AI video generation and processing server with intelligent model routing, multi-step pipelines, memory persistence, and asset management
vidmcp has 14 tools with explicit schemas and descriptions, but quality is inconsistent. Naming is mostly verb-driven (route_, generate_, create_, run_, get_, upload_, animate_, restyle_, process_) but some tools lack parameter descriptions and have vague schemas. The 'create_pipeline' tool has an underspecified 'steps' array with no item schema. Several tools (remember, recall, learn_from) lack parameter descriptions. Output schemas are not documented. Error handling is absent from visible code. Average description length ~120 chars (within baseline 194), but parameter annotations are sparse. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write semantics.
Animate a static image into a video using AI. Pass asset_id from upload_asset.
Creates a multi-step video workflow. Steps can be: generate, quality_check, audio, merge. Returns a pipeline_id to track progress.
Generate voiceover audio from text. Auto-detects ElevenLabs if API key present, otherwise uses free Edge TTS. Indian English voice by default.
Generates a video from a text prompt. Automatically uses the best available API — Kling, Fal.ai, or free fallback. Just pass a prompt and it handles the rest.
Check progress of a running pipeline.
Rate a generation 1-5 to improve future routing decisions.
create_pipeline 'steps' parameter has no item schema. Accepts array but does not specify what each step object must contain (type, params structure). LLM cannot validate step structure before calling.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear semantics. Tools like 'remember', 'learn_from', 'generate_video' are destructive/write operations but unmarked. Agents cannot distinguish safe reads from state-modifying calls.
Output schemas not documented. Tools return results but LLM does not know field names, types, or structure. E.g., route_generation returns 'model' and 'intent' but this is inferred from description text, not a formal schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Upload a raw video or audio recording. Automatically transcribes speech, detects and lists filler words like um, uh, like, you know. Returns transcript and asset ready for editing. Use this when user uploads any raw footage.
Scores a generated video clip from 0 to 5.
Retrieve relevant memories and similar past generations for the current context.
Store a user preference, brand detail, or style choice. Persists across all future sessions.
Apply AI style transfer to existing video footage. Strength 0-1.
Automatically picks the best AI video model for any prompt. Just pass the prompt — no need to mention budget or quality unless user specifically asks. Handles everything automatically.
Runs a created pipeline asynchronously. Handles all API waits in background. Use get_pipeline_status to check progress.
Upload a local image, video, or audio file for use in pipelines. Returns an asset_id.
run_pipeline requires 'routing_decision' parameter (object type) with no description or schema. LLM cannot infer what fields this object should contain.
No error handling guidance in visible code. Tools do not return actionable error messages (e.g., 'Invalid quality: must be draft|good|best'). Agents cannot self-correct on failures.