Migrating MCP Servers from Cursor to Claude Code

This guide helps you migrate MCP (Model Context Protocol) server configuration from Cursor to the Claude Code VSCode extension running in Cursor.

Quick Start: Copy-Paste Prompt
Manual Migration Instructions
Useful CLI Commands
Troubleshooting
MCP Tool Permission Configuration

Quick Start: Copy-Paste Prompt

For fastest setup: Copy everything between the arrows below and paste it into a new Claude Code chat window.

⬇️ COPY FROM HERE ⬇️

I need you to migrate my MCP servers from Cursor's configuration to work with the Claude Code VSCode extension. Here's what I need you to do:

Context: I'm using the Claude Code VSCode extension inside Cursor. Cursor has MCP servers configured in ~/.cursor/mcp.json, but the Claude Code extension doesn't automatically read that file for stdio-based servers. I need to migrate them using the Claude CLI.

Step 1: Find the Claude CLI

Use bash to dynamically locate the Claude CLI binary in my Cursor extensions directory
Path pattern: ~/.cursor/extensions/anthropic.claude-code-*/resources/native-binary/claude
Store it in a variable for reuse

Step 2: Read my existing Cursor MCP configuration

Check if I have ~/.cursor/mcp.json (this is Cursor's MCP config file)
Read it and show me all the MCP servers I currently have configured
For each server, note:
- Server name and command
- Whether it's HTTP-based (can skip these - they already work) or stdio-based (needs migration)
- Environment variables that need to be preserved
- Any custom arguments or flags
- Whether it uses local paths (e.g., ~/my-server/build/index.js)

Step 3: Show me the migration plan

List all stdio-based servers that need to be migrated
HTTP servers (if any) don't need migration - tell me they'll continue working
For each stdio server, show me the exact CLI command you'll use to add it
Ask me to confirm before proceeding

Step 4: Handle credentials For servers that need credentials (check the Cursor config for env vars):

If the Cursor config has hardcoded tokens/keys, warn me about security risks
Ask me if I want to provide fresh credentials or reuse the existing ones
Never hardcode credentials - always use --env KEY=VALUE in the CLI command
Tell me what each credential is for

Step 5: Execute the migration For each stdio-based server from my Cursor config:

Use the Claude CLI to add it:

$CLAUDE_CLI mcp add --transport stdio <name> [--env KEY=VALUE ...] -- <command> [args...]

Preserve the exact command, arguments, and environment variables from Cursor's config
For locally installed servers, preserve the full path exactly as it appears in Cursor
Show me each command as you run it

Step 6: Verify the migration

Run $CLAUDE_CLI mcp list to check all servers connected successfully
Show me the results with connection status (✓ Connected or ✗ Failed)
If any servers failed to connect, help me troubleshoot them
Tell me to reload Cursor: Press Cmd+Shift+P (Mac) or Ctrl+Shift+P (Windows/Linux) → type "Developer: Reload Window" → Enter

Step 7: Post-migration setup

If Memory Bank was migrated, remind me to initialize it after reload with: mcp__memory-bank__initialize_memory_bank
Tell me the migration is complete and the servers are ready to use

Important guidelines:

Preserve the exact configuration from Cursor's mcp.json file
Don't change package managers (if Cursor uses yarn dlx, keep using yarn dlx)
For locally installed servers, confirm the path exists before adding
Never hardcode credentials - always use --env flags
Skip HTTP servers - they don't need migration

Now, start by finding the Claude CLI binary and reading my Cursor MCP configuration file.

⬆️ COPY TO HERE ⬆️

What Happens After Pasting

Claude finds the CLI - Automatically locates the Claude CLI binary in your Cursor extensions
Claude reads Cursor config - Reads ~/.cursor/mcp.json and shows you all configured MCP servers
Migration plan - Claude presents which servers need migration (stdio) vs which work already (HTTP)
Credentials handling - Claude asks for any missing credentials (tokens, API keys)
Migration execution - Claude runs CLI commands to add each stdio server
Verification - Claude checks that all servers connected successfully
You reload Cursor - Reload the window to activate the migrated servers
Done! - Your Cursor MCP servers now work with Claude Code extension

Manual Migration Instructions

If you prefer to migrate servers manually or need detailed reference documentation, follow these instructions.

