Model Context Protocol (MCP) in Practice: Connecting AI Agents to Brand Context

Why Chat Windows Are a Bottleneck for Agentic Workflows

Most teams use AI by opening a chat tab, pasting three paragraphs of background, and asking for a caption. The chat model has zero visibility into what is already scheduled in your queue, what performed well over the last 30 days, what articles exist on your website, or what your active Brand Voice forbids. Model Context Protocol (MCP) solves this by turning your brand workspace into a structured tool server.

1. Designing a Clean 3-Tier MCP Tool Surface

When exposing a content and marketing system through Model Context Protocol (MCP), group tools into three distinct security and operational tiers:

Tier 1 — Read & Inspection Tools (`get_workspace_context`, `get_content_strategy`, `get_analytics_overview`, `list_content_performance`, `list_articles`, `list_channels`, `preview_schedule`): Allow the agent to ground itself in live facts without mutating state. Tier 2 — Safe Draft & Audit Tools (`create_publication_draft`, `create_article_draft`, `create_content_strategy_draft`, `save_dashboard_audit`): Let the agent write visible drafts inside the workspace that never go live automatically. Tier 3 — Gated Action Tools (`prepare_publication_action`, `confirm_publication_action`): Require a two-phase token handshake before scheduling or publishing.

MCP SECURITY PRINCIPLE

Creating a draft (`create_publication_draft`) must never publish to a live social channel. Drafts are safe workspace artifacts for inspection.

// Step 1: Agent reads active workspace context & strategy
{ "tool": "get_workspace_context", "arguments": { "workspace": "fieldwork" } }
{ "tool": "get_content_strategy", "arguments": { "workspace": "fieldwork" } }

// Step 2: Agent checks upcoming publishing slots to avoid collisions
{ "tool": "preview_schedule", "arguments": { "channel": "instagram", "horizonDays": 7 } }

// Step 3: Agent creates a visible draft grounded in Strategy v2 & Brand Voice
{
  "tool": "create_publication_draft",
  "arguments": {
    "workspace": "fieldwork",
    "channel": "instagram",
    "caption": "Two days. Four litres.\n\nJacket, snacks and a repair kit. On your bars, instead of your back.\n\nSee the full packing layout in our guide.",
    "sourceRef": "art_drift4l_packing_guide"
  }
}

2. Combining MCP Tools with Portable Agent Skills

Raw tools tell an agent *what* API calls exist; Agent Skills (`SKILL.md`) teach the agent *how* to run a disciplined workflow. For example, a `mumro-analyze-performance` or `mumro-plan-campaign` skill instructs the agent to always call `get_workspace_context` and `list_content_performance` first, cite measured cohort numbers rather than inventing metrics, and save its proposal as a reviewable draft.

3. Scoped Authentication: API Keys, OAuth, and Workspace Isolation

An agent should only ever access the specific workspace and tool scopes you grant (`AG01`, `AG04`). Whether connected via scoped API keys or OAuth, every MCP call is bound to a single brand workspace so agency operators never risk cross-client data leakage.

Browser Chat Copy-Pasting vs. Connected MCP Agent Architecture