mcp-server

name	mcp-server
description	MCP server for AI agents to create video ads via Streamable HTTP. Primary auth is OAuth 2.1 + Dynamic Client Registration (the modern Claude Desktop / Claude.ai / ChatGPT / Cursor "Connect" flow); long-lived API keys remain supported for headless clients.
version	1.1.0

Prizmad MCP Server

Prizmad exposes a remote Model Context Protocol (MCP) server. Any MCP-compatible host (Claude Desktop, Claude.ai web, ChatGPT, Cursor, Continue, Zed, Claude Code CLI, n8n, custom SDKs) connects and drives the full UGC video studio programmatically.

• Endpoint: https://prizmad.com/api/mcp
• Transport: streamable-http
• Server card: https://prizmad.com/.well-known/mcp/server-card.json
• Source code (stdio bridge for stdio-only clients): https://github.com/prizmad/Prizmad-MCP-server

Connect (primary path — OAuth Connect button)

Modern MCP clients click Add custom connector → enter the MCP URL and the OAuth flow runs automatically:

1. Client probes https://prizmad.com/api/mcp → 401 with WWW-Authenticate: Bearer resource_metadata="…".
2. Client reads /.well-known/oauth-protected-resource and /.well-known/oauth-authorization-server.
3. Client POST /oauth/register (RFC 7591 Dynamic Client Registration) — no manual app pre-registration needed.
4. Browser opens https://prizmad.com/oauth/authorize → user signs in (NextAuth) → consent screen → redirect back with code.
5. Client exchanges code for an RS256 JWT (audience https://prizmad.com/api/mcp).
6. Client calls https://prizmad.com/api/mcp with Authorization: Bearer <jwt>. Refresh tokens rotate; PKCE S256 mandatory.

Compatible clients: Claude Desktop, Claude.ai, ChatGPT, Cursor, Zed, Claude Code CLI. Same-origin design (authorize/token/register all live on prizmad.com) sidesteps the Claude.ai web bug that ignores cross-origin authorization endpoints.

Streamable HTTP config (when manual config is needed)

{
  "mcpServers": {
    "prizmad": {
      "transport": "streamable-http",
      "url": "https://prizmad.com/api/mcp"
    }
  }
}

The Connect flow handles the Bearer token automatically. To pre-fill an existing API key instead, add "headers": { "Authorization": "Bearer przmad_sk_live_..." }.

stdio-only clients (via the published bridge)

{
  "mcpServers": {
    "prizmad": {
      "command": "npx",
      "args": ["-y", "@prizmad/mcp-server"],
      "env": { "PRIZMAD_API_KEY": "przmad_sk_live_..." }
    }
  }
}

Or via the generic bridge: ["mcp-remote", "https://prizmad.com/api/mcp"].

Authentication options at a glance

Path	When	Token shape
OAuth 2.1 Authorization Code + PKCE + DCR	Interactive clients (Claude Desktop, etc.)	RS256 JWT (1 h), refresh rotated (30 d)
OAuth 2.0 client_credentials	Headless / server-to-server	HS256 JWT (1 h)
API key (Bearer)	Quickest for dev, scripts, internal tools	przmad_sk_live_...

All three resolve to the same per-user identity. Discovery: /.well-known/oauth-authorization-server, /.well-known/oauth-protected-resource, /.well-known/jwks.json.

Available tools

Tool	Auth	Purpose
list_templates	No	Full template catalog with features and token costs
list_avatars	No	Built-in avatar presets with recommended voices
recommend_template	No	Pick top-3 templates from intent + constraints (voice / avatar / duration / budget). Use this before create_video instead of guessing from the catalog.
list_my_videos	Yes	Recent projects with projectUrl / shareUrl / downloadUrl
upload_image	Yes	Upload an image (URL or base64) → returns prizmad.com-hosted URL for use as productImages / avatar
create_video	Yes	Start a render. Returns videoId.
get_video_status	Yes	Snapshot status by default; with wait: true it blocks server-side and streams notifications/progress until terminal (up to 10 min) — preferred over polling.
get_download_url	Yes	Authenticated download URL on prizmad.com for a completed video
create_video_batch	Yes	Launch up to 20 renders in parallel; pre-checks total token cost

list_templates and list_avatars are deliberately public so an agent can browse before authenticating.

Customization knobs on create_video (and per-item in create_video_batch)

All optional; omit any field for a randomised pick at render time.

captionStyle — on-video subtitles

Value	Look
classic	TikTok yellow highlight, white sans-serif (default fallback)
bold-impact	MrBeast / Hormozi — uppercase Anton with heavy black stroke and green highlight
karaoke	Active word colour-sweeps left-to-right as it's spoken
pop	Active word springs up in scale, inactive words dim
bounce	Active word hops up with a low-damping spring (Fredoka)
neon	Pink/cyan glowing halo (Audiowide)
typewriter	Characters reveal one-by-one with a blinking caret (JetBrains Mono)
glow	Soft white halo, gentle scale on the active word (Montserrat)

musicStyle — background music

energetic, friendly, professional, luxury, funny, cinematic, lo-fi, hip-hop, acoustic.

ctaStyle — end-card

Value	Look
classic	Solid dark backdrop, centered title + price, green "LINK BELOW" pill
blurred-photo	First creative blurred behind minimal cream title + price + circled down-arrow
dark-solid	Same minimal layout on a solid warm-dark backdrop

imageStyle — lighting/palette preset for AI creatives (10 options)

warm-golden, bright-neutral, cool-diffused, window-light, earthy-ambient, studio-clean, moody-dramatic, pastel-soft, nordic-minimal, sunset-warm.

Free-text prompt hints

imagePromptHint, videoPromptHint, musicPromptHint — each ≤ 400 chars, layered on top of style presets.

Output URLs (what to give the user)

get_video_status returns three URLs once rendering is complete — they map cleanly:

Field	Goes to
projectUrl	https://prizmad.com/projects/<id> — owner-only dashboard with player, remix, edit, asset library. Primary link when handing the result back to the signed-in user.
shareUrl	https://prizmad.com/share/<token> — public share page, useful only when forwarding the video to someone outside the account.
downloadUrl	https://prizmad.com/api/v1/videos/<id>/download — authenticated mp4 stream proxied via prizmad.com.

The raw Vercel Blob URL is never surfaced to the agent.

Typical flow

recommend_template ─► create_video ─► get_video_status (wait: true) ─► projectUrl + downloadUrl

1. (No auth) recommend_template({ intent, hasVoiceover, hasAvatar, targetDurationSec, maxTokens }) — pick a template.
2. (Optional) upload_image if the agent wants to attach product photos that aren't yet on the web.
3. create_video with templateId + productUrl (scraped) or explicit productTitle / productDescription / productImages + any customisation. Returns videoId.
4. get_video_status({ videoId, wait: true }) — blocks server-side, streams progress, returns the final payload when ready (3–8 minutes typical, capped at 10 min).
5. Hand projectUrl to the user.

For bulk campaigns use create_video_batch (1–20 per request).

Plan + token rules

• API video generation requires a Pro subscription (/api/v1/videos returns 403 with a clear upgrade message otherwise).
• Tokens come from the user's monthly plan first, then any top-up balance.
• Insufficient balance returns a 402 with a top-up URL.
• The MCP layer renders both errors as plain English so the assistant can repeat them verbatim.

Rate limits

• Per-key / per-OAuth-client rate limit on the MCP endpoint.
• Each successful create_video debits the caller's Prizmad token balance.