Delv
IDEby Anysphere★ 4.3

Cursor

The AI-first fork of VS Code that made "agent mode" mainstream. Broad MCP support, a polished config UI, and deep IDE integration.

Homepage MCP docs
B
Safety & Trust

Delv Safety Grade: B

Score 72/100 · assessed 2026-04-19

Maintainer75
Permissions55
Supply chain80
Transparency60
Incidents90

Cursor is a well-funded commercial IDE from Anysphere, a venture-backed startup that's gained significant traction in the developer tools space. The MCP integration is polished and the config UI is genuinely helpful for debugging server connections. However, as a closed-source fork of VS Code, you're trusting Anysphere with full filesystem access, shell execution, and network capabilities inherent to any IDE. The company has raised substantial funding and appears professionally run, but lacks the transparency of open-source alternatives. No known security incidents, and the MCP implementation follows Anthropic's reference patterns closely. The main risk is the broad permission surface: any IDE can read your entire codebase, execute arbitrary commands, and access environment variables including secrets. Cursor's proprietary AI features add an additional data flow consideration beyond standard VS Code.

Green flags

  • Polished MCP config UI with helpful error messages and server management
  • Well-funded startup with professional team and active development
  • Follows Anthropic's MCP reference implementation patterns closely
  • Standard installer distribution via official website
  • No known security incidents or credential leaks

Red flags

  • Closed source - no public repo to audit MCP implementation or data handling
  • Full filesystem and shell access inherent to IDE functionality
  • Proprietary AI features with unclear data retention policies
  • Relatively young company (founded 2022) compared to Microsoft/VS Code

Permissions requested

Read filesWrite filesDelete filesShell executeOutbound networkRead envAccess secretsExternal LLM call
Assessed by Delv Editorial using public metadata. Grades are advisory and update as the ecosystem changes. They do not replace your own review of permissions and code before granting an agent access to sensitive systems.

MCP capabilities

  • Tools
  • Resources
  • Prompts
  • Sampling

Platforms

macOSWindowsLinux

Config location

~/.cursor/mcp.json · or via Settings -> MCP

Review

I switched to Cursor six months ago and haven't looked back. It's VS Code underneath, but Anysphere rebuilt the AI layer from scratch. The MCP integration is the most polished I've used outside Claude Desktop itself. Config happens in two ways: either edit `~/.cursor/mcp.json` directly (same schema as Claude) or use the Settings UI. The UI is a genuine advantage. You can toggle servers on and off, see which tools are exposed, and restart connections without leaving the editor. When a server fails to start, the error messages actually tell you what's wrong. I've debugged half a dozen broken Python paths this way. MCP support covers tools and resources but not prompts or sampling. That's fine. I rarely need prompts in an IDE context, and sampling would be redundant given Cursor's own agent mode. The agent can call MCP tools mid-task, which is where this setup shines. I keep a filesystem server running for cross-project searches and a Postgres server for schema queries. When I ask Cursor to "add a migration for the users table", it fetches the current schema via MCP, writes the migration, and I'm done. The catch: Cursor's agent mode is opinionated. It auto-applies edits unless you explicitly review them. If you're used to Claude Desktop's cautious, step-by-step flow, this feels reckless at first. I've learned to pause and read the diff before accepting. The agent also doesn't expose which MCP tools it's considering, so debugging a failed call means checking server logs manually. Performance is solid. I run four MCP servers simultaneously (filesystem, git, Postgres, Slack) and haven't noticed lag. Cursor caches resource lists aggressively, so repeated queries are instant. The only hiccup: if you update an MCP server's code, you must restart Cursor entirely. The UI restart button only reconnects, it doesn't reload the server binary. One workflow I rely on: I have a custom MCP server that wraps our internal API docs. When I'm implementing a new endpoint, I ask Cursor to "check the auth middleware spec". It fetches the resource, pastes it into context, and I'm coding against the right contract. This beats switching to a browser or digging through Notion. If you live in VS Code and want MCP without the ceremony, Cursor is the obvious choice. The UI polish alone justifies the switch.
Verdict

Best MCP client for developers who want IDE-native AI without leaving their editor. If you need prompt templates or prefer explicit tool approval, stick with Claude Desktop.

Good at

  • Settings UI makes MCP config and debugging trivial
  • Agent mode calls MCP tools mid-task without manual approval
  • Handles multiple servers simultaneously with no performance hit
  • VS Code compatibility means your extensions and keybinds carry over
  • Error messages for failed servers are actually useful

Watch out

  • No support for MCP prompts or sampling
  • Agent auto-applies edits, which can feel reckless
  • Server binary updates require full Cursor restart
  • No visibility into which MCP tools the agent is considering
  • Opinionated workflow won't suit everyone