MCP vs Skill vs API: How to Connect ViralNote to Your Agent

TL;DR: Use MCP when your host is Claude Desktop, Claude Code, or Cursor (15 native tools). Use the viralnote-skill when your agent reads SKILL.md and calls REST endpoints (OpenClaw, Claude skills CLI). Use the REST API directly for scripts, Zapier, Make, and n8n. All three hit the same ViralNote backend — pick by how your agent consumes tools.

ViralNote exposes three integration surfaces. Choosing the wrong one means extra config, missing tools, or fighting your host’s expectations. This guide is the decision tree we give every builder.

Quick decision tree

Does your client support MCP natively?
├─ YES → MCP server (@viralnote/mcp-server or HTTP endpoint)
└─ NO → Does it load Agent Skills (SKILL.md)?
    ├─ YES → npx skills add viralnote/viralnote-skill (or ClawHub)
    └─ NO → REST API (curl, SDK, Zapier, n8n)

Option 1: MCP server (recommended for Claude & Cursor)

What it is: A Model Context Protocol server — 15 typed tools (create_post, list_analytics, etc.) your host discovers automatically.

Best for:

• Claude Desktop / Claude Code (HTTP or OAuth)
• Cursor (Cursor Directory plugin)
• Any MCP-compatible IDE or agent runtime

Install highlights:

Host	How
Cursor	cursor.directory/plugins/viralnote
Claude + Smithery	smithery.ai/servers/viralnote/viralnote
Generic HTTP	https://dashboard.viralnote.app/api/mcp/mcp + Authorization: Bearer vnd_...
Stdio npm	npx -y @viralnote/mcp-server

Pros: Richest tool surface, no prompt engineering for endpoint shapes, works in chat UIs.

Cons: Requires MCP support in the host (not every runtime has it yet).

Deep dive: MCP server for social media.

Option 2: Agent skill (viralnote-skill)

What it is: An open-source SKILL.md pack that teaches the agent ViralNote’s REST API — when to call it, how to auth, workflow patterns, and guardrails (confirm before publish).

Best for:

• Claude Code / Claude Desktop with npx skills add
• OpenClaw via ClawHub: clawhub install @viralnote/viralnote
• Custom agents that execute HTTP but don’t speak MCP

Install:

npx skills add viralnote/viralnote-skill
export VIRALNOTE_API_KEY="vnd_..."

Pros: Works anywhere the agent can read skills and issue HTTP requests. MIT-licensed, forkable.

Cons: The model chooses endpoints from instructions — slightly more variance than MCP tool schemas. No automatic tool discovery UI.

Docs: viralnote.app/developers/skill · viralnote.app/agents

Option 3: REST API (scripts & no-code)

What it is: Versioned HTTP API at https://dashboard.viralnote.app/api/v1 with OpenAPI spec at /api/v1/openapi.

Best for:

• Cron jobs and internal scripts
• Zapier / Make / n8n (no agent reasoning layer)
• Mobile apps and custom dashboards
• Agents you build yourself with explicit function tools

Minimal example:

curl https://dashboard.viralnote.app/api/v1/posts \
  -H "x-api-key: $VIRALNOTE_API_KEY"

Pros: Universal, deterministic, easy to test and audit.

Cons: You wire each endpoint (or generate a client from OpenAPI). No chat-native “just ask” experience unless you add an agent layer on top.

Docs: viralnote.app/developers/docs

Comparison table

	MCP	Skill	REST API
Tools exposed	15 MCP tools	Agent reads API docs in skill	All v1 endpoints
Auth	API key or OAuth	VIRALNOTE_API_KEY	API key / Bearer
Host examples	Claude, Cursor	OpenClaw, Claude skills	Zapier, Python, n8n
Distribution	npm, Smithery, Cursor Directory	GitHub, ClawHub	OpenAPI
MIT open source	mcp-server	viralnote-skill	N/A (API product)

Can you use more than one?

Yes — common pattern:

• MCP in Cursor for day-to-day scheduling while you work
• REST + webhooks for n8n when a post publishes
• Skill on OpenClaw for overnight autonomous runs

Use separate API keys per environment so you can revoke independently.

Runtime-specific cheat sheet

Runtime	Use
OpenClaw	MCP HTTP (preferred) or ClawHub skill — OpenClaw setup
Hermes Agent	~/.hermes/config.yaml HTTP MCP — Hermes setup
Zapier / Make	REST API — no-code guide
n8n	REST + webhooks — n8n triggers

Get started

1. Create an API key
2. Pick your surface from the tree above
3. Try one workflow from three agent workflows

When in doubt: MCP if the client supports it, skill if it loads SKILL.md, API if you’re wiring automation yourself.