MCP server that lets Claude read and edit CapCut desktop draft projects.
capcut-mcp has 17 tools with explicit Zod schemas and descriptions. However, most parameter descriptions are null or missing (visible in schema definitions: 'description':null). Tool descriptions are present but brief (avg ~80 chars). No output schemas documented. Error handling wraps exceptions generically without recovery guidance. Naming is verb-noun compliant (add_*, set_*, etc.) but some tools like capcut_raw_patch expose low-level implementation details. The server lacks idempotency guarantees, confirmation patterns for destructive ops, and structured error messages.
Add an audio clip at a time on a track.
Add an image at a time on a track.
Add a text overlay. Requires a text template (a draft with a text layer; see CAPCUT_TEMPLATE_DRAFT).
Add a new track (video | audio | text | sticker).
Add a video clip at a time on a track. Session edit; call capcut_save to persist.
Copy a draft folder to a new name (valid scaffolding). empty:true clears all clips/tracks for a fresh build.
Remove a segment.
Parameter descriptions are null or missing across media placement tools (add_video, add_image, add_audio). Zod schema shows 'description':null for atSec, durSec, srcStartSec, trackIndex, trackRenderIndex, scale, posX, posY, rotation, opacity, volume, speed. LLMs cannot infer parameter meaning without explicit descriptions.
No output schemas documented. Tools return JSON via ok() wrapper but LLMs have no way to know what fields to expect (e.g., does capcut_read_timeline return canvas, fps, tracks, segments?). This forces agents to guess downstream field names.
Error handling is generic. All errors wrapped with 'ERROR: ' prefix and isError:true flag. No recovery guidance (e.g., 'file not found, check DRAFTS_DIR' or 'CapCut is open, use force:true to override'). Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Drop unsaved session edits and reload the draft from disk.
List CapCut drafts in ${DRAFTS_DIR} with duration and lock status.
Move a segment to a new start time and optionally another track.
Advanced escape hatch: deep-merge a JSON patch into draft_content (undocumented ops). Use with care.
Read a draft: canvas, fps, tracks and every segment (id, media, times, layer). Read-only.
Write session edits to disk (backs up .mcpbak, validates). Refuses if CapCut is open unless force:true.
Set transform / opacity / volume / speed / visibility on a segment.
Split a segment into two at a timeline time.
Change a segment start / duration / source in-point (seconds).
Check the (in-session) draft for overlaps, duplicate ids/render_index, missing media.
Destructive operations (capcut_delete_segment, capcut_save with force:true) lack confirmation or dry-run patterns. Agents can irreversibly delete segments without a safety gate. No idempotency guarantees documented.
capcut_raw_patch exposes low-level JSON patching with undocumented schema ('patch': z.record(z.any())). This is an escape hatch that bypasses validation and invites malformed edits. Naming and description do not convey risk clearly enough.