Where is the Cursor MCP config file?

Global: ~/.cursor/mcp.json (macOS / Linux) or %USERPROFILE%\.cursor\mcp.json (Windows). Project-scoped: /.cursor/mcp.json, which overrides the global one for that project. The project-scoped file is useful for team-wide MCP configs you want to commit to the repo — everyone on the team gets the same servers when they open the project.

What is the Cursor MCP config format?

JSON with one top-level key mcpServers — each entry has a name, command (the executable to launch), args (array of CLI args), and env (object of environment variables). Example: { mcpServers: { github: { command: 'npx', args: ['-y', '@modelcontextprotocol/server-github'], env: { GITHUB_PERSONAL_ACCESS_TOKEN: 'ghp_...' } } } }. Restart Cursor (or hit the reload button in Settings → MCP) to pick up changes.

Which MCP servers should I install first?

Three for immediate value: GitHub MCP (for PRs, issues, and comments — Composer can read your team's review style and follow it), a scoped filesystem MCP (the agent can read project files outside the open editor without you pasting them in), and a browser MCP — either Chrome MCP for your daily-driver browser or Human Browser MCP for protected-site work. Add others as you find pain — Linear MCP, Notion MCP, Stripe MCP, your own internal-service MCPs.

Why don't my MCP tools show up in Cursor?

Three common causes: (1) the launch command in your config failed silently — open Settings → MCP, look at the server status; red means launch error. (2) the env vars are missing — MCP servers usually need an API key in env; the launch will succeed but the server won't advertise tools. (3) you need to restart Cursor — config changes don't always live-reload reliably, especially on Windows. The Settings → MCP page has a 'Reload' button per server that's faster than restarting the whole IDE.

June 15, 2026 · 9 min read · AI Agents · Cursor

Cursor MCP (2026): Setting Up Model Context Protocol Servers in Your IDE

Q: Does Cursor support MCP?

Yes, since Cursor v0.42 (October 2024). You add MCP servers via Settings → MCP, or by editing ~/.cursor/mcp.json directly. Once a server is wired, the tools it exposes show up to the chat and Composer agents as callable tools. The Cursor MCP support is feature-complete with the official spec — same servers that work in Claude Desktop / Claude Code work in Cursor with no code changes.

Q: Can I share an MCP config between Cursor and Claude Code?

Yes — both clients use the same MCP server format and ecosystem. Keep one canonical file (~/.config/mcp/shared.json), and reference it from both ~/.cursor/mcp.json (Cursor) and Claude Code's config. New servers added to the shared file propagate to both. Reduces per-client config drift, especially for teams that use both tools.

Cursor IDE at the centre with three MCP server tiles plugged into it labelled GitHub, Browser, and Filesystem — Cursor MCP setup illustration

If you searched cursor mcp, you're probably trying to do one of two things: figure out how to wire your first MCP server into Cursor (this post covers the end-to-end recipe), or debug why a server you already added isn't showing up as tools the agent can call. We cover both. The MCP support in Cursor is feature-complete — the same servers that work in Claude Desktop, Claude Code, or any other MCP client work here, no code changes.

Disclosure: we make Human Browser, which ships an MCP server for stealth browser work. We'll cover the OpenClaw + Cursor + HB MCP setup as one of the three recommended first installs.

What MCP is, in one paragraph
Where Cursor's MCP config lives
The config format
Three servers worth installing first
Debugging tools that don't show up
Pairing Cursor with Human Browser MCP
FAQ

What MCP is, in one paragraph

The Model Context Protocol is a standard for giving an LLM agent typed tools. An MCP server is a process that exposes a set of named tools (github.create_issue, filesystem.read_file, browser.navigate). An MCP client — Cursor in our case, but also Claude Code, Claude Desktop, ChatGPT Desktop, Continue.dev — calls those tools generically. The protocol handles handshake, capability discovery, calls, and streaming. The point is that you don't bake in support for every API; you discover tools at runtime.

Cursor and the MCP ecosystem stay separately maintained. Servers improve; your IDE doesn't have to ship to use them.

Where Cursor's MCP config lives

Two locations:

Global: ~/.cursor/mcp.json on macOS and Linux, %USERPROFILE%\.cursor\mcp.json on Windows. Applied to every project you open.
Project-scoped: <repo>/.cursor/mcp.json. Overrides the global file for that project. Useful when you want a team-shared MCP config committed to the repo so everyone gets the same servers when they open the project.

The Settings UI (Cursor → Settings → MCP) edits the global file behind the scenes — fine for quick changes, but the JSON file is the source of truth and works better for any non-trivial setup (env vars referencing other files, comments, version control).

The config format

One top-level key, mcpServers. Each entry has command, args, and env:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..."
      }
    },
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/you/code",
        "/Users/you/Documents/notes"
      ]
    },
    "human-browser": {
      "command": "npx",
      "args": ["-y", "@virixlabs/humanbrowser-mcp"],
      "env": {
        "HB_TOKEN": "hb_live_..."
      }
    }
  }
}

That's the entire format. The minor extras:

disabled: true on an entry skips it without removing the JSON. Useful for "turn off Stripe MCP for this debugging session."
autoApprove: ["read_file", "list"] (per-server) — array of tool names the agent can call without prompting you. Use for read-only operations. Leave writes off the list so you confirm them.

