Masky API keys are generated at https://masky.ai/docs. No key? See "First: get an API key" in https://masky.ai/skill.md — an AI assistant can connect its user's account with one link, no developer page needed.
Masky MCP#
Masky exposes its capabilities as native MCP tools for any MCP client — Claude Desktop, Claude Code, Cursor, or Anthropic API tool use — two ways:
- Hosted endpoint —
https://masky.ai/api/mcp(streamable HTTP), nothing to install - Local server — the undefined npm package (stdio)
Both expose the same tools:
- List your avatars and publicly renderable avatars from other users
- Create avatars (image + personality + voice) and generate images
- Start live conversations and inject turns whose audio/video can be embedded in OBS or any web page
- Generate one-shot avatar TTS / talking-head clips and freeform videos
Hosted endpoint (no install)#
Claude Code, one line:
claude mcp add --transport http masky https://masky.ai/api/mcp --header "Authorization: Bearer mky_YOURKEY"
Any client that takes an HTTP/streamable-HTTP server config:
{
"mcpServers": {
"masky": {
"type": "http",
"url": "https://masky.ai/api/mcp",
"headers": { "Authorization": "Bearer mky_YOURKEY" }
}
}
}
Details:
- Stateless streamable HTTP, JSON responses — each
POST /api/mcpcall stands alone; no session to manage. initializeandtools/listwork without auth, so clients can connect and discover tools before a key exists.- Every
tools/callneedsAuthorization: Bearer mky_...— the same keys, OAuth access tokens, and billing as the REST API.
Local server (npm)#
Claude Code, one line:
claude mcp add masky -e MASKY_API_KEY=mky_... -- npx -y masky-mcp-server
Claude Desktop (or any client that takes a JSON server config):
{
"mcpServers": {
"masky": {
"command": "npx",
"args": ["-y", "masky-mcp-server"],
"env": { "MASKY_API_KEY": "mky_..." }
}
}
}
Restart the client; the Masky tools appear in the MCP tools menu. Requires Node ≥18.
Environment variables:
MASKY_API_KEY(required) — bearer token from https://masky.ai/docsMASKY_API_BASE(optional) — defaults tohttps://masky.ai
Tools#
list_my_avatars/list_public_avatars— your avatars, and avatars marked publicly renderable across all userscreate_avatar—{ displayName, imageUrl?, personalityPrompt?, voiceId?, knowledgeBaseUrls?, publiclyRenderable? }generate_image—{ prompt, aspectRatio?, model? }→ publicimageUrl(feed it tocreate_avatar).model: "ideogram"renders legible in-image text at 3x cost; withmodelomitted, "double-quoted" text in the prompt auto-routes there.edit_image—{ prompt, imageUrl|imageUrls[], aspectRatio?, model? }— edit/compose from 1–5 reference images;model: "qwen"for legible in-image text.generate_video/get_video/get_video_pricing— async freeform video renders (scenes, animated images, generated sound/music). Never for lip-sync — itsspeechTextis voice-over, not mouth-sync. Talking to camera →avatar_speakor a speak-mode turn; talking inside a moving scene → compose the frame withedit_image, then the REST speak endpoint withscenePrompt(see the "talking avatar in a complex scene" recipe in https://masky.ai/masky-api.md).list_voices/set_avatar_voice— preset + your trained voices; attach one to an avatarset_avatar_publicly_renderable— share / unshare an avatarstart_conversation—{ avatarOwnerUserId, avatarId }→liveUrl(embed in OBS / web) +conversationIdinject_turn—{ conversationId, userText, output?, mode?, speakerAvatarId?, userOutput?, avatarText?, voiceStyle? }. For a two-avatar podcast:speakerAvatarId+userOutput: "video"renders the user turn as a second avatar of yours, andavatarTextscripts the exact reply — both sides in one call.userTextmax 500 chars (≈ ~35s of speech); replies are auto-split into ≤3-sentence clips — a latency feature for the realtime/liveplayer, so the first clip plays while the rest renders.set_conversation_visibility— public/private toggleavatar_speak/get_avatar_speak— say one line (it wraps a one-turn conversation and returns theliveUrlplayer). TakestextORaudioUrl(exactly one):audioUrlis a public MP3/WAV URL the avatar lip-syncs directly — sung vocals, any pre-recorded voice, no TTS, no voice needed on the avatar, max 3 minutes, always one unsplit clip. When you want a single unsplit media file with no realtime player at all, call the REST one-shot endpoint instead —POST /avatars/{avatarId}/speak(also takestextoraudioUrl), pollGET /avatars/speak/{generationId}(see https://masky.ai/masky-api.md).
Voice delivery: speak/turn calls take an optional voiceStyle: { emotion, style, prosody: [] } (falls back per-field to the avatar's Voice-tab default), and the spoken text may embed inline cues like <|sfx:laughter|> or <|prosody:pause|> (stripped from captions). An sfx tag must immediately precede matching onomatopoeia to vocalize — <|sfx:laughter|>Haha! laughs; a bare tag before ordinary words is silently skipped. Full tag vocabulary: the Voice delivery section of https://masky.ai/masky-api.md.
register_oauth_client/get_sso_integration_guide— add "Login with Masky" SSO to your own site (full guide: https://masky.ai/masky-sso.md)
Example prompts#
- "List my Masky avatars."
- "Create a dragon avatar from this image: https://... Make it gruff and a bit medieval."
- "Start a conversation with my dragon avatar and ask what it had for breakfast — give me the embeddable live URL."
- "Make a 30-second podcast where my two avatars argue about tabs vs spaces."
Prefer raw HTTP instead? The same capabilities (plus translations, brain memory, the local-agent bridge, tribes, and credit gifts) are documented in the full API reference at https://masky.ai/masky-api.md, with a curl quickstart at https://masky.ai/masky-cli.md.
Generate or manage your Masky API key any time at https://masky.ai/docs