How do you set up MCP servers in Cursor for a whole team?

Short answer

Cursor reads MCP servers from two files: .cursor/mcp.json in the project, which you can commit so the team shares it, and ~/.cursor/mcp.json in your home folder for yourself. It supports stdio, SSE and Streamable HTTP, OAuth for remote servers, and ${env:NAME} for secrets. Team admins can distribute shared servers; Enterprise admins can configure an MCP allowlist.

Two files, one for the team and one for you

Cursor keeps MCP configuration in plain JSON files.

File Applies to Share it?
.cursor/mcp.json in the project Everyone who opens this project Yes, commit it to the repo
~/.cursor/mcp.json in your home folder All your projects No, it is yours

Each entry is a server. Cursor supports three transports:

Transport Runs Configured with
stdio Locally, started and managed by Cursor, one user A shell command
SSE Local or remote, multiple users A URL
Streamable HTTP Local or remote, multiple users A URL

For remote servers Cursor handles OAuth: either the server registers Cursor dynamically, or you put a client ID, secret and scopes in an auth object. Sign-in uses the authorization code flow with PKCE, and Cursor renews tokens itself.

The same files work outside the editor. Cursor’s documentation says “MCP in the CLI uses the same configuration as the editor”, with the same order of precedence. In the terminal, agent mcp list shows which servers are configured and where each comes from, and agent mcp list-tools <name> shows what a server offers. A server the team commits to the repo is therefore available to everyone who runs Cursor’s agent in a terminal, too.

Setting it up for a team, step by step

  1. Pick remote servers where you can. A remote server with OAuth means nobody has a key in a file, and everyone signs in as themselves. Many vendors run one: see the GitHub MCP server or the Notion MCP server.
  2. Write .cursor/mcp.json in the repo with the servers the project needs.
  3. Keep secrets out. Use ${env:NAME} in env, url or headers, and let each developer set the variable. For local stdio servers you can point to an envFile.
  4. Admins set the allowlist. On Enterprise, go to Team Settings > MCP Configuration and allow servers by command or URL, and optionally by tool. Allowlisting does not install a server; it limits what is allowed to run.
  5. Distribute shared servers. Under Dashboard > Plugins & MCPs admins can make servers available to the team; Cursor notes these are available to Cloud Agents.
  6. Decide on auto-run. By default Cursor asks before every MCP tool call. If you switch to a run mode where allowlisted tools run without asking, make sure write tools are not on that list by accident.

Official integrations are also in the Cursor Marketplace, where “Add to Cursor” installs and signs in. Community servers are listed at cursor.directory; vet those as you would any third-party code, see where to find trustworthy MCP servers.

The common misunderstanding

“Our .cursor/mcp.json is in the repo, so everyone’s AI works the same.” Everyone’s AI has the same tools. The MCP config says nothing about what the agent knows or has learned; that lives in rules, AGENTS.md and Cursor’s own features, and project rules in .cursor/rules are shared only because someone wrote them down. See Cursor memory was removed and how to share Cursor context with your team.

The other one: the file only works in Cursor. A teammate on Claude Code needs .mcp.json; one on Codex needs config.toml. Same servers, three files. See MCP servers in Claude Code for a team.

What this means for a team

A committed .cursor/mcp.json plus an admin allowlist is a clean setup: the same tools for everyone, nothing unapproved. Keep the list short, because each server adds tools the model has to choose from. And treat the config as part of the codebase: review changes to it like you would review a new dependency.

Where Bombos fits

Bombos connects to Cursor as one MCP server, and to Claude Code, Codex, Copilot and the other AI tools the same way, so every tool in the team reads and adds to one shared memory: team knowledge, rules the AI learned from corrections, and what already sits in repos, documents and mail, with a trail of what each answer rests on. It runs in our cloud or on your own server, and one developer can try it first with their own agents.

Related questions

Where is the Cursor MCP config file?

Per project in .cursor/mcp.json, and globally in ~/.cursor/mcp.json in your home folder.

Does Cursor ask before it runs an MCP tool?

Yes, by default Cursor asks for approval before using MCP tools. MCP follows the same run modes as terminal commands; in Auto-review mode, allowlisted MCP tools run right away.

Can I use environment variables in Cursor's mcp.json?

Yes. Cursor resolves ${env:NAME}, ${workspaceFolder}, ${userHome} and a few others in command, args, env, url and headers. Local stdio servers can also load an envFile.

Can a Cursor admin restrict which MCP servers people use?

Yes, on Enterprise. Under Team Settings > MCP Configuration an admin sets an MCP allowlist by command, URL and tool. An allowlist does not install anything; it only limits what may run.

Does Cursor support MCP resources and prompts, or only tools?

Cursor's documentation lists tools, prompts, resources, roots, elicitation and the Apps extension as supported.

Sources

  1. Model Context Protocol (Cursor docs)
  2. MCP in the Cursor CLI (Cursor docs)
The first half hour · free

What does your AI forget?

  1. 01You tell us which AI you use and where it keeps starting over.
  2. 02We show what a shared memory changes there, on your own examples.
  3. 03You keep the plan, with or without Bombos.
Piet Baudoin
Piet BaudoinYou talk to Piet, from the Bombos teamDutch team from Utrecht and Nijmegen · data in Europe
We reply the same working day.