Mcp Setup

MCP Setup

Overview

MCP turns AI agents into tool-using operators. The risk is that teams connect every tool, expose write access, and then wonder why the agent behaves unpredictably. Good MCP setup is not "give the agent more tools." It is tool curation, permission design, and auditability.

This skill configures MCP for GTM workflows where agents research accounts, enrich records, inspect CRM, draft messaging, and route tasks without uncontrolled external side effects.

When to Use

Use this skill when the user asks to "set up MCP for sales tools", "connect our CRM to Claude/Cursor", "give my agent enrichment capabilities", "build an MCP server for GTM tools", "configure MCP permissions", "connect LeadMagic MCP", "orchestrate multiple MCP servers", or "make agent tool use safer".

Authoritative Foundations

Model Context Protocol Specification

MCP standardizes how agents discover tools, resources, and prompts. The key design principle is clear tool descriptions and predictable tool inputs/outputs.

Anthropic — Tool Use Design Patterns

Tools should be narrowly described, unambiguous, and paired with verification steps. Agents need clear rules for when to call a tool and when to ask for confirmation.

Least Privilege Access Control

Agents should receive the minimum access needed for the current task. Read-only first, write access only where the business process explicitly requires it.

Audit Logging

Log every tool call: tool name, who triggered it, timestamp, and a short summary of inputs/outputs. Without this, you cannot debug mistaken CRM writes or sequence enrollments.

Prerequisites

• Agent platform that supports MCP
• List of tools to expose
• API credentials stored securely
• Permission model: read, draft, write, send
• Audit log destination
• Human approval policy for write actions

Step-by-Step Process

Phase 1: Inventory Tool Categories

Phase 2: Define Tool Scope

For each MCP server, document available tools, inputs required, outputs returned, side effects, rate limits, permission level, and human approval requirement. If a tool has side effects, its name and description should make that obvious.

Phase 3: Configure MCP Client

Typical configuration fields:

jsonCopy code
{
  "mcpServers": {
    "gtm-tools": {
      "command": "node",
      "args": ["server.js"],
      "env": {
        "API_KEY": "stored-in-secret-manager"
      }
    }
  }
}

Never hardcode API keys in repo files. Use environment variables or the agent platform's secret manager.

Plain MCP (BYOAI support): Remote server https://mcp.plain.com/mcp — OAuth via Plain account (same permissions as the user). Tools cover thread search, customer lookup, help center, draft replies (addGeneratedReply), and state changes (assign, label, done) with human approval before customer-facing sends. Pair with headless-support → references/byoai-headless-stack.md when stacking Attio + Plain + agent IDE.

Phase 4: Add Guardrails

Minimum policies: read-only by default, confirmation before CRM writes, confirmation before sequence enrollment, no autonomous sends, rate limits per tool, audit log for every tool call, and surfaced errors.

Phase 5: Pair with n8n for Batch Execution

MCP agents should not loop hundreds of enrich or CRM writes in chat. For approved batch jobs, agents POST to n8n webhook MCP-01 (n8n-toolkit) with HMAC + one-time approval_token. n8n runs deterministic pipelines with audit logs; agent uses MCP read tools to verify results.

Phase 6: Test End-to-End

Run a test workflow: research one account → enrich one contact → read CRM record → draft outreach → stop before sending → verify logs contain every tool call.

Output Format

markdownCopy code
# MCP Configuration Plan

## Servers
| Server | Tools | Permission | Side Effects | Approval Required |
|---|---|---|---|---|

## Tool Policies
- Read-only tools:
- Draft-only tools:
- Write tools:
- Forbidden actions:

## Client Config
[Configuration block with secrets referenced via env vars]

## Test Workflow
1.
2.
3.

## Observability
- Log destination:
- Fields captured:
- Alert conditions:

Quality Check

Before delivering, verify:

• Secrets are referenced through env vars, not hardcoded
• Tool permissions are least-privilege
• Side-effecting tools require confirmation
• Every tool has a clear description and input schema
• Audit logging captures tool name, input, output summary, timestamp, and actor
• Test workflow proves tools work end-to-end

Common Pitfalls

1. Exposing too many tools. Agents choose poorly when tool scope is noisy. Fix: only expose tools needed for the workflow.
2. Write access by default. A research agent should not mutate CRM. Fix: read-only first.
3. No confirmation gates. Agents can send or enroll by mistake. Fix: require user confirmation for external side effects.
4. Secrets in config. Public repos and logs leak keys. Fix: env vars or secret manager.
5. No logging. You cannot debug a tool call after the fact. Fix: log every call.
6. Ambiguous tool descriptions. Agents misuse tools with overlapping names. Fix: make tool names and descriptions specific.

Execution Artifacts

• references/framework-notes.md — named frameworks, citation anchors, and operating assumptions
• templates/output-template.md — copy-paste deliverable structure for the user
• scripts/check-output.py — local checklist validator for required sections This skill includes lightweight artifacts the agent can load on demand: Use the artifacts when the user asks for an implementation-ready deliverable, a repeatable workflow, or a quality check rather than generic advice.

Related Skills

• ai-sdr-setup — guardrails, pilot scope, and human handoff for agent workflows
• leadmagic-mcp — LeadMagic-specific MCP setup
• headless-support — Plain BYOAI stack, deflection funnel, KB requirements
• crm-integration — CRM read/write patterns and field mapping
• api-enrichment — enrichment API usage and batch patterns