MODEL CONTEXT PROTOCOL

Let your agent make videos.

The edets MCP server exposes video generation to any agent that speaks the protocol. Claude, Cursor, Grok Bot — they all become video editors.

Install the MCP server

v1.0.0
NPXnpx @edets/mcp-server
PIPComing
CLAUDE DESKTOP CONFIG
{
  "mcpServers": {
    "edets": {
      "command": "npx",
      "args": ["@edets/mcp-server"],
      "env": {
        "EDETS_API_KEY": "edets_live_…"
      }
    }
  }
}

Requires Node 18+. Python package coming. Get your API key from edets.com/api/keys.

Tools your agent can call

generate_video

Create a single video clip

Upload a still, describe motion, pick edets Max or edets Motion. Returns a generation ID to poll.

agent: "Make a 5-second clip of this product photo. The model turns and looks at camera. Use edets Max at 720."
edets: Generation queued. ID: 7c1e4d2a-9b0f-4c3e-8a11-2f6d9e4b81c0. 5 seconds deducted from balance. Polling… Succeeded in 34s. Download: https://edets.com/api/v1/generations/…/clip
check_status

Poll generation status

Check if a clip is queued, running, succeeded, or failed. Failed jobs return seconds to the pool.

create_campaign

Multi-shot campaign

Generate multiple shots, then compose a master and export platform packs (ig_9x16, tt_9x16, fb_4x5).

compose_campaign

Stitch master + packs

After clips succeed, stitch the campaign master and platform packs. Campaigns do not auto-compose.

check_balance

Check remaining seconds

Agent checks prepaid seconds before generating. Returns seconds remaining and the dollar equivalent.

What your agent can do

Auto-generate ad variants

Agent reads your product sheet, generates 6 clips in different styles, and drops them in your Slack for review.

Video replies to email

Agent sees a support ticket, generates a personalized walkthrough video, and drafts the reply with the clip attached.

Batch social content

Agent takes your blog post, extracts 5 key points, generates one video per point, and schedules them across platforms.

A/B test thumbnails

Agent generates the same scene with 3 different motions, downloads all three, and reports which one it predicts will perform best.

Build on the protocol

Open source. MIT licensed. PRs welcome.

GitHub →Documentation