Cursor MCP Not Working / 'No Tools Found' Fix

"No tools found" almost always means the server never started or started and immediately crashed. Cursor launches the command you gave it; if the path is wrong or a required API key is missing, the process dies before it can report any tools. The other common case is a transport mismatch: a remote server wants a URL, not a command to run.

• The command/path is wrong or the package isn't installed.
• Missing env vars / API keys, so the server crashes on start.
• stdio vs remote mismatch (a remote server needs a URL, not a command).
• The server started but exposes no tools. Too many tools may also be enabled across servers.

The list above misses the cause that costs the most time. The config you are editing is not always the config that loaded. A project .cursor/mcp.json and a global ~/.cursor/mcp.json can both define a server, and the project file wins when the names clash. Editing the global one while a project entry shadows it produces a fix that changes nothing, which is miserable to debug because everything you are looking at is correct.

On a managed install the answer may not be in your config either.

Enterprise admins can allowlist which MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. servers members may connect to, and set each tool inside a server to auto-run or to wait for manual approval. When a server you added never appears at all, ask whether it is on the approved list before you spend an hour on the JSON.

Read the crash report first, then prove the server runs outside Cursor. Cursor writes each server's startup output to a dedicated log, and most failures show up right there as a missing binary or a crash on a missing key. Once it runs by hand, match the transport in your config, reload and confirm the tools come up green.

1. 1Check the MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. Logs: open the Output panel (⌘+⇧+U) and pick 'MCP Logs' from the dropdown. Startup crashes and auth failures land there.
2. 2Run the exact start command in a terminal and confirm it launches.
3. 3Add the required env vars / API keys to the MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. config.
4. 4Match the transport: command for stdio, url for remote (SSE or streamable HTTP).
5. 5Reload Cursor and check the server shows enabled (green) tools.
6. 6Trim to the tools you need so the agent isn't overwhelmed.

Step one is first because most of these failures look identical from outside the log: no tools, and nothing saying why. In the log they stop looking identical, and a missing binary reads nothing like a crash on a missing key.

The terminal test feels like the decisive one, so it is where most people start. Start with the log instead. It is cheaper and it usually names the failure outright, so keep the terminal for the case the log does not settle. A server that starts in your shell and dies under Cursor points at the environment rather than the command, and that is the argument for putting keys in the MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. config instead of your shell profile.

The last step reads like housekeeping and is doing real work. A long tool list spread across several servers gives the agent more ways to pick wrong, and a short relevant one produces steadier behaviour. It is also the only step here that helps in the case where everything technically connects and the agent still reaches for the wrong tool.

Outside the one-click Cursor or VS Code experience, MCPModel Context Protocol. A standard that lets an AI agent pull in context from outside the repo, like Jira tickets or internal docs. Press Enter for the full definition. needs manual setup and each host has its own trap. The server config is the same shape, but where you put it and what you have to enable changes.