Create MCP

What This Command Does

In plain English: A guided wizard that helps you create and integrate an MCP server into Dex. No coding knowledge required - you describe what you want, we build it together.

When to use it:

• You want to connect Dex to an external service (calendar, email, CRM, API)
• You have data somewhere that would be useful in Dex
• You want to automate interactions with a tool you use regularly

How to run it:

/create-mcp                    # Starts the wizard from the beginning
/create-mcp "calendar"         # Jumps ahead with a service hint

Why MCP Matters: Probabilistic AI vs Deterministic Tools

The Problem with AI Alone

AI models like Claude are fundamentally probabilistic - they generate responses by predicting the most likely next token based on patterns in their training data. This is powerful for reasoning and language, but dangerous for facts:

Without MCP, AI can only:

• Guess based on general knowledge
• Hallucinate plausible-sounding but wrong answers
• Admit it doesn't have access

What MCP Actually Does

MCP (Model Context Protocol) provides guardrails and structure for AI interactions with external systems:

┌─────────────────────────────────────────────────────────────┐
│                    YOUR QUESTION                            │
│         "What features are customers using most?"           │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    AI REASONING                             │
│  "I need support ticket data. I have a Zendesk MCP tool    │
│   called `get_ticket_stats`. Let me call it..."            │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    MCP TOOL CALL                            │
│  Tool: get_feature_usage                                    │
│  Input: { "days": 30, "limit": 10 }                        │
│  ─────────────────────────────────────────────────────────  │
│  │ GUARDRAILS:                                          │  │
│  │ ✓ Defined input schema - can't send bad data         │  │
│  │ ✓ Authenticated connection - uses real credentials   │  │
│  │ ✓ Structured output - returns consistent format      │  │
│  │ ✓ Deterministic - same query = same results          │  │
│  └──────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    REAL DATA RESPONSE                       │
│  { "features": [                                            │
│    { "name": "Dashboard", "usage": 89% },                   │
│    { "name": "Reports", "usage": 67% },                     │
│    { "name": "Guides", "usage": 45% }                       │
│  ]}                                                         │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    AI SYNTHESIS                             │
│  "Your top 3 features by usage are Dashboard (89%),        │
│   Reports (67%), and Guides (45%). Dashboard dominates -   │
│   consider investing more there."                          │
└─────────────────────────────────────────────────────────────┘

The Key Insight

MCP doesn't make AI smarter - it gives AI reliable ways to get real data. The AI still does reasoning, synthesis, and explanation. But the facts come from deterministic tool calls, not probabilistic generation.

Entry Point

If no arguments provided:

🔌 **MCP Server Creation Wizard**

MCP (Model Context Protocol) lets Dex connect to external tools and services. Instead of guessing or saying "I don't have access", AI can query real data and give you accurate answers.

**The difference MCP makes:**
- ❌ Without: "I'd estimate you have around 10-15 support tickets..."
- ✅ With: "Zendesk shows 32 tickets, 12 high priority, avg response time 2.3hrs"

**Examples of what you can build:**
- 📅 Calendar → "What meetings do I have tomorrow? Who's attending?"
- 📧 Email → "Find emails from Sarah about the Q1 roadmap"
- 💬 Slack → "What did #product-team discuss today?"
- 📊 Analytics → "Show me feature adoption for our enterprise tier"
- 🔗 Any API → If it has an API, we can probably connect it

**Benefits:**
- Real data, not AI guessing
- Guardrails prevent hallucination
- Live queries, not stale training data
- Structured tools with validation

**This wizard will:**
1. Help you describe what you want to connect
2. Design the integration together
3. Generate the MCP server code
4. Integrate it into Dex
5. Update all documentation so future sessions know how to use it

Ready to get started? **What would you like to connect Dex to?**

(Just describe it in plain English — e.g., "my Google Calendar", "Notion database", "company CRM")

If service hint provided:

Skip education and jump to Phase 1 with the hint as starting context.

Phase 1: Understand the Connection

Goal: Get clarity on what service and what data

Ask these questions (adapt based on what's already known):

Question 1: Service identification

What service or tool do you want to connect?

Examples:
- A specific app (Google Calendar, Notion, Salesforce)
- A type of data (my emails, my tasks, my documents)
- An API you have access to

Your answer:

Question 2: Authentication

How do you currently access this service?

