Cursor Bridge MCP

A local MCP server that gives Claude Code (and any MCP-compatible AI tool) native understanding of Cursor's context ecosystem — rules, agents, and skills.

Why

Most teams aren't all-in on one AI tool. Some developers prefer Cursor, others use Claude Code, and some switch between both depending on the task. But Cursor's context system — rules, agents, and skills — only works inside Cursor. Claude Code ignores .cursor/ entirely and reads only CLAUDE.md.

That leaves teams with a choice: maintain parallel context systems that drift apart, or pick one tool and lose the other's strengths. Cursor Bridge eliminates this by making Cursor's context files consumable by any MCP-compatible AI tool. Write your conventions once in .cursor/rules/, define your agents in .cursor/agents/, package your skills in .cursor/skills/ — and every developer on the team gets the same context regardless of which tool they use.

Related MCP server: Sub-Agents MCP

What It Does

Cursor Bridge reads .cursor/rules/, .cursor/agents/, and .cursor/skills/, applies Cursor's activation logic (glob matching, always-apply, description-based discovery), and serves the right context to any AI tool on demand.

• Rules are routed intelligently — always-apply rules inject automatically, glob-matched rules activate based on the files you're working on, and description-matched rules surface when the task is relevant

• Agents are served as loadable role instructions — when a task matches an agent's description, the AI can adopt that agent's behavioral guidance

• Skills are served as on-demand knowledge packages — the AI browses the index and loads what it needs, including supporting scripts and references

Quick Start

The fastest way to get running — no install needed:

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

Add this to .claude/settings.json (Claude Code) or .cursor/mcp.json (Cursor). That's it — npx downloads and runs the latest version automatically.

Other Install Options

Install globally

npm install -g cursor-bridge-mcp

Then use "command": "cursor-bridge-mcp" in your MCP config instead of npx.

Build from source

git clone https://github.com/theamazingwolf/cursor-bridge-mcp.git
cd cursor-bridge-mcp
npm install
npm run build
npm link

Then use "command": "cursor-bridge-mcp" in your MCP config, or point directly to the build:

{
  "mcpServers": {
    "cursor-bridge": {
      "command": "node",
      "args": ["/path/to/cursor-bridge-mcp/dist/index.cjs"]
    }
  }
}

Tools

get_active_rules

Returns rules that apply to your current files and/or task. This is the primary tool — call it when starting work or switching file context.

{
  "files": ["src/components/Button.tsx"],
  "task_description": "Refactoring the button component"
}

Rules are matched by activation mode:

• always — alwaysApply: true rules are always included

• auto — rules with globs patterns are included when files match

• agent-requested — rules with no globs are included when the task description overlaps with the rule's description

get_context_index

Returns a lightweight menu of all available rules, agents, and skills — descriptions only, no content. Use this to discover what's available.

{
  "types": ["rule", "agent", "skill"]
}

get_rule

Fetches the full content of a specific rule by name.

{ "name": "data-models-relationships" }

get_agent

Fetches the full definition of a Cursor agent — including its system prompt / instructions.

{ "name": "security-auditor" }

get_skill

Fetches the full content of a skill package.

{
  "name": "pdf-processing",
  "include_supporting_files": true
}

get_skill_file

Fetches a specific supporting file from within a skill directory.

{
  "skill_name": "pdf-processing",
  "file_path": "scripts/extract-text.py"
}

Usage with Claude Code

Add a CLAUDE.md to any project with a .cursor/ directory to tell Claude Code to use the bridge automatically. Here's a template:

## Cursor Bridge MCP

When the cursor-bridge MCP is connected, use it to read Cursor context files
instead of reading .cursor/ files directly.

- On session start, call `get_context_index` to discover available context
- When working on files, call `get_active_rules` with current file paths and/or
  a task description to get relevant rules with full content
- Use `get_rule`, `get_agent`, `get_skill` to load specific items by name
- Use `get_skill_file` to load supporting files from skill packages

The bridge applies Cursor's activation logic (glob matching, always-apply,
description relevance) so you get the right context automatically.

This way Claude Code knows to call the MCP tools rather than manually reading and interpreting .cursor/ files. The bridge handles the routing logic — which rules apply, which don't — so the AI doesn't have to.

Cursor Context Formats

Rules (.cursor/rules/)

Standard .mdc or RULE.md files with YAML frontmatter:

---
description: "React component conventions"
globs: "src/components/**/*.tsx"
alwaysApply: false
---

# Component Patterns

Your conventions here...

Both .mdc files directly in the rules directory and RULE.md files in subdirectories are supported.

Agents (.cursor/agents/)

Markdown files with agent metadata in frontmatter:

---
description: "Reviews code for security vulnerabilities"
model: inherit
readonly: true
is_background: true
---

You are a security specialist...

Skills (.cursor/skills/)

Directory-based packages with a SKILL.md and optional scripts/, references/, assets/ subdirectories:

.cursor/skills/pdf-processing/
├── SKILL.md
├── scripts/extract-text.py
└── references/form-fields.md

Configuration

Optionally create .cursor-bridge/config.yaml to customize directory paths:

rules_directory: ".cursor/rules"
agents_directory: ".cursor/agents"
skills_directory: ".cursor/skills"

All paths are relative to the workspace root. If no config file exists, the defaults above are used.

How It Works

• No file watchers — the index rebuilds from disk on each tool call, parsing only YAML frontmatter (not full file content). This keeps every call fresh and avoids platform-specific reliability issues.

• Lazy content loading — full markdown bodies are only read when a tool specifically needs them (e.g., get_rule, get_active_rules for matched rules). The index itself never loads content.

• No network access — pure local file parsing. No AI calls, no external APIs.

• No code execution — skill scripts are served as text content, never executed.

Development

npm install
npm run build
npm test