MCP server for content-aware landscape to vertical (9:16) video reframing — analyse, preview, then render, entirely locally.
Strong foundation with well-structured schemas and detailed parameter descriptions. All 8 tools have explicit descriptions (avg 180 chars) and comprehensive input schemas with type constraints. Tool names follow verb_noun pattern (analyze_shots, preview_shots, render_reel). However, output schemas are not documented in the visible code, and error handling guidance is minimal. The `doctor` tool lacks parameter descriptions (empty input object). Some parameter descriptions are verbose but lack actionable constraints (e.g., cut_thresh explanation is 200+ chars but buried in prose rather than structured constraints).
Detect scene cuts and compute a per-shot treatment plan from saliency plus on-screen text. Returns each shot's suggested mode, raw metrics, and warnings explaining where the suggestion is risky — plus a labelled contact sheet of shot mid-frames. LOOK AT THE CONTACT SHEET: the metrics cannot tell you that a crop cuts a headline in half. Results are cached, so re-planning the same clip is free.
Cancel a running render or preview job. Returns true if the job was cancelled, false if it was already finished.
Report which parts of the toolchain are present and what each missing piece would unlock. Call this first if anything behaves unexpectedly — the server runs in a degraded mode rather than failing loudly.
Poll a render or preview job for status and progress. Returns the current job state; if status is 'completed', output_path contains the result.
Render a few frames from each shot with a proposed treatment, so you can check a plan before committing to a full render. Returns a contact sheet of the results.
Output schemas not documented. Tools return complex objects (contact sheets, job state, frame data) but LLMs cannot see expected response structure. Forces agents to guess field names and types.
doctor tool has empty input schema ({}). No parameters documented, yet description suggests it reports toolchain state. Unclear what the tool accepts or returns.
Error handling lacks recovery guidance. No tool description explains what to do if a video is unsupported, a job fails, or a path is invalid. Agents cannot self-correct.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
Read a video's dimensions, frame rate, duration and audio presence, plus rough analysis/render time estimates. Cheap — call it before committing to anything slow.
Extract a single frame from a video at a given timestamp, as a JPEG. Useful for inspecting specific moments or building custom contact sheets.
Render the full reframed video with a shot plan. Returns a job handle immediately; poll get_job to wait for completion. This is the only slow call — use preview_shots to iterate on the plan first.
Parameter descriptions embed constraints in prose (e.g., cut_thresh explanation is 200+ chars). Should use JSON Schema constraints (minimum, maximum, enum) and keep descriptions concise (10 - 100 chars).
No confirmation or dry-run pattern for destructive operations. render_reel and preview_shots accept shot_plan overrides that could produce unwanted output, but no preview-before-commit safety mechanism.