1. I log in with username/password
2. I have an API key
3. It uses OAuth (Google, Microsoft login)
4. It's a local file or database
5. Something else

This helps me understand what authentication we'll need.

Question 3: Data of interest

What specific information do you want Dex to access?

Be specific about:
- What types of data (events, messages, records)
- What you'd want to read vs. write
- Any specific fields that matter most

Example: "I want to see my calendar events - title, time, attendees. Just reading, no need to create events."

Question 4: Use cases

How would you actually use this in practice?

Give me 2-3 example questions or commands you'd want to ask:
- "Show me today's meetings"
- "Find emails from [person] about [topic]"
- "What's the status of [account]?"

This shapes what tools we'll build.

After Phase 1, summarize:

**Understood. Here's what we're building:**

📦 **Service:** [service name]
🔐 **Auth method:** [auth type]
📊 **Data access:** [read/write + what data]
🎯 **Primary use cases:**
1. [use case 1]
2. [use case 2]
3. [use case 3]

Does this capture what you want? (yes / let me clarify)

Phase 2: Design the Tools

Goal: Define the specific MCP tools to build

Based on use cases, propose tool designs:

**Proposed MCP Tools**

Based on your use cases, here's what I suggest building:

| Tool Name | What It Does | Example Usage |
|-----------|--------------|---------------|
| `[tool_1]` | [description] | "[natural language example]" |
| `[tool_2]` | [description] | "[natural language example]" |
| `[tool_3]` | [description] | "[natural language example]" |

**Input parameters for each:**

### `[tool_1]`
- `param_1` (required): [description]
- `param_2` (optional): [description]

### `[tool_2]`
...

**Questions:**
1. Do these tools cover your use cases?
2. Should any tool do more or less?
3. Are there additional scenarios I missed?

Iterate until user confirms design

Keep refining based on feedback. Ask focused questions:

• "Should [tool] also support [capability]?"
• "What happens if [edge case]?"
• "Do you need to filter by [field]?"

Confirm before implementation:

**Final Tool Design**

We're building an MCP server called `[server-name]` with:

| Tool | Purpose | Inputs |
|------|---------|--------|
| [tool] | [purpose] | [inputs summary] |

**Authentication:** [method + what user needs to provide]
**Configuration:** [any env vars or config needed]

Ready to build? (yes / let me adjust)

Phase 3: Implementation

Goal: Generate the MCP server code

Step 3.1: Create the server file

Generate Python code following the pattern in core/mcp/work_server.py:

#!/usr/bin/env python3
"""
MCP Server for [Service Name]
[Brief description of what this server does]

Tools:
- [tool_1]: [description]
- [tool_2]: [description]
"""

import os
import json
import logging
from typing import Dict, List, Optional, Any
from datetime import datetime

from mcp.server import Server, NotificationOptions
from mcp.server.models import InitializationOptions
import mcp.server.stdio
import mcp.types as types

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Configuration from environment
[RELEVANT_CONFIG_VARS]

# ============================================================================
# SERVICE CLIENT
# ============================================================================

class [ServiceName]Client:
    """Client for interacting with [Service]"""
    
    def __init__(self):
        [initialization code]
    
    [methods for each operation]

# ============================================================================
# MCP SERVER
# ============================================================================

app = Server("[server-name]-mcp")
client = [ServiceName]Client()

@app.list_tools()
async def handle_list_tools() -> list[types.Tool]:
    """List all available tools"""
    return [
        types.Tool(
            name="[tool_name]",
            description="[tool description]",
            inputSchema={
                "type": "object",
                "properties": {
                    [property definitions]
                },
                "required": [required fields]
            }
        ),
        # ... more tools
    ]

@app.call_tool()
async def handle_call_tool(
    name: str, arguments: dict | None
) -> list[types.TextContent | types.ImageContent | types.EmbeddedResource]:
    """Handle tool calls"""
    
    if name == "[tool_name]":
        [implementation]
        return [types.TextContent(type="text", text=json.dumps(result, indent=2))]
    
    # ... more tool handlers
    
    return [types.TextContent(type="text", text=f"Unknown tool: {name}")]

async def _main():
    """Async main entry point"""
    logger.info("Starting [Service] MCP Server")
    
    async with mcp.server.stdio.stdio_server() as (read_stream, write_stream):
        await app.run(
            read_stream,
            write_stream,
            InitializationOptions(
                server_name="[server-name]-mcp",
                server_version="1.0.0",
                capabilities=app.get_capabilities(
                    notification_options=NotificationOptions(),
                    experimental_capabilities={},
                ),
            ),
        )

def main():
    """Sync entry point"""
    import asyncio
    asyncio.run(_main())

if __name__ == "__main__":
    main()

