cursor-cli-mcp

Use every model in your Cursor subscription from any AI agent. No API keys needed.

An MCP (Model Context Protocol) server that bridges any MCP-compatible tool to Cursor's full model roster — GPT-5.x, Claude 4.x, Gemini 3.x, Grok, and more — using your existing Cursor subscription credits.

Why?

AI coding agents like Claude Code, Codex CLI, and others are powerful — but they're locked to a single model provider. Meanwhile, your Cursor subscription gives you access to 40+ models across OpenAI, Anthropic, Google, and others.

cursor-cli-mcp bridges this gap. It lets any MCP-compatible agent call any Cursor model as a tool, using your subscription. No API keys, no separate billing, no rate limits beyond your plan.

The value

| Without cursor-cli-mcp | With cursor-cli-mcp | |------------------------|---------------------| | Each model needs its own API key | Zero API keys — use your Cursor subscription | | Separate billing per provider | One subscription covers all models | | Rate limits per API | Use your existing Cursor allowance | | Locked to one model per agent | Switch models per-task from any agent |

What you get

| Tool | Description | |------|-------------| | cursor_prompt | Send any prompt to any Cursor model and get a response | | cursor_review | Code review using any model — checks bugs, security, performance, quality | | cursor_review_file | Review a file by path using any model | | cursor_list_models | List all models available in your subscription |

Prerequisites

1. Cursor installed with an active subscription (cursor.com)
2. Cursor CLI agent installed at ~/.local/bin/agent (Cursor installs this automatically — run cursor from terminal to verify)
3. Node.js >= 18

Custom agent path? Set the CURSOR_AGENT_PATH environment variable to override the default location.

Install

Option 1: npx (no install)

Use directly in your MCP config — no global install needed:

{
  "mcpServers": {
    "cursor-cli": {
      "command": "npx",
      "args": ["-y", "cursor-cli-mcp"]
    }
  }
}

Option 2: Global install

npm install -g cursor-cli-mcp

Then reference it in your MCP config:

{
  "mcpServers": {
    "cursor-cli": {
      "command": "cursor-cli-mcp"
    }
  }
}

Option 3: From source

git clone https://github.com/MazRadwan/cursor-cli-mcp.git
cd cursor-cli-mcp
npm install

{
  "mcpServers": {
    "cursor-cli": {
      "command": "node",
      "args": ["/path/to/cursor-cli-mcp/index.js"]
    }
  }
}

Setup by client

Claude Code

claude mcp add cursor-cli -- npx -y cursor-cli-mcp

Or add to your project .mcp.json:

{
  "mcpServers": {
    "cursor-cli": {
      "command": "npx",
      "args": ["-y", "cursor-cli-mcp"]
    }
  }
}

Codex CLI

Add to your MCP configuration:

{
  "mcpServers": {
    "cursor-cli": {
      "command": "npx",
      "args": ["-y", "cursor-cli-mcp"]
    }
  }
}

Any MCP-compatible client

cursor-cli-mcp uses stdio transport — it works with any client that supports the MCP standard. Just point your client's MCP server config to npx -y cursor-cli-mcp or the installed binary.

Use cases

Cross-model code review

Use Claude Code as your primary agent, but send reviews to GPT-5.4 or Gemini for a second opinion:

"Review this auth module for security issues"
→ cursor_review_file({ file_path: "/src/auth.ts", model: "gpt-5.4-high" })

Model comparison

Test the same prompt across different models to compare outputs:

→ cursor_prompt({ prompt: "Explain this regex: ...", model: "gpt-5.4-high" })
→ cursor_prompt({ prompt: "Explain this regex: ...", model: "sonnet-4.6-thinking" })

Overflow / fallback

When your primary agent's API hits rate limits or quota, route to Cursor as a fallback without changing your workflow.

Specialized tasks

Route tasks to the best model for the job:

• Deep reasoning → opus-4.6-thinking
• Fast iteration → gpt-5.4-high-fast
• Code generation → gpt-5.3-codex-high
• Planning → sonnet-4.6-thinking

Available models

Run cursor_list_models to see your full list. Typical subscriptions include:

• OpenAI: GPT-5.4, GPT-5.3 Codex, GPT-5.2, GPT-5.1
• Anthropic: Claude 4.6 Opus, Claude 4.6 Sonnet, Claude 4.5
• Google: Gemini 3.1 Pro, Gemini 3 Pro, Gemini 3 Flash
• Others: Grok, Kimi K2.5, and more

Model availability depends on your Cursor subscription tier.

Configuration

| Environment variable | Default | Description | |---------------------|---------|-------------| | CURSOR_AGENT_PATH | ~/.local/bin/agent | Path to the Cursor CLI agent binary |

How it works

flowchart LR
    A["MCP Client<br/>(Claude Code, Codex CLI, etc.)"] -- "stdio / MCP protocol" --> B["cursor-cli-mcp"]
    B -- "exec / agent binary" --> C["Cursor CLI<br/>(40+ models)"]
    C -- response --> B
    B -- response --> A

    style A fill:#2d333b,stroke:#539bf5,stroke-width:2px,color:#adbac7
    style B fill:#2d333b,stroke:#57ab5a,stroke-width:2px,color:#adbac7
    style C fill:#2d333b,stroke:#daaa3f,stroke-width:2px,color:#adbac7

1. Your MCP client sends a tool call over stdio
2. cursor-cli-mcp translates it into a Cursor CLI agent invocation
3. Cursor routes to the requested model using your subscription
4. Response flows back through the same path

Contributing

Contributions welcome! Please open an issue or PR.

License

See LICENSE for details.