Understanding the Migration

Why migration is needed:

Cursor stores MCP server configuration in ~/.cursor/mcp.json
The Claude Code VSCode extension doesn't automatically read this file for stdio-based servers
Stdio servers must be added via the Claude CLI to work with Claude Code extension
HTTP-based servers (like Figma) continue working without migration

What gets migrated:

Server commands and arguments
Environment variables
Credentials (you'll need to provide these during migration)
Locally installed server paths

Migration Steps

1. Locate the Claude CLI Binary

The Claude binary is located in your Cursor extensions directory. The exact path depends on the version installed.

Find it dynamically:

find ~/.cursor/extensions -name "claude" -path "*/native-binary/claude" 2>/dev/null | head -n1

Example path:

~/.cursor/extensions/anthropic.claude-code-2.0.20-darwin-arm64/resources/native-binary/claude

Note: The version number (e.g., 2.0.20) and architecture (e.g., darwin-arm64) will vary based on your system and installed version.

2. Check Your Existing Cursor MCP Configuration

Before migrating, check what servers you have configured in Cursor:

# View your Cursor MCP configuration
cat ~/.cursor/mcp.json

Identify which servers are:

stdio-based (have "command" field) - these need migration
HTTP-based (have "url" field) - these work automatically, skip them

3. Migrate MCP Servers Using the CLI

Use the following command pattern to migrate stdio-based MCP servers from Cursor:

/path/to/claude mcp add --transport stdio <server-name> [--env KEY=VALUE] -- <command> [args...]

For locally installed MCP servers: If you have MCP servers installed locally (e.g., cloned from GitHub), use the path to the executable:

/path/to/claude mcp add --transport stdio <server-name> -- node /path/to/local-mcp-server/build/index.js

Note: Preserve the exact path from your Cursor configuration.

4. Set Up Your Environment

For convenience, set the Claude CLI path as a variable:

# Find and set the Claude CLI path
CLAUDE_CLI=$(find ~/.cursor/extensions -name "claude" -path "*/native-binary/claude" 2>/dev/null | head -n1)

# Verify it exists
echo $CLAUDE_CLI

5. Migration Examples

Below are examples of migrating common MCP servers from Cursor to Claude Code.

Example: Migrate GitHub MCP Server

From Cursor config:

{
  "Github": {
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-github"],
    "env": {
      "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_xxx..."
    }
  }
}

Migration command:

$CLAUDE_CLI mcp add \
  --transport stdio github \
  --env GITHUB_PERSONAL_ACCESS_TOKEN=YOUR_GITHUB_TOKEN \
  -- npx -y @modelcontextprotocol/server-github

Note: Replace YOUR_GITHUB_TOKEN with your actual token (get it from your Cursor config or generate a new one).

Example: Migrate Memory Bank MCP Server

From Cursor config:

{
  "aakarsh-sasi-memory-bank": {
    "command": "yarn",
    "args": ["dlx", "@aakarsh-sasi/memory-bank-mcp"],
    "env": {
      "MEMORY_BANK_ROOT": "/path/to/workspace/.cursor"
    }
  }
}

Migration command:

$CLAUDE_CLI mcp add \
  --transport stdio memory-bank \
  --env MEMORY_BANK_ROOT="$(pwd)/.cursor" \
  -- yarn dlx @aakarsh-sasi/memory-bank-mcp

Note: Adjust MEMORY_BANK_ROOT to match your workspace path.

Example: Migrate Atlassian MCP Server

⚠️ CRITICAL: DO NOT USE SMITHERY OR DIRECT mcp-atlassian PACKAGE ⚠️

Known Issues:

❌ Smithery wrapper (@smithery/cli) - Causes hanging and zombie processes
❌ Direct mcp-atlassian npm package - Missing jsdom dependency
❌ Python uvx mcp-atlassian - Has Pydantic configuration bugs
✅ ONLY USE: @aashari/mcp-server-atlassian-jira - This is the working Node.js version

Working configuration for .mcp.json:

{
  "mcp-atlassian": {
    "command": "npx",
    "args": ["-y", "@aashari/mcp-server-atlassian-jira"],
    "env": {
      "ATLASSIAN_SITE_NAME": "your-company",
      "ATLASSIAN_USER_EMAIL": "[email protected]",
      "ATLASSIAN_API_TOKEN": "YOUR_API_TOKEN"
    }
  }
}

