---
type: "synthesis"
sources: ["cross-day"]
tags: ["mcp", "integration-pattern", "architecture"]
id: "arc-mcp-connective-tissue"
---
## MCP appears in 4 of 6 videos under different names

The Model Context Protocol is the single most important *invisible* primitive in the corpus. It's the reason Claude can stop being a chatbot and start being an orchestrator.

## The four MCP appearances

- **[[concept-higgsfield-mcp]] (Alex, Day 1)** — image/video generation models exposed as Claude tools. Setup: Settings → Connectors → add custom connector → paste URL → authenticate.
- **[[concept-mcp]] (Sabrina, Day 3)** — the protocol itself, with multiple servers in play: Claude for Chrome MCP, Perplexity MCP, Blotato MCP.
- **[[concept-custom-connectors-mcp]] (MAG, Day 4)** — the Blotato MCP at `https://mcp.blotato.com/mcp` added as a custom connector. Same install pattern as Alex's.
- **[[concept-claude-cowork]] (Dara, Day 6)** — uses *Connectors* (Chrome, Slack, Canva). The word "MCP" is not used explicitly but the architecture is identical. See [[prereq-chrome-connector]].

## CCC is the outlier that proves the rule

CCC (Day 2) does **not** use MCP for its core transcription bridge. Instead it uses [[concept-webhook-integration]] — Claude POSTs to an n8n webhook, n8n does the work, returns text. This is the pre-MCP integration pattern.

Reading the corpus chronologically by depth, MCP is *the answer* to the integration problem that CCC solves with raw webhooks. Both work; MCP is the more native, more discoverable, more revocable path.

## What every MCP server has in common

1. **A remote URL** the user pastes once.
2. **An authentication step.**
3. **A schema** Claude reads to know what tools the server offers.
4. **A natural-language invocation** — the user says "use Blotato to schedule…" and Claude routes the call.

This is the same pattern across Higgsfield, Blotato, Perplexity, and the Chrome connector. Once you understand one, you understand all.

## The integration-pattern lesson

MCP is the answer to the question CCC's architecture begs: *"isn't there a less brittle way to bridge Claude to external services than custom webhooks?"* Yes — and that's the trajectory of the entire corpus. See [[arc-skills-primitive-three-flavors]] for the *content* layer of this same architectural shift.

## Risks the corpus understates

- API changes break MCP servers silently.
- Auth tokens expire.
- Rate limits are per-server and unpublished.
- A misconfigured MCP can leak credentials.

None of the six videos benchmarks MCP reliability under load. Treat MCP-driven autonomy as load-tested only in demos, not in production.