Plug Switch into any AI.
The Switch MCP server lets Claude, ChatGPT, Cursor, Grok, Muse and any MCP client talk directly to your Switch studio. Generate images and video, lip sync and talking avatar clips, reuse your reference images, search your library, and check your balance, right from the chat. No copy and paste, no tab switching.
https://www.switchapp.ai/mcpConnect your AI.
Every one of these connects to the same endpoint. Pick yours and follow its steps.
Claude on the webThe fastest setup. One custom connector, sign in once.- Claude DesktopNative connector. No bridge, no terminal.
Claude Code (CLI)Add once at user scope, then sign in inside a session.
Codex (CLI)Native remote MCP with OAuth. Use one canonical connection only.
Kimi Code (Kimi 3 and Kimi 2)One MCP connection works with every Kimi model selected in Kimi Code.
Gemini CLINative remote HTTP MCP with automatic OAuth discovery.
CursorAdd an HTTP MCP server, for every project or for one.
Grok (grok.com and Grok Bot)Custom connector on grok.com. Grok Bot cannot open a browser, so it uses a Switch API key.- ChatGPT connectorRequires a paid ChatGPT plan with Developer mode enabled.
Muse (Meta)Ask Muse to create a Custom Connector for Switch.
DeepSeek modelsDeepSeek is a model API, not a remote MCP client by itself.
BytePlus Managed AgentAttach Switch as an MCP server plus the MCP toolset in a Managed Agent.
Inside Adobe.
Two free panels put the same brain inside the apps you already edit in. Each has its own page with the download and the steps.
What you can make
Every tool runs on your own Switch account after you sign in: images, video, audio, references, styles, and your whole library, straight from a chat.
Generate images across the Switch model lineup: GPT Image 2.5, Switch Pro, Switch Ultra, Nano Banana 2.1 (Google's newest), Nano Banana 2 and Switch Vintage (film look, works from a character it has been taught). Lock a subject, face, or product with reference images.
generate_image · explore_models
Text-, image-, frame-, or reference-to-video across Kling, Seedance, WAN 2.7, MiniMax H3 Max, and Switch Video. plus Switch Cinema (remake a clip with the person in your pictures, keeps the motion), motion control, omni audio-video, clip edits, and Topaz upscale.
generate_video · list_video_models · get_video_status · list_my_videos
Lip-sync a clip to audio, drive a talking-avatar video, and manage your custom voices.
lip_sync_video · talking_avatar_video · voice
Upload an image, video, or audio reference once and reuse it across generations. Image and video references sync into your live Studio reference strip.
upload_reference_asset · get_my_active_references · upload_media
Create and inspect persistent Editor projects, upload media directly, list the named editor workflows, run the Editor agent, approve guarded paid steps, and render through HyperFrames.
create_editor_project · get_editor_project · upload_editor_media · add_editor_library_assets · list_editor_workflows · run_editor_agent · get_editor_run · approve_editor_run · cancel_editor_run · render_editor_project
One-tap photographic and cinematic looks. apply a finished style stack to your prompt without hand-tuning.
apply_iphone_realism · apply_movie_scene · apply_high_fashion_editorial · apply_cinematic_anamorphic · apply_magic_hour_portrait · apply_product · apply_travel · apply_ugc · apply_wellness · apply_graphic_editorial_portrait
Browse folders and assets, search your library, show any asset inline, track generations, check your balance, and cancel a running task.
list_my_folders · list_my_assets · search_my_library · show_media · show_generation · list_generations · check_job_status · check_balance · cancel_my_task
Try these prompts
Say one of these to your connected AI. Each maps to a real tool.
Model limits
Straight from the engine. Accurate, not aspirational.
Switch Ultra and the image-edit models use your first 4 reference images.
5 or 10-second clips. Omni mode runs 3 to 15 seconds, takes up to 7 image references and one reference clip, and generates native audio.
4 to 15 seconds. Up to 9 image, 3 video, and 3 audio references (15 total); reference and omni modes need at least one image or video anchor.
5, 10, or 15-second clips. the longest durations Switch offers.
5 or 10-second clips (text, image, or clip-edit).
Whole seconds 5 to 15 at 480p or 768p, with native audio on every clip (dialogue, ambience, lip-sync; no off switch). Text, image (start + end frame) and reference modes; reference takes up to 9 pictures, 3 clips and 3 audio files (12 total), and the reference pictures and clip seconds are billed on top of the output.
Preview. The same 5 to 15 whole seconds at 480p or 768p with native audio, about twice as fast at half the price, targeting the 97th percentile of H3 Max quality. Text and image (start + end frame) only; there is no Turbo reference lane, so reference pictures, clips and audio go through MiniMax H3 Max.
The Switch editor (HyperFrames) combines clips and renders HTML-based motion graphics into one MP4.
Under the hood.
For the person wiring it up. Everyone else can skip this.
- Transport
- Streamable HTTP
- JSON-RPC over POST. SSE for long running tool calls. Clients negotiate with the
mcp-protocol-versionheader. - Sign in
- OAuth 2.1 with PKCE
- Dynamic client registration. Bearer tokens scoped per person. API keys also work for clients that cannot open a browser. No raw secrets in chat.
- Spec
- MCP 2025-06-18
- Follows the current MCP spec. Discovery via
/.well-known/oauth-protected-resource. Every tool call is logged for auditing.