CLI Migration command (if needed):

$CLAUDE_CLI mcp add \
  --transport stdio mcp-atlassian \
  --env ATLASSIAN_SITE_NAME="your-company" \
  --env ATLASSIAN_USER_EMAIL="[email protected]" \
  --env ATLASSIAN_API_TOKEN="YOUR_API_TOKEN" \
  -- npx -y @aashari/mcp-server-atlassian-jira

Getting your API token:

Go to https://id.atlassian.com/manage-profile/security/api-tokens
Click "Create API token"
Name it (e.g., "Claude Code")
Copy the token immediately

Available tools:

mcp__mcp-atlassian__jira_get_issue - Get issue details (param: issueIdOrKey)
mcp__mcp-atlassian__jira_ls_projects - List Jira projects
mcp__mcp-atlassian__jira_ls_issues - Search issues with JQL
Additional tools for comments, boards, sprints

CRITICAL: Correct Parameter Names

The @aashari/mcp-server-atlassian-jira package uses camelCase parameter names, NOT snake_case:

✅ Correct:

mcp__mcp-atlassian__jira_get_issue({
  issueIdOrKey: "PROJ-123"
})

❌ Wrong (will fail with -32602 error):

mcp__mcp-atlassian__jira_get_issue({
  issue_key_or_id: "PROJ-123"  // Wrong parameter name
})
mcp__mcp-atlassian__jira_get_issue({
  owner: "...",     // These are GitHub MCP parameters
  repo: "...",      // NOT Atlassian parameters
  issue_number: 123
})

Common parameter names:

issueIdOrKey - For getting/updating issues (e.g., "PROJ-123")
projectKeyOrId - For project operations (e.g., "PROJ" or "10001")
commentBody - For adding comments
jql - For JQL queries

Troubleshooting:

If not working after reload, check both .mcp.json AND ~/.claude.json use aashari package
Kill zombie processes: pkill -9 -f "smithery.*atlassian"
Verify: ~/.cursor/extensions/anthropic.claude-code-*/resources/native-binary/claude mcp list
If you get MCP error -32602 "Invalid arguments", check parameter names are camelCase

6. Other Common Server Migrations

Here are migration commands for other popular MCP servers:

# Context7 - Library documentation
$CLAUDE_CLI mcp add \
  --transport stdio context7 \
  -- yarn dlx -q @upstash/context7-mcp@latest

# Toolbox - MCP server search and discovery
$CLAUDE_CLI mcp add \
  --transport stdio toolbox \
  -- yarn dlx @smithery/cli@latest run @smithery/toolbox \
  --key YOUR_SMITHERY_KEY \
  --profile YOUR_PROFILE

# Shrimp Task Manager - Advanced task planning
$CLAUDE_CLI mcp add \
  --transport stdio shrimp-task-manager \
  --env DATA_DIR="$HOME/mcp-shrimp-task-manager/data" \
  --env TEMPLATES_USE=en \
  --env ENABLE_GUI=false \
  -- yarn dlx mcp-shrimp-task-manager

# Clear Thought - Structured thinking tools
$CLAUDE_CLI mcp add \
  --transport stdio clear-thought \
  -- yarn dlx @waldzellai/clear-thought

# Example: Locally installed MCP server
$CLAUDE_CLI mcp add \
  --transport stdio my-local-server \
  -- node $HOME/my-mcp-servers/custom-server/build/index.js

Migration Notes:

Preserve the exact command and package manager from your Cursor config
Copy environment variables and credentials from Cursor's mcp.json
For locally installed servers, preserve the full path exactly as it appears in Cursor
Check your Cursor config for the correct DATA_DIR paths for servers like Shrimp Task Manager

7. Verify Migration Status

After migrating your servers, check that they all connected successfully:

$CLAUDE_CLI mcp list

Expected output showing all your migrated servers:

Checking MCP server health...

mcp-atlassian: ... - ✓ Connected
github: ... - ✓ Connected
memory-bank: ... - ✓ Connected
context7: ... - ✓ Connected
toolbox: ... - ✓ Connected
shrimp-task-manager: ... - ✓ Connected
clear-thought: ... - ✓ Connected

If any server shows ✗ Failed to connect, see the Troubleshooting section.

