and pick "MCP Logs" from the dropdown. Cursor's docs say the logs show server initialization, tool calls and error messages.