An AI agent can create finished videos when you give it video tools over MCP. The Model Context Protocol (MCP) is the open standard that lets an AI assistant call real software, and ViralMint — a free, open-source desktop video pipeline from the makers of OpenClaw Easy — ships an MCP server built in. It exposes 100+ tools: download a source video, extract the most viral clips, generate AI video, images, music and voiceovers, burn in captions, and export platform-ready files. Connect that server to the OpenClaw Desktop App and those tools become reachable from the messaging apps you already use — you text your agent on WhatsApp or Telegram, and it drives the video pipeline on your computer. This tutorial walks the exact setup we verified end to end, including the two gotchas that cost us time: a trailing slash and a tool allowlist.
What you're building
The flow has three parts, all running locally on your machine:
- OpenClaw Easy runs your AI agent and connects it to WhatsApp, Telegram, Slack, Discord, Feishu, or Line.
- ViralMint runs the video pipeline and exposes it as a local MCP server on
http://localhost:16888/mcp/. - MCP is the wire between them: OpenClaw's MCP client discovers ViralMint's tools and hands them to the model.
Once wired, a message like “clip the best three shorts out of this YouTube link and caption them” becomes a series of real tool calls — download_video, extract_viral_clips, tool_captions, export_video — and you get files back, not a description of files.
Step 1: Install both apps
Both are free, open-source desktop apps. Download OpenClaw Easy for macOS or Windows, and ViralMint for macOS, Windows, or Linux. Launch both. ViralMint's backend (and its MCP server with it) starts with the app — there is no separate server process to manage.
Step 2: Copy the ViralMint MCP token
ViralMint protects its MCP endpoint with a Bearer token generated on first run. Open ViralMint's MCP page (in Settings) — the Setup tab shows the endpoint and the token, with one-click install snippets for several MCP clients. Copy the token.
Step 3: Register the server with OpenClaw
From a terminal (or paste the equivalent into the OpenClaw Control UI under Settings › MCP):
$ openclaw mcp add viralmint \
--url "http://localhost:16888/mcp/" \
--transport streamable-http \
--header "Authorization=Bearer <your-token>"The trailing slash is not optional. ViralMint mounts its MCP server at /mcp/. Register http://localhost:16888/mcp (no slash) and the probe fails with {"detail":"Method Not Allowed"}. This was the first thing that bit us.
Step 4: Verify the connection
$ openclaw mcp probe viralmintA healthy probe reports the tool inventory — when we ran it, 118 tools, plus MCP resources and prompts, with zero diagnostics. You'll see names like viralmint__extract_viral_clips, viralmint__generate_ai_video, viralmint__tool_captions, and viralmint__quick_balance.
Step 5: Allow the tools (the second gotcha)
OpenClaw configs can pin the agent's tool surface with a tools.allow list — OpenClaw Easy does this by default to keep the messaging agent's tool list tight. When that list exists, MCP tools are silently stripped until you allow them. Add "viralmint__*" to tools.allow in your OpenClaw config (openclaw.json, or the Control UI's config editor):
"tools": {
"allow": ["exec", "group:web", "viralmint__*"]
}If your config has no tools.allow list, skip this step — configured MCP tools appear in the normal tool profiles automatically.
Step 6: Ask for work from any channel
Message your agent on any connected channel. Start with a free, read-only call to prove the loop — this is the exact test we ran, and the agent answered with live account data (balance, video library count, and the three most recent renders):
You: Check my ViralMint balance and list my last 3 videos.
Agent: Your cloud balance is $0.19. You have 419 videos; the 3 most
recent are 5-second 9:16 AI clips created July 12 ...Then ask for real output. Good first requests, mapped to the tools they trigger:
| You say | ViralMint tools the agent calls |
|---|---|
| “Clip the best 3 shorts out of this YouTube link” | download_video → extract_viral_clips |
| “Turn this article into a 30-second vertical video” | tool_url_to_short |
| “Make a brainrot-style short about X” | tool_brainrot_video |
| “Caption and reframe this to 9:16, then export” | tool_captions → tool_reframe → export_video |
| “What's trending in my niche this week?” | scout_trending (free) |
What it costs
The wiring, the apps, and the MCP server are free. Read-only tools — listing videos, balance checks, cost estimates, trend scouting, channel analysis — cost nothing to call. AI generation and rendering spend ViralMint's prepaid cloud credits, and the agent can call tool_cost_estimate first, so you can tell it to always check the price before rendering. There is no subscription on either side.
Troubleshooting
| Symptom | Cause and fix |
|---|---|
Method Not Allowed on add/probe |
Missing trailing slash — the URL must end in /mcp/ |
| Probe succeeds, agent says it has no viralmint tools | Your config's tools.allow strips them — add "viralmint__*" (Step 5) |
| Probe can't connect at all | ViralMint isn't running — its MCP server lives inside the app |
| 401 Unauthorized | Token mismatch — re-copy it from ViralMint's MCP setup tab |
| Server seems to vanish after idle time | Normal — idle MCP runtimes are reaped and reconnect on next use |
Frequently asked questions
Can an AI agent really create videos on its own?
Yes, with the right tools attached. A model alone only writes text; give it a video pipeline over MCP and it can download sources, pick viral clips, generate AI footage and voiceovers, caption, and export finished files. The agent decides which tools to call; you decide which tools exist.
Is this setup free?
The wiring is free — both apps are free and open-source, and read-only tools cost nothing. AI generation and rendering spend ViralMint's prepaid credits, with a cost-estimate tool the agent can call before rendering. No subscriptions.
Does this work with Claude Code or Cursor instead of OpenClaw?
Yes — ViralMint's MCP server is standard Streamable HTTP and its setup tab generates config snippets for Claude Code, Claude Desktop, Cursor, Windsurf, and Gemini CLI. OpenClaw's difference is the messaging layer: the same tools become reachable from WhatsApp, Telegram, Slack, or Discord.
Which AI models can drive the video tools?
Any model your OpenClaw agent runs — Claude, GPT, Gemini, or a local Ollama model. MCP tool calls are model-agnostic; we verified this tutorial's flow with Gemini as the agent model.
Can I run video jobs on a schedule?
Yes. Combine this setup with OpenClaw Easy's cron jobs — for example, a weekly job that scouts trending topics in your niche and drafts a short for your review.