skillZs
★ LIVE SKILL TAGS ★
>>> LIVE SKILLS INDEX <<<
* OPEN SOURCE *
NO LOGIN, NO TRACKING
※ REAL INSTALL DATA ※
← back to all skills
samuelgursky/davinci-resolve-mcp61 installs

resolve-edit

Editing, cutting, trimming, pacing, and timeline restructuring in the DaVinci Resolve MCP. Apply when duplicating/moving clips, copying ranges, building variants, tightening or restructuring a cut, editing selects, or generating an editorial changelist/turnover — live in a running Resolve OR offline against .drt timeline files. Routes to the live edit tools, the offline editorial tools, and the project's editorial craft guidance.

How do I install this agent skill?

npx skills add https://github.com/samuelgursky/davinci-resolve-mcp --skill resolve-edit
view source ↗

Is this agent skill safe to install?

  • Gen Agent Trust Hubpass

    The skill defines a safe interface for DaVinci Resolve editorial tasks, bridging craft guidance with technical tool kernels. It emphasizes safety through API version-gating and enforces human-in-the-loop review for transcript-based automated editing.

  • Socketpass

    No alerts

  • Snykpass

    Risk: LOW · No issues

What does this agent skill do?

Resolve Timeline Edit

Bridges editorial craft to this repo's tools. Open the right manual; don't re-derive it here.

  • Craft / story — docs/guides/editorial-decision-guide.md. The global editor / assistant-editor skills add editorial philosophy and cutting-room practice; use them for why to cut, not tool mechanics.
  • Live tool mechanics — docs/kernels/timeline-edit-kernel.md (the timeline edit-kernel boundary: duplicate/copy/move, range ops, item-state copy).
  • Offline timeline / changelist — resolve-advanced/README.md → drt (timeline file authoring) and editorial (interchange + turnover).

Two servers

JobServerTools
Restructure a running timelinedavinci-resolve (Python, live)timeline (edit kernel), timeline_item, edit_engine, timeline_markers
Author/diff a .drt file, or parse/compare editorial interchange with no Resolve opendavinci-resolve-advanced (Node)drt, editorial

What this build cannot do (check before you offer it)

The scripting API changes per patch release, so "Resolve 21" is not a usable label. Read resolve_control get_version → build.unavailable_on_this_build before offering a gated surface; check_version_support asks about one named symbol. Gated in this domain:

SurfaceNeedsIf absent
Timeline.GetSelectedClips21.0.4No selection readback. Identify clips by track/index or id instead — this repo's selection helper duck-probes three names, so it degrades to "no selection" rather than erroring
TimelineItem.SetName20.2Clip names are read-only from a script

An empty unavailable_on_this_build means nothing recorded is missing, not that everything exists — most of the API has never been version-bisected. A symbol with no gate returns unknown, which means probe it. Probe with name in dir(obj), never bare hasattr: on a Resolve object hasattr returns True for every name, real or invented, so it can only say yes.

Distinguish gated from absent. A gated surface arrives with a newer build; an absent one never does, and upgrading will not help. Clip speed / retime is absent on every build — see the edit-kernel essentials below before you offer set_retime for it.