8. Reload Cursor

After successfully migrating your MCP servers:

Press Cmd+Shift+P (or Ctrl+Shift+P on Windows/Linux)
Type "Developer: Reload Window"
Press Enter

9. Post-Migration Setup

After reload, your migrated MCP servers are ready to use!

If you migrated Memory Bank: Initialize it with the same path you used in the migration:

Ask Claude to run: mcp__memory-bank__initialize_memory_bank with path: <your-workspace-path>/.cursor

Note: Use the same MEMORY_BANK_ROOT path you configured during migration.

Useful CLI Commands

List all configured servers

claude mcp list

Get details about a specific server

claude mcp get <server-name>

Remove a server

claude mcp remove <server-name>

Reset project-scoped server approvals

claude mcp reset-project-choices

Configuration File Location

MCP servers are saved to: ~/.claude.json

This is a user-level configuration file that persists across projects.

Troubleshooting

Server shows as "Failed to connect"

Check that the command/package exists and is accessible
Verify environment variables are set correctly
Try running the command manually to see error messages

No MCP tools available after reload

Verify servers show as "Connected" in claude mcp list
Reload Cursor window again
Check the Output panel (View → Output → Claude Code) for errors

Cursor's mcp.json file not being read by Claude Code

This is expected behavior and why migration is necessary. The Claude Code extension doesn't automatically read Cursor's ~/.cursor/mcp.json file for stdio servers. You must migrate them using the CLI as described in this guide.

How do I know which servers need migration?

stdio servers (have "command" field in Cursor config) - need migration
HTTP servers (have "url" field in Cursor config) - work automatically, no migration needed

Can I keep using Cursor's native MCP servers alongside Claude Code?

Yes! Cursor's native MCP integration and Claude Code's MCP servers operate independently. You can use both simultaneously.

MCP Servers Connected But Tools Return "No Data" or "Not Found"

Problem: MCP server shows as "✓ Connected" in claude mcp list but API calls return no data:

Jira MCP returns "Issue not found" for valid issues
Lists return 0 items when data should exist
Authentication appears to work but queries fail

Root Causes:

Expired API Tokens - Most common cause
Project-specific MCP configuration not loaded - Server registered globally but not in project config
Environment variables not properly passed - Credentials not reaching the MCP server process

Solution Steps:

Step 1: Verify API Token is Valid

Test the token directly with curl to verify authentication:

# For Atlassian Jira
curl -s -u "[email protected]:YOUR_API_TOKEN" \
  -H "Accept: application/json" \
  "https://yourcompany.atlassian.net/rest/api/3/myself"

# Should return your user info (name, email)
# If it returns "Client must be authenticated" → token is invalid/expired

If token is invalid:

Go to https://id.atlassian.com/manage-profile/security/api-tokens (for Atlassian)
Create a new API token
Update both .mcp.json AND ~/.claude.json with the new token
Reload Cursor window

Step 2: Verify Project-Specific Configuration

Check if the MCP server exists in your project's configuration:

cat ~/.claude.json | jq '.projects["/your/project/path"].mcpServers["mcp-atlassian"]'

If it returns null, the server is not registered for this project even though it exists globally.

Fix: Add the server configuration to the project-specific config in ~/.claude.json:

# Example: Add mcp-atlassian to project config
cat ~/.claude.json | jq '.projects["/your/project/path"].mcpServers["mcp-atlassian"] = {
  "type": "stdio",
  "command": "npx",
  "args": ["-y", "@aashari/mcp-server-atlassian-jira"],
  "env": {
    "ATLASSIAN_SITE_NAME": "yourcompany",
    "ATLASSIAN_USER_EMAIL": "[email protected]",
    "ATLASSIAN_API_TOKEN": "YOUR_NEW_TOKEN"
  }
}' > /tmp/claude_updated.json

# Backup and replace
mv ~/.claude.json ~/.claude.json.backup
mv /tmp/claude_updated.json ~/.claude.json

After editing ~/.claude.json, reload Cursor for changes to take effect.