Save to: core/mcp/[service_name]_server.py

Step 3.2: Update requirements

Add any new dependencies to core/mcp/requirements.txt

Step 3.3: Create launcher script (if needed)

Create core/mcp/run_[service_name].sh:

#!/bin/bash
# Launch [Service] MCP server
cd "$(dirname "$0")"
source venv/bin/activate 2>/dev/null || true
python [service_name]_server.py

Make executable: chmod +x run_[service_name].sh

Tell user what was created:

**Server Created!** ✅

Files generated:
- `core/mcp/[service_name]_server.py` — The MCP server
- `core/mcp/requirements.txt` — Updated with dependencies

**Before we integrate, you'll need to:**

[Auth-specific instructions based on Phase 1]

Examples:
- For API key: "Add [SERVICE]_API_KEY to your environment or .env file"
- For OAuth: "Run the OAuth setup script: `python core/mcp/setup_[service]_auth.py`"
- For local: "No additional setup needed"

Let me know when you're ready to integrate.

Phase 4: Integration

Goal: Connect the MCP server to Dex

Step 4.1: Update CLAUDE.md

Add to the Integration Options section or create new MCP section:

### [Service Name] Integration

**Server:** `core/mcp/[service_name]_server.py`
**Purpose:** [what it does]

**Available Tools:**

| Tool | What It Does | Example |
|------|--------------|---------|
| `[tool]` | [description] | "[natural example]" |

**Configuration Required:**
- `[ENV_VAR]`: [description]

**Usage examples:**
- "[natural language request]" → uses `[tool]` tool
- "[another request]" → uses `[another_tool]` tool

Step 4.2: Add MCP Instructions (if not present)

Check if CLAUDE.md has mcp_instructions section. If not, add:

<mcp_instructions>
### [service-name]-mcp

[Description of the server and when to use it]

**Tools:**
- `[tool_name]`: [when to use and what it returns]

</mcp_instructions>

Step 4.3: Update System Guide

Add to 06-Resources/Dex_System/Dex_System_Guide.md under Integration Options:

| **[Service]** | [Brief description of capabilities] |

And add a new section if significant:

### [Service] MCP Server

Server: `core/mcp/[service_name]_server.py`

#### Available Tools

| Tool | Purpose |
|------|---------|
| `[tool]` | [description] |

#### Usage

[How to use in natural language, what to expect]

#### Configuration

| Variable | Description |
|----------|-------------|
| `[ENV_VAR]` | [what it's for] |

Phase 5: Verification

Goal: Ensure everything is properly connected

Only a custom-* entry whose command is the current sys.executable and whose args contain exactly one local .py file can use either startup check below. remote, HTTP, npm, npx, and binary entries cannot be blessed; explain that they remain structural-only.

Step 5.1: Offer a one-off startup proof

Ask exactly once:

Want me to prove it starts? This runs it once from a private copy, with your user permissions, and trusts whatever it imports. The entry's configured env is ignored. (yes / no)

Only after an explicit yes, issue a fresh token bound to this one entry and pass it to the check. The checker consumes and deletes the token before validating or launching anything, so it cannot be reused. The token prevents the automatic/recurring health checks from ever launching a one-off custom server and makes each explicit approval single-use. It is not protection against another program running as you, which could run your code directly regardless:

DEX_MCP_ONCE_TOKEN=$(./.venv/bin/python core/utils/smoke.py \
  --issue-mcp-once-consent custom-[server-name]) || exit 1
./.venv/bin/python core/utils/smoke.py \
  --check-mcp-once custom-[server-name] \
  --consent-token "$DEX_MCP_ONCE_TOKEN"

Show the command result honestly. A refusal or failed handshake is not permission to try another command shape or issue another token. On no or an ambiguous answer, do not issue a token and continue without running anything. The one-off check always uses a temporary vault for both cwd and VAULT_PATH; it never launches the custom code against the live vault as its working directory.

Step 5.2: Offer recurring startup checks

First inspect the eligible entry without executing it:

./.venv/bin/python -m core.utils.trust_registry --inspect-mcp custom-[server-name]

Show the returned MCP name, vault-relative path, and sha256. Then ask:

Trust this exact file for recurring startup checks? This runs <vault-relative path> with your user permissions (nightly and in deep scans), and trusts whatever it imports. Dex will run only a private copy of the exact content whose sha256 is <sha256>. If the file changes, Dex refuses to run it until you bless it again.

