The video studio your AI agent can drive
ReelMint ships a built-in MCP server. Connect Claude Desktop, Claude Code, or Cursor with a workspace-scoped token, and “make a 60-second reel from my post” becomes a working command — the agent scripts, composes editable scenes, previews its own frames, narrates, and publishes to YouTube or Instagram Reels. You review; it does the clicking. A Short that takes an afternoon to hand-edit takes about ten minutes of your attention here — and the agent can draft overnight, leaving a reel for your morning review instead of posting blind.
What is MCP video generation?
MCP video generation is letting an AI agent build and publish a video by calling a video studio's tools over the Model Context Protocol— an open standard from Anthropic that lets assistants like Claude, Claude Code, and Cursor call external tools during a task. Instead of one “generate” button that returns a finished file, the studio exposes each step of the pipeline — script, scenes, narration, render, publish — as a separate tool the agent calls in order, handing back an editable result at every step so it (and you) can catch mistakes before anything ships. ReelMint's built-in MCP server works exactly this way: the agent drafts and composes; you review and approve; nothing posts unattended. Read the full explainer →
Built-in MCP server — works with Claude Desktop, Claude Code, Cursor, and any MCP-capable agent, scoped to your workspace.
Which agents can drive it?
Your token lives on the MCP & Agents page inside the studio — scoped to one workspace, rotatable in one click. In Claude Code: claude mcp add --transport http reelmint https://reelmint.io/api/mcp --header "Authorization: Bearer YOUR_TOKEN".
What can the agent actually do?
| Tool | What it does |
|---|---|
| draft_script | Write a scene-by-scene script from a topic, URL, or your knowledge base |
| author_scenes | Compose each line into motion-graphic scenes from a 100+ component catalog |
| preview_scene | Render film-strip frames and LOOK at them — the agent checks its own work |
| synthesize_scene_voice | Narrate any scene in the project's voice |
| sync_scene_duration | Lock a scene's timing to its narration — no manual nudging |
| learn_reference_style | Study a reference reel and distill its look into reusable style rules |
| compose_video | Render the full 9:16 Short on a fair queue, within your daily budget |
| publish_reel | Publish to YouTube (draft or scheduled) or Instagram Reels — under your review |
One tool = one job — your agent orchestrates the order. Long renders are the only background job (it polls for completion); every other step returns its result immediately.
What it is — and isn't
✓ Editable motion-graphic scenes (code, terminals, charts, diagrams) — not generative pixels.
✓ Publishes to YouTube and Instagram Reels today. TikTok is on the roadmap.
✓ Supervised autopilot: the agent drafts; you approve. Nothing auto-posts behind your back.
✓ BYOK — your own Gemini / OpenAI / ElevenLabs keys, at cost. No markup, no expiring credits, no watermark.
✗ No Sora/Veo-style generative footage, no avatars, no unattended posting. That's deliberate.
Questions agents ask
Which agents work with it?
Anything that speaks MCP: Claude Desktop, Claude Code, Cursor, Cline, Continue, or your own agent via the open protocol. One config entry connects them.
What can an agent actually do?
The full pipeline: draft scripts, author scenes from the component catalog, preview frames, synthesize narration, compose the video, and publish to YouTube or Instagram Reels — the same guarded endpoints the UI uses.
Is it safe to let an agent render and publish?
Yes — by design it's supervised, not blind. Renders run on a fair queue within a per-workspace daily budget, so a runaway loop can't melt the studio. Publishing lands as a draft you approve (or a scheduled post), never an unattended auto-post. Tokens are workspace-scoped and rotate in one click.
Is it generative video?
No — and that's the point. Every scene is a layered, editable motion-graphic composition (code cards, terminals, charts, diagrams), not baked pixels. Re-open any scene next month and change a variable name or chart value, then re-render. No Sora/Veo-style footage, no avatars.
Does it cost extra?
No. The MCP server is part of every workspace, including the free pilot. Generation runs on your own API keys (Gemini / OpenAI / ElevenLabs) at provider list price — zero markup, no credits, no watermarks.
Point your agent at a real studio
Free pilot, workspace-scoped token, daily budgets on. No credit card.
Claim a pilot spot