Step 3: Verify .mcp.json is Being Loaded

If you copied .mcp.json from another project:

Check that .mcp.json exists in project root
Verify it's not in .gitignore (though credentials should be)
Check ~/.claude.json for "enabledMcpjsonServers" array for this project:

cat ~/.claude.json | jq '.projects["/your/project/path"].enabledMcpjsonServers'

If empty [], the .mcp.json servers aren't enabled. This is usually resolved by reloading after the servers show as connected.

Step 4: Common Configuration Issues

/* WRONG - Using Smithery wrapper (causes issues) */
{
  "mcp-atlassian": {
    "command": "npx",
    "args": ["-y", "@smithery/cli@latest", "run", "mcp-atlassian"]
  }
}

/* CORRECT - Direct package */
{
  "mcp-atlassian": {
    "command": "npx",
    "args": ["-y", "@aashari/mcp-server-atlassian-jira"],
    "env": {
      "ATLASSIAN_SITE_NAME": "yourcompany",
      "ATLASSIAN_USER_EMAIL": "[email protected]",
      "ATLASSIAN_API_TOKEN": "YOUR_TOKEN"
    }
  }
}

Step 5: Full Reset Procedure

If all else fails, perform a clean reset:

# 1. Remove the problematic MCP server
~/.cursor/extensions/anthropic.claude-code-*/resources/native-binary/claude mcp remove mcp-atlassian -s local

# 2. Verify it's gone
~/.cursor/extensions/anthropic.claude-code-*/resources/native-binary/claude mcp list

# 3. Manually edit ~/.claude.json to remove any project-specific entries
# (use jq or text editor to remove the mcpServers entry for this project)

# 4. Update .mcp.json with fresh token

# 5. Add via CLI with fresh credentials
CLAUDE_CLI=~/.cursor/extensions/anthropic.claude-code-*/resources/native-binary/claude
$CLAUDE_CLI mcp add --transport stdio mcp-atlassian \
  --env ATLASSIAN_SITE_NAME="yourcompany" \
  --env ATLASSIAN_USER_EMAIL="[email protected]" \
  --env ATLASSIAN_API_TOKEN="YOUR_NEW_TOKEN" \
  -- npx -y @aashari/mcp-server-atlassian-jira

# 6. Manually add to ~/.claude.json project config (see Step 2)

# 7. Reload Cursor window

# 8. Test with a simple query
# Ask Claude: "Use mcp__mcp-atlassian__jira_ls_projects to list Jira projects"

Debugging Checklist:

API token tested directly with curl - authenticates successfully
claude mcp list shows server as "✓ Connected"
~/.claude.json contains server in project-specific mcpServers object
.mcp.json has correct credentials (if using project .mcp.json file)
Cursor window reloaded after configuration changes
Correct package used (e.g., @aashari/mcp-server-atlassian-jira, not Smithery wrapper)
Environment variables use correct keys (e.g., ATLASSIAN_SITE_NAME, not ATLASSIAN_BASE_URL)

Figma MCP: "No session found for sessionId" Error

Problem: When trying to access Figma designs via MCP, you get:

{"jsonrpc":"2.0","error":{"code":-32002,"message":"No session found for sessionId"},"id":null}

Root Cause: The Figma desktop MCP server session state is out of sync between Figma and the MCP client.

Solution (CRITICAL - Use HTTP transport + Toggle):

Ensure Figma desktop app is running with MCP server enabled:
- Open Figma desktop app
- Open the design file you want to access
- Switch to Dev Mode (Shift+D)
- In the inspect panel (right side), find the "MCP server" section
- Ensure "Enable desktop MCP server" is toggled ON
- Verify server is running: lsof -i :3845

Use HTTP transport (NOT SSE):

SSE transport consistently fails to connect with Figma MCP. The working configuration uses HTTP transport:

/* Remove SSE configuration if present */
$CLAUDE_CLI mcp remove figma-desktop -s local

/* Add with HTTP transport (this works) */
$CLAUDE_CLI mcp add --transport http figma http://127.0.0.1:3845/mcp

Update .claude/settings.local.json:

Ensure the enabled server list references the HTTP server name:
```
{
  "enabledMcpjsonServers": [
    "figma"
  ]
}
```
Reload Cursor after configuration changes
Toggle the MCP Server in Figma (CRITICAL FIX):