Restart Cursor (or hit Settings → MCP → Reload on the affected server) after editing. The Settings → MCP page also shows per-server health — green for "launched and advertising N tools," red for "launch failed," yellow for "launched but no tools yet."

Three MCP server tiles plugged into a central Cursor IDE icon — GitHub, Chrome, and Browser servers each with a gear

Three servers worth installing first

From the wider MCP ecosystem, three give immediate, durable value in a Cursor setup:

1. GitHub MCP

@modelcontextprotocol/server-github. The agent can list issues, read PR diffs, post review comments, create branches, merge PRs (with permissions). The first "summarise the failing CI on PR #1234" or "find the open issues mentioning 'timeout'" query justifies it.

2. Filesystem MCP (scoped)

@modelcontextprotocol/server-filesystem with the allowed roots restricted to the directories the agent should touch. The agent can grep, summarise, and edit files outside the open editor without you pasting them. Pair with a permission model that requires confirmation for writes — or use autoApprove on the read-only tools only.

3. A browser MCP

Two natural picks:

Chrome MCP (@modelcontextprotocol/server-puppeteer or the official Chrome MCP extension) — attaches to your real signed-in Chrome. Great for Gmail summarisation, Calendar triage, internal dashboards where you're already logged in.
Human Browser MCP (@virixlabs/humanbrowser-mcp) — covered below. Stealth Chromium on a residential IP for sites that fight bots.

You can install both — they expose different tool names (chrome.* vs human-browser.*) so the agent picks the right one based on context.

Debugging tools that don't show up

Single most common Cursor MCP question we get: "I added a server but the agent doesn't know about it." The diagnostic flowchart:

Settings → MCP — is the server green? Red means launch failed. Click the server to see the error. 90% of the time: missing env var, wrong path in args, or the package name is misspelled.
Did Cursor pick up the config? Edit the JSON, save, then hit Reload on Settings → MCP. If the new server doesn't appear, the JSON is malformed — paste it into jq . in your terminal to validate.
Is the server advertising tools? Some MCP servers launch successfully but advertise no tools until they have credentials. Check the launch log in Settings → MCP — "0 tools" usually means the server is up but auth failed.
Are you using the right tool name in your prompt? The agent picks tools by name + description. If two tools have similar names (list_issues in GitHub MCP and Linear MCP), the description matters. Mention the server name in your prompt: "use the github server to list issues".

Pairing Cursor with Human Browser MCP

The browser-MCP slot is where the choice matters. The Chrome MCP is fine for "use my real session." It fails the moment your target is Cloudflare-, DataDome-, or PerimeterX-protected — the same data-center-IP and CDP-fingerprint walls every stock Chromium hits.

Human Browser ships an MCP server (@virixlabs/humanbrowser-mcp) that exposes the same browser-shaped tool surface (navigate, click, type, screenshot, evaluate) but the browser on the other end is a stealth-patched Chromium running on a residential IP, with a CAPTCHA solver and human-paced input baked in. From Cursor's side, it's just another MCP entry:

{
  "mcpServers": {
    "human-browser": {
      "command": "npx",
      "args": ["-y", "@virixlabs/humanbrowser-mcp"],
      "env": {
        "HB_TOKEN": "hb_live_...",   // from /install on humanbrowser.cloud
        "HB_DEFAULT_COUNTRY": "us"    // optional: residential-IP country
      }
    }
  }
}

Add this entry next to your GitHub + filesystem servers, reload, and Composer can drive a stealth browser whenever your task needs one. The agent picks human-browser.* when the target is protected and a regular browser MCP otherwise.

Try the Human Browser MCP — $1 for about 4 typical tasks

Pay-as-you-go: $0.10/hr browser time, $4/GB residential proxy, $0.005/captcha, $0.02 per agent step with the AI included. No subscription, unused balance never expires. humanbrowser.cloud.

FAQ

Does Cursor support MCP?

Yes, since v0.42 (Oct 2024). Same ecosystem as Claude Desktop / Claude Code.

Where is the config file?

Global: ~/.cursor/mcp.json. Project-scoped: <repo>/.cursor/mcp.json.

What's the config format?

JSON with mcpServers top-level key. Each entry has command, args, env. Optional disabled and autoApprove.

Which servers should I install first?

GitHub, scoped filesystem, and a browser MCP (Chrome MCP for daily-driver Chrome, Human Browser MCP for protected sites).

Why don't my tools show up?

Check Settings → MCP for server status. Red = launch error (missing env, bad command). Yellow = launched but no tools (often auth-failed). Restart Cursor or hit Reload after config changes.

Can I share an MCP config between Cursor and Claude Code?

Yes — same format and ecosystem. Keep one canonical file and reference it from both clients to avoid drift.

Cursor vs Claude Code (2026) — Comparison
Cursor Browser (2026): Built-in vs MCP vs Remote-CDP — AI Agents
Browser MCP: Connect Claude / Cursor to a Real Browser — MCP
OpenClaw and MCP (2026) — AI Agents