Live edit-kernel essentials

  • Duplicate/relocate: duplicate_clips (modes same_time/offset/ at_playhead/track_above/after_source/next_gap), copy_clips (alias), move_clips (duplicate-then-delete). include_linked=True carries linked audio.
  • Ranges: copy_range, duplicate_range, overwrite_range, lift_range. No public razor/split — partial overlaps in lift_range are blocked unless allow_partial_item_delete=True (whole-item delete, not a trim).
  • Item state copy: copy_properties (transform/crop/composite/audio/retime/ markers/flags/grades/takes/keyframes …); scope with a group list.
  • Reading the user's selection is best-effort and version-dependent. The selection helper duck-probes three method names; the documented one, Timeline.GetSelectedClips, is 21.0.4+ (issue #131), so on older builds selection resolves by luck or not at all. Never build a destructive operation on "the selected clips" without reading back what you actually got — an empty or partial selection is indistinguishable from a small one.
  • There is no clip-speed API at any version. set_retime sets retime quality only (RetimeProcess, MotionEstimation) and returns True for doing so, which reads as if the retime succeeded. Setting a % speed, reversing a clip, and speed ramps are all unreachable — SetProperty('Speed'|'PlaybackSpeed'| 'RetimeSpeed'|'ClipSpeed') returns False and the matching GetProperty returns None on every build measured. Say so and route the user to the UI; do not offer set_retime as if it answered the question (issue #132).
  • Routing to the UI is where issue #132 actually went wrong — and it was not a tool lie. Having been told to hand off, the assistant improvised the handoff: it sent the user hunting for a dropdown "in the lower left of the clip" and never mentioned the keyframe tray. That direction came from nowhere in this repo. This repo verifies API behavior, not UI geography, and has no mechanism to version-guard a UI claim — Resolve's controls move between builds, pages and layouts, and every ledger entry here is stamped with the build it was measured on precisely because unstamped claims rot. So: never improvise UI geography. Say what the user needs to do ("set the clip's speed / build the speed ramp in Resolve's own retime controls") and say plainly that you cannot see their screen. Do not invent where a control sits, what it looks like, which corner of a clip to click, or a menu path or keyboard shortcut recalled from general knowledge — an invented location costs the user more than "I don't know where it is on your build" ever does, and it reads as authoritative because everything else you told them was measured. The line is improvised vs. written down here. A UI pointer that already exists in these skills (e.g. the playback-frame-rate path in resolve-rough-cut) was authored and reviewed deliberately — use it verbatim. If no pointer exists, that is not an invitation to supply one: name the operation and point at Blackmagic's manual for their build. The same rule covers every other "do it in the UI" handoff in these skills.
  • edit_engine drives higher-level selects/tighten/swap flows (plan → confirm → execute); tighten variants can carry audio via keep_ranges mirror / include_audio. For a full dead-air pass over one long single-take recording — transcribe → plan → review → execute, with the source/record coordinate trap spelled out — use the resolve-tighten-recording skill.

Show the gaps before cutting them (edit_engine)

  • plan_dead_space_markers(timeline_name?, tightness?) — the review gate. Same calibrated detection as plan_silence_ripple, but it proposes markers instead of assembling a variant, so a human sees every gap before agreeing to lose it. Use this whenever the ask is "mark the dead space so I can review and approve" — that is a request for this verb, not for a tighten. Red = confident, Yellow = the gate only just cleared its separation floor. Nothing is written; pair the marker specs with timeline_markers.
  • tightness on the dead-space planner: generous (default) | balanced | tight. Default is deliberately the loosest. A first assembly is supposed to be long — trimming a generous cut is fast and visible, while recovering material the machine already discarded is slow and invisible. Move to tight only when the editor asks for it.
  • Guard bands are floored regardless of tightness: a removed gap always leaves frames before it (the outgoing word's decay) and after it (the incoming word's attack). tight removes more gaps, never more speech.

Word-level editing and spoken search (edit_engine)

Silence-driven tightening cannot remove an audible "um" mid-phrase or a stammered restart. These work on the words instead and are complementary to plan_silence_ripple, not a replacement.

  • plan_transcript_tighten(clip_ref) — fillers, false starts and long pauses, at word boundaries. Emits the same keep_ranges shape the variant assembler already takes, and every removal states its reason, so the plan can be argued with rather than only accepted. False starts are a flagged heuristic: tight repeats only, so "no, no, no" as emphasis survives.
  • search_spoken_content(query, mode) — phrase / all_words / regex across every transcribed clip, returning timestamped hits with context plus selects (in/out with handles). A lexical axis; find_similar is the semantic-visual one. Use both — a shoot should be searchable either way. Deterministic order (clip name, then time), so an identical search gives an identical selects list.
  • Needs transcript_words: run transcription, then strata backfill. Both actions say so rather than returning an empty plan.

Offline editorial (editorial actions)

  • parse_interchange — EDL / OTIO / XMEML (AAF = an honest refuse, not a fake).
  • turnover_changelist — moved / retimed / trimmed / replaced / new / gone between two cuts PLUS the junction diff (transitions added / dropped / changed, fade in/out, span deltas — carrier lines and BL fade legs never read as sources), with timing silent-lie guards (it flags what it cannot verify).
  • conform_manifest, marker_roundtrip.

Use these to answer "what changed between v3 and v4" or to hand a conform an accurate change list without opening either timeline in Resolve. For carrying a conform across a re-edit, see the resolve-conform skill.

Source-media safety (AGENTS.md)

Edit operations reference existing Media Pool items — they never transcode, render, proxy, or create derivatives of source media. Keep it that way. Treat generated probe reports as local scratch artifacts, not committed files.

Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.

<a href="https://skillzs.dev/skills/samuelgursky/davinci-resolve-mcp/resolve-edit">View resolve-edit on skillZs</a>