When you get the "No session found" error even with HTTP transport:
- In Figma Dev Mode inspect panel → MCP server section
- Toggle OFF "Enable desktop MCP server"
- Wait 2-3 seconds
- Toggle ON "Enable desktop MCP server"
- You should see a confirmation message at the bottom
- Try the MCP tool again immediately
Verify connection:
```
$CLAUDE_CLI mcp list
```
Should show: figma: http://127.0.0.1:3845/mcp (HTTP) - ✓ Connected

Why HTTP works but SSE doesn't:

Testing shows that while Figma's MCP server technically responds with SSE headers, the HTTP transport with periodic session reset (toggle) is more stable than SSE transport in Claude Code. The SSE configuration consistently shows "✗ Failed to connect" while HTTP shows "✓ Connected".

Alternative Session Reset:

You can also reset the session via Figma Preferences:

Go to Figma > Preferences dropdown
Select "Desktop MCP server settings"
Toggle the server off and then back on

Using Figma MCP Tools:

Extract node ID from Figma URL:

https://www.figma.com/design/FILE-KEY/Design-Name?node-id=1234-5678
                                                            ^^^^^^^^^
Node ID: 1234-5678 (preserve hyphen)

Get design context (code):

mcp__figma__get_design_context({
  nodeId: "1234-5678",
  clientLanguages: "typescript,javascript",
  clientFrameworks: "react"
})

Get screenshot:

mcp__figma__get_screenshot({
  nodeId: "1234-5678",
  clientLanguages: "typescript,javascript",
  clientFrameworks: "react"
})

Configuration file location: ~/.claude.json

Correct SSE configuration:

{
  "projects": {
    "/path/to/your-project": {
      "mcpServers": {
        "figma-desktop": {
          "type": "sse",
          "url": "http://127.0.0.1:3845/mcp"
        }
      }
    }
  }
}

Common mistakes:

❌ Using --transport http instead of --transport sse
❌ Not reloading Claude Code session after configuration change
❌ Figma desktop app not running or MCP server not enabled
❌ Using ReadMcpResourceTool instead of specific mcp__figma__* tools

MCP Tool Permission Configuration

The Permission Prompt Problem

After migrating your MCP servers, you will be prompted for permission every time Claude Code tries to use an MCP tool (e.g., mcp__clear-thought__sequentialthinking, mcp__memory-bank__get_memory_bank_status, mcp__mcp-atlassian__jira_get_issue). This creates significant friction and interrupts your workflow.

Why this happens:

Claude Code has a security model where MCP tools default to requiring approval
This is intentional to protect against untrusted MCP servers
However, for frequently-used, trusted MCP tools, this becomes "prompt fatigue"

⚠️ CRITICAL: Wildcard Permissions Don't Work

Known Bug: MCP tool wildcard permissions (mcp__<server>__*) are NOT honored in the VSCode/Cursor extension as of October 2025.

GitHub Issue: anthropics/claude-code#3107 - MCP wildcard permissions not being honored
Patterns like mcp__clear-thought__* or mcp__memory-bank__* will NOT prevent permission prompts
You MUST add each specific tool name individually
Status: This is a confirmed bug with no ETA for a fix as of October 2025

Solution: Add Specific Tool Names to Configuration

You must add each specific MCP tool name you want to use to .claude/settings.local.json.

Step 1: Discover All Available Tool Names

To find all available MCP tools for your configured servers, ask Claude Code:

List all available MCP tools from the following servers and provide their exact tool names in the format mcp__<server>__<tool>:
- clear-thought
- memory-bank
- mcp-atlassian
- figma

For each server, examine the function definitions available in the environment and extract the complete list of tool names.

Alternatively, check the tools list at the top of any Claude Code conversation - you'll see all available tools listed in the format mcp__<server>__<toolname>.