Default: No. (yes / no)

On no or an ambiguous answer, do not create or modify System/trusted-mcps.yaml.

On an explicit yes, create the user-owned registry from its shipped template if absent, then bind the consent to the sha256 that was shown:

if [ ! -e System/trusted-mcps.yaml ]; then
  cp -- System/trusted-mcps.example.yaml System/trusted-mcps.yaml
fi
./.venv/bin/python -m core.utils.trust_registry \
  --bless-mcp custom-[server-name] --expected-sha256 <sha256>

If either command refuses, report its reason and leave the entry unblessed. Never hand-add a remote, HTTP, npm, npx, binary, flagged Python (-c or -m), absolute, or .. path.

Step 5.3: Finish the verification checklist

Run verification checklist:

**Integration Complete!** 🎉

Let me verify everything is in place:

✅ Server created: `core/mcp/[service_name]_server.py`
✅ Dependencies updated: `core/mcp/requirements.txt`
✅ CLAUDE.md updated with integration docs
✅ System Guide updated

**To start using it:**

1. Install dependencies (if new ones added):

cd core/mcp && pip install -r requirements.txt

2. Add your credentials:

export [ENV_VAR]="your-value-here"

Or add to your shell config / the vault-root `.env` file (keep `.env` owner-only: `chmod 600 .env`).

3. Configure Claude Desktop to use the server:
Edit `~/Library/Application Support/Claude/claude_desktop_config.json`:
```json
{
  "mcpServers": {
    "[server-name]": {
      "command": "python",
      "args": ["[full-path]/core/mcp/[service_name]_server.py"],
      "env": {
        "VAULT_PATH": "[full-path-to-dex]"
      }
    }
  }
}

4. Test it! Try asking:
   • "[example natural language query]"

The one-off and recurring checks above are always optional.

---

## Example Walkthrough

User: /create-mcp

Claude: [Shows education intro]

User: I want to connect my Todoist

Claude: Great choice! Let me understand what you need...

[Phase 1 questions]

User: I want to see my tasks, add new ones, and mark them complete. Example: "What's on my Todoist today?" or "Add 'buy groceries' to Todoist"

Claude: Understood. Here's what we're building:

📦 Service: Todoist 🔐 Auth method: API key (Todoist provides this) 📊 Data access: Read tasks, create tasks, update task status 🎯 Primary use cases:

1. "What's on my Todoist today/this week?"
2. "Add [task] to Todoist"
3. "Mark [task] as done in Todoist"

Does this capture what you want?

User: Yes, but also let me filter by project

Claude: Got it! Adding project filtering...

[Phase 2 - Tool design]

Proposed MCP Tools

[Continues through phases...]

---

## Behaviors

### Always Do
- Start with education for new users
- Confirm understanding before building
- Generate complete, working code
- Update ALL documentation (CLAUDE.md, System Guide)
- Provide clear setup instructions
- Offer to help test

### Never Do
- Skip the design phase
- Generate partial/placeholder code
- Forget to update documentation
- Assume authentication works without explaining setup
- Create tools without clear use cases

### If stuck on technical details
- Search for the service's API documentation
- Check if an existing Python library handles auth
- Offer simpler alternatives if complexity is high

---

## Integration Checklist

After completion, verify:

- [ ] Server file created in `core/mcp/`
- [ ] Requirements.txt updated
- [ ] CLAUDE.md has integration documentation
- [ ] System Guide updated with new capabilities
- [ ] Setup instructions are clear and complete
- [ ] Example queries provided for testing

### Analytics (Required)

- [ ] Events defined for key tools (e.g., `{tool}_used`)
- [ ] Added checkbox to `System/usage_log.md` (Integrations section)
- [ ] Privacy verified: only tracks that feature was used, not content

See `.claude/reference/skill-analytics-checklist.md` for detailed guidance.
- [ ] User knows how to configure Claude to use the server

---

## Track Usage (Silent)

Update `System/usage_log.md` to mark MCP creation as used.

**Analytics (Silent):**

Call `track_event` with event_name `mcp_created` and properties:
- (no properties — do NOT include service names)

This only fires if the user has opted into analytics. No action needed if it returns "analytics_disabled".