MCP & Skills Orchestration

Chain AI tools in declarative, human-readable workflows

The Problem

Modern AI agents have access to powerful tools:

MCP (Model Context Protocol) — External capabilities like databases, APIs, browsers
Skills — Knowledge files that guide how to use tools effectively

But there's no standard way to chain multiple tools, install them on-demand, or document the orchestration in a human-readable format.

IntentFlow as Orchestrator

IntentFlow provides a declarative way to orchestrate AI tooling — install, configure, and chain MCP servers and Skills in a single workflow.

Installing MCP Servers

## Step 1: Setup Database Access

### Dependencies
```bash
npm install -g @anthropic/mcp-server-postgres
```

### Configuration
Create MCP config at `~/.config/claude/`:
- Host: Use `DB_HOST` environment variable
- Credentials: `DB_USER` / `DB_PASSWORD`

### Task
Verify connection by listing available tables.

Using MCP in Workflows

## Step 2: Query Customer Data

### Task
Using the Postgres MCP connection from Step 1:
- Query all customers created in the last 30 days
- Include their order history
- Calculate lifetime value

### Save as
`/tmp/workflow/customers.json`

Combining MCP with Skills

## Step 3: Generate Report

### Dependencies
Read the skill at `/mnt/skills/public/docx/SKILL.md`

### Task
Following the docx skill instructions, create a professional
Word document with:
- Customer acquisition summary
- Revenue breakdown chart
- Recommendations section

### Save as
`/tmp/workflow/customer_report.docx`

Multi-Tool Pipelines

A single workflow can orchestrate multiple MCP servers:

Example: Competitive Analysis Pipeline

# Workflow: Competitive Analysis

## Step 1: Web Research
### Dependencies
npx @anthropic/mcp-server-puppeteer

### Task
Using the browser MCP, visit competitor websites and extract
pricing, features, and recent announcements.

---

## Step 2: Enrich with News
### Dependencies
pip install news-api-mcp

### Task
Search for recent news about each competitor.
Add sentiment analysis for each article.

---

## Step 3: Store in Database
### Dependencies
npm install -g @anthropic/mcp-server-postgres

### Task
Insert the collected data into the `competitive_intel` table.

---

## Step 4: Generate Insights
### Flexibility [autonomous]
Analyze the data and create visualizations that best
communicate market positioning and opportunities.

Why This Matters

❌ Traditional Approach

# Hard-coded, brittle
from mcp_postgres import connect
from mcp_browser import Browser
from news_api import search

db = connect(...)
browser = Browser()
# ... 200 lines of code

✅ IntentFlow Approach

## Step 1: Get data
### Task
Query the sales table...

## Step 2: Enrich
### Task
Search for context...

## Step 3: Report
### Task
Create a summary...

The IntentFlow workflow is:

Readable by non-programmers
Flexible — AI chooses implementation details
Self-documenting — the workflow IS the documentation
Adaptable — AI handles edge cases gracefully

Limitations

MCP Registration Timing

Most MCP servers need to be registered at client startup. IntentFlow handles this by:

Installation step — Install the MCP server package
Configuration step — Generate necessary config files
Session restart — User restarts client (if required)
Usage steps — Workflow continues with MCP available

For MCP servers that work as CLI tools or HTTP services, no restart is needed.

Skill Loading

Skills are loaded via the view tool at runtime. Explicitly instruct the LLM to read the skill:

### Dependencies
First, read the skill at `/mnt/skills/public/xlsx/SKILL.md`

Best Practices

Group related MCP setup — Install and configure in the same step
Verify before proceeding — Add success criteria to confirm tools work
Document requirements — Use the Meta section to list required MCP servers
Provide fallbacks — If an MCP fails, specify alternatives

## Meta
requires: Claude with computer use
mcp_servers: postgres, puppeteer, news-api

## Step 1: Setup
### If something goes wrong
- MCP not available → install it in Dependencies
- Connection failed → check credentials, retry once
- Still failing → skip this step, note in final report