Step 2: Common Tools for This Project

Each MCP server provides specific tools. Here are the commonly used tools:

clear-thought server:

mcp__clear-thought__sequentialthinking
mcp__clear-thought__mentalmodel
mcp__clear-thought__debuggingapproach
mcp__clear-thought__collaborativereasoning
mcp__clear-thought__decisionframework
mcp__clear-thought__metacognitivemonitoring
mcp__clear-thought__scientificmethod
mcp__clear-thought__structuredargumentation
mcp__clear-thought__visualreasoning

memory-bank server:

mcp__memory-bank__get_memory_bank_status
mcp__memory-bank__read_memory_bank_file
mcp__memory-bank__write_memory_bank_file
mcp__memory-bank__list_memory_bank_files
mcp__memory-bank__track_progress
mcp__memory-bank__update_active_context
mcp__memory-bank__log_decision

mcp-atlassian server:

mcp__mcp-atlassian__jira_get_issue
mcp__mcp-atlassian__jira_ls_issues
mcp__mcp-atlassian__jira_get_project
mcp__mcp-atlassian__jira_ls_projects
mcp__mcp-atlassian__jira_add_comment
mcp__mcp-atlassian__jira_ls_comments

figma server:

mcp__figma__get_design_context
mcp__figma__get_screenshot
mcp__figma__get_metadata
mcp__figma__get_variable_defs
mcp__figma__get_code_connect_map
mcp__figma__add_code_connect_map
mcp__figma__create_design_system_rules

Note: This list may not be exhaustive. New tools may be added by MCP server updates. Use Step 1 to discover all available tools in your environment.

Step 3: Edit `.claude/settings.local.json`

Add the specific tool names to the permissions.allow array:

Location: .claude/settings.local.json (in your project root)

{
  "permissions": {
    "allow": [
      // ... existing permissions ...

      // Clear-thought tools (add the ones you use)
      "mcp__clear-thought__sequentialthinking",
      "mcp__clear-thought__mentalmodel",

      // Memory-bank tools (add the ones you use)
      "mcp__memory-bank__get_memory_bank_status",
      "mcp__memory-bank__read_memory_bank_file",
      "mcp__memory-bank__write_memory_bank_file",

      // Jira tools (add the ones you use)
      "mcp__mcp-atlassian__jira_get_issue",
      "mcp__mcp-atlassian__jira_ls_issues",

      // Figma tools (add the ones you use)
      "mcp__figma__get_design_context",
      "mcp__figma__get_screenshot"
    ],
    "deny": [],
    "ask": []
  }
}

Step 4: Reload Cursor

After editing the configuration:

Press Cmd+Shift+P (Mac) or Ctrl+Shift+P (Windows/Linux)
Type "Developer: Reload Window"
Press Enter

Important: Settings changes do NOT take effect until you reload. If tools still prompt after adding them, you likely forgot to reload.

Required MCP Tools for Mandatory Supervisor Protocol

If your project uses the Mandatory Supervisor Protocol (see .cursor/rules/mandatory-supervisor-protocol.mdc), you MUST add at minimum these specific tools:

{
  "permissions": {
    "allow": [
      "mcp__clear-thought__sequentialthinking",     // REQUIRED: Used before every action
      "mcp__memory-bank__get_memory_bank_status",   // REQUIRED: Task state checking
      "mcp__memory-bank__track_progress",           // REQUIRED: Progress tracking
      "mcp__mcp-atlassian__jira_get_issue",        // REQUIRED: Jira ticket access
      "mcp__figma__get_design_context"              // REQUIRED: Design access for UI tasks
    ]
  }
}

Why these are required:

sequentialthinking - The protocol mandates using this tool before ANY action
memory-bank tools - Required for documenting analysis and maintaining task state
jira_get_issue - Protocol requires reading Jira tickets before planning (dependency order step 0)
get_design_context - Protocol requires checking Figma designs for UI tasks (dependency order step 1)

Adding Tools Incrementally

Recommended approach: Add tools as you need them rather than adding every tool upfront.

Method 1: Learn from permission prompts

When Claude tries to use an MCP tool and gets prompted
Note the exact tool name from the prompt (e.g., mcp__memory-bank__write_memory_bank_file)
Add that specific tool to .claude/settings.local.json
Reload Cursor
The tool will now work without prompting

Method 2: Proactive discovery

Ask Claude Code: "List all MCP tools available in my environment with their exact names"
Claude will examine available functions and provide the complete list
Add all tools you trust to .claude/settings.local.json
Reload Cursor once

Method 3: Check tool list in conversation

At the top of any Claude Code conversation, you can see all available tools
Scroll through the list and identify tools starting with mcp__
Copy the exact names to your configuration

Verifying Permission Configuration

After configuring permissions:

