SSD-MCP Server

Endpoints (1 live)

• POST /mcp — create_text — $0.07 USDC (0.01 USDC on Base)

MCP tools (21)

soundside-mcp https://mcp.soundside.ai/mcp

• analyze_media — Analyze media: technical metadata, vision_qa, transcribe, detect_segments, detect_shots, export_edl.
• apply_effect — Cinematic effects: ken_burns (image→video motion), speed_ramp, film_grain, vignette. Returns resource_id.
• compose_media — Layer text, images, or video onto a base video: add_text, overlay, split_screen. Returns resource_id.
• compose_video — Compose a complete video from a timeline plan. Accepts a plan (sparse brief or detailed timeline), enriches it, generates assets in parallel, and assembles with transitions, audio mix, and overlays. Auto-creates a project/collection if none provided. Plans: minimal (just brief+style), moderate (segment outlines), or detailed (providers, prompts, timing). Server fills gaps. Async: returns pending resource_id. Check completion with lib_list (the resource's status); an MCP notification may also arrive, but don't rely on it. Revision: pass reuse_segments mapping indices to prior resource_ids (from metadata.composition.segment_resources). Only changed segments regenerate. Revise (sugar): pass revise_from=<parent composition UUID> with plan={} and regenerate=[indices]. The server rebuilds the plan and reuse map from the parent's durable checkpoint (owner-only): every segment NOT listed in regenerate reuses the parent's clip, and the parent's monolithic narration/music carry over as supp
• create_artifact — Create a business artifact: presentation (slides), chart (data viz), document (sections), or diagram (mermaid). Bundle mode: pass outputs list for multi-format generation. Supports brand kits. Provider 'gamma' for AI presentations.
• create_audio — Create audio: TTS, sound effects, voice cloning/design, voice listing, or the deprecated transcription compatibility shim. Use analyze_media(analysis_type='transcribe') for new transcription integrations.
• create_image — Create image from text prompt or reference. Sync (vertex, minimax, grok, creative_freedom) or async (alibaba — poll with lib_list).
• create_music — Create music from lyrics/prompt. Lyria is synchronous; Creative Freedom is authenticated-only and asynchronous. MiniMax music is unavailable and is not part of the public contract.
• create_text — Generate text via LLM chat completions. Pass messages or prompt.
• create_video — Create video from text/image. Async: returns pending resource_id, poll with lib_list until completed (5-10s intervals, 600s timeout). Supports character_reference for consistency.
• edit_audio — Edit audio on video: mix_audio, replace_audio, pad_audio. Returns resource_id.
• edit_video — Edit video: trim, concat, crossfade, adjust_speed, loop, color_grade, custom, burn_subtitles. Returns resource_id of new media.
• extract_media — Extract from media: extract_frame, extract_frames, extract_audio. Returns resource_id(s).
• lib_list — List library entities: projects, collections, resources, lineage, brand_kits, credits, or dashboard (composite load). Filter by project_id, collection_id, tags, mime_type_prefix. Supports pagination (limit/offset), sorting, and search.
• lib_manage — CRUD for library entities (projects, collections, resources, brand_kits). Large file upload: upload_initiate → PUT to signed URL → upload_complete. To share a resource: set visibility="public" → response includes a permanent short_url. For listing, use lib_list.
• lib_share — Manage project sharing: share by email, list shares, or revoke access. For download URLs, use lib_list with resource_id.
• list_adapters — List LoRA adapters mirrored into the Soundside library.
• manage_adapter — Manage a LoRA adapter: inspect, deploy, undeploy, delete, or select_checkpoint.
• publish_content — Read and manage a connected X account. action defaults to publish. Read actions return bounded posts/profiles and pagination tokens synchronously. Writes (publish/reply/edit/delete/like/unlike/repost/unrepost) require a unique idempotency_key and return a durable receipt: check its status with lib_list; a push may also arrive, but don't rely on it. Reuse a key only with the exact same request; never retry an unknown write using a new key. API-key/OAuth-client automation needs an explicit read, publish (publish/reply), or manage (other writes) grant. Authenticated credits only. Post text max 280 weighted characters; up to four owned JPEG/PNG/WEBP UUIDs (5 MiB each) or one MP4 UUID (512 MiB, 140s, H.264/yuv420p, AAC if audio). made_with_ai is a client choice, default true for compatibility. Editing is limited by X's live edit_controls; delete affects the edit chain. Self-service replies require X's mention/quote eligibility. No DMs/bookmarks. Examples: {"action":"get_account"}; {"action"
• remix_video — Shot-for-shot video reskin or recast ("remix") of a film you own. Detects cuts in an owned video, shares cast and setting references across shots, transforms the cast and world (mode="reskin") or the performers alone (mode="recast"), normalizes clips and verifies the delivered frame count, and restores the original soundtrack. It preserves edit boundaries and duration; fidelity of generated motion, mouths, hands and lettering is checked rather than guaranteed. It does not clone or replace voices. Import a direct video URL into your library with lib_manage(content_url=...) first. Two recipes, chosen with `recipe`: - `recipe="source_edit"` (default): edit each source shot with shared references, check it against source frames, and attempt targeted repairs within the quoted allowance. The best take is retained. Remaining shot defects are named in `needs_review`. The final synchronized source-left/result-right video is reviewed at 4 sampled frames/second in windows up to 10 seconds; its
• train_adapter — Train a LoRA adapter from library media. Each training_data item needs first_frame_resource_id + video_resource_id (kf2v also needs last_frame_resource_id). Backends: dashscope (wan* models) or modal (hunyuan/ltx/wan-selfhosted/HF repos).