Reload Cursor (required for changes to take effect)
Test a tool:
- Ask Claude to use a tool you added (e.g., mcp__clear-thought__sequentialthinking)
- It should execute WITHOUT prompting for permission
- If prompted, check your configuration again

Check active configuration:

cat .claude/settings.local.json | jq '.permissions.allow'

Security Best Practices

When to allowlist:

✅ Trusted MCP servers you or your team installed
✅ Protocol-required tools (clear-thought, memory-bank, etc.)
✅ Frequently-used tools that interrupt workflow

When NOT to allowlist:

❌ Experimental or untested MCP servers
❌ Third-party MCP servers from unknown sources
❌ MCP tools that perform destructive actions without review

Remember: Allowlisting means Claude Code can use these tools without asking for permission. Only allowlist tools you trust completely.

Troubleshooting Permission Issues

Problem: Tools still prompt after adding to configuration

Solution 1: Reload Cursor window (Cmd+Shift+P → "Developer: Reload Window")
Solution 2: Check JSON syntax is valid (no trailing commas, matching quotes)
Solution 3: Verify tool name matches EXACTLY (case-sensitive, double underscores)

Problem: Not sure what the exact tool name is

Solution: Let the prompt happen once, note the EXACT tool name shown in the prompt, then add it to config

Problem: Too many tools to add manually

Workaround: This is a known limitation due to bug anthropics/claude-code#3107
Current status: Wildcards don't work; specific tool names are required
Subscribe to updates: Follow the GitHub issue for notification when wildcards are fixed
Tip: Use the proactive discovery method (Step 1) to get all tool names at once, then copy-paste into config

Problem: How do I find ALL available MCP tools?

Solution: Ask Claude Code to list all MCP tools: "List all available MCP tools in this environment with exact tool names in format mcp__<server>__<tool>"
Claude will examine the function definitions and provide a complete list
This is more reliable than checking documentation, as it reflects your actual environment

Security Note

Never commit sensitive tokens to version control!

Store tokens in environment variables
Use .gitignore to exclude configuration files with secrets (especially .claude/settings.local.json)
Regenerate tokens immediately if accidentally exposed
Add .claude/settings.local.json to .gitignore for local-only configuration

mpalpha/migrate-mcp-servers.md

Migrating MCP Servers from Cursor to Claude Code

Table of Contents

Quick Start: Copy-Paste Prompt

⬇️ COPY FROM HERE ⬇️

⬆️ COPY TO HERE ⬆️

What Happens After Pasting

Manual Migration Instructions

Understanding the Migration

Migration Steps

1. Locate the Claude CLI Binary

2. Check Your Existing Cursor MCP Configuration

3. Migrate MCP Servers Using the CLI

4. Set Up Your Environment

5. Migration Examples

Example: Migrate GitHub MCP Server

Example: Migrate Memory Bank MCP Server

Example: Migrate Atlassian MCP Server

6. Other Common Server Migrations

7. Verify Migration Status

8. Reload Cursor

9. Post-Migration Setup

Useful CLI Commands

List all configured servers

Get details about a specific server

Remove a server

Reset project-scoped server approvals

Configuration File Location

Troubleshooting

Server shows as "Failed to connect"

No MCP tools available after reload

Cursor's mcp.json file not being read by Claude Code

How do I know which servers need migration?

Can I keep using Cursor's native MCP servers alongside Claude Code?

MCP Servers Connected But Tools Return "No Data" or "Not Found"

Figma MCP: "No session found for sessionId" Error

MCP Tool Permission Configuration

The Permission Prompt Problem

⚠️ CRITICAL: Wildcard Permissions Don't Work

Solution: Add Specific Tool Names to Configuration

Step 1: Discover All Available Tool Names

Step 2: Common Tools for This Project

Step 3: Edit .claude/settings.local.json

Step 4: Reload Cursor

Required MCP Tools for Mandatory Supervisor Protocol

Adding Tools Incrementally

Verifying Permission Configuration

Security Best Practices

Troubleshooting Permission Issues

Security Note

Reference

mpalpha commented Oct 23, 2025

Uh oh!

mpalpha commented Oct 23, 2025

Uh oh!

mpalpha commented Nov 14, 2025

Uh oh!

Step 3: Edit `.claude/settings.local.json`