| name | resolve-mcp |
| description | Orientation and index for DaVinci Resolve MCP work — grading, editing, conforming, delivery, media analysis, and .drp/.drt/.drx file work, live in a running Resolve or offline with none open. Load this for a map of the domain skills, the live-vs-offline servers, and the cross-cutting safety rules. The per-domain skills (resolve-color / resolve-edit / resolve-conform / resolve-delivery / resolve-media-analysis) carry the depth and self-trigger on their own descriptions; use this as the map, or when a task spans several domains. |
DaVinci Resolve MCP — Index
Orientation for any Resolve MCP task. This is the map, not the depth — each
domain skill below carries its own routing and triggers on its own description.
Open the one that matches, or use this when a task spans several domains. This
skill does not auto-load the others; it points at them.
Two servers — compute offline, apply live
- Live —
davinci-resolve (Python): drives a running Resolve via the
scripting API.
- Advanced / offline —
davinci-resolve-advanced (Node): authors
.drp/.drt/.drx files and patches the project DB with no Resolve open.
Rule of thumb: compute grades, QC, and conform math offline; apply the result
live. Node never drives Resolve.
Domains
| Task | Skill | Kernel | Any-MCP-client prompt |
|---|
| Grading, looks, shot match, LUT/CDL/DRX | resolve-color | color-grade-kernel.md | /color_grade_workflow |
| Cutting, trimming, ranges, variants, changelist | resolve-edit | timeline-edit-kernel.md | /timeline_edit_workflow |
| Conform, relink, finishing QC, grade tracing | resolve-conform | timeline-conform-interchange-kernel.md | /conform_workflow |
| Render, deliverable QC, media/provenance | resolve-delivery | render-deliver-kernel.md | /delivery_workflow |
| Fusion comps (titles, MG, VFX) | resolve-fusion | fusion-composition-kernel.md | /fusion_workflow |
| Audio / Fairlight (tracks, buses, loudness) | resolve-audio | audio-fairlight-kernel.md | /audio_workflow |
| Media pool ingest / organize / multicam | resolve-media-pool | media-pool-ingest-kernel.md | /media_pool_workflow |
| Reading/analyzing source media | resolve-media-analysis | media-analysis-guide.md | /analyze_media |
Less-common domains (no dedicated skill — go straight to the kernel/tool)
These have real coverage but low enough traffic that they route through this
index rather than their own skill:
- Project lifecycle — create/export/import/archive/restore projects,
databases, settings, presets:
project_manager compound tool →
docs/kernels/project-lifecycle-kernel.md. Offline DB read/patch: advanced
project_read / project_db (project CLOSED + quit/relaunch).
- Review / annotations — timeline markers, review reports, annotation
copy/move/scope:
timeline_markers → docs/kernels/review-annotation-kernel.md.
Offline audit/lineage: advanced provenance.
- Extension authoring — install/remove Fuse/DCTL/Lua-Python plugins:
script_plugin → docs/kernels/extension-authoring-kernel.md,
docs/authoring/.
- Pipeline (DB-as-truth) — YAML-authored canonical project DB, staged runs
with gates + provenance + drift: advanced
pipeline tool →
resolve-advanced/README.md.
Read the build before you promise a capability
The scripting API changes per PATCH release, not per major one. "Resolve 21"
is not a fine enough label to reason from: GetFairlightPresets exists on 20.2.2
and not on 19.1.3, and three surfaces reported in 21.0.4 are absent from 21.0.2.
Guidance given without knowing the build is how an agent ends up insisting a
method is there when it is not (issue #132).
So, before describing what is possible:
resolve_control get_version → note both version_string (the Resolve
build) and mcp.version (this server).resolve_control check_version_support → every recorded gate this build does
not clear. With a symbol, it answers for one method.resolve_control api_truth "<topic>" → each fact now carries a
version_context saying whether it was measured on an older, newer, or
identical build. An older measurement is a prior, not a finding.
Two failure modes worth naming, because neither announces itself:
unknown is the common answer, and it means probe — not yes. Only
surfaces with recorded evidence are gated; most of the API has never been
version-bisected. Absence from the unavailable list is not a promise that a
method exists.mcp.version is the version this process STARTED with. A running server
keeps executing it, so git pull does not refresh the ledger until restart —
an api_truth miss may mean "stale server", not "no known issues".
When a build is too old for what was asked, say so and name the floor. Do not
offer the call and let it fail.
Cross-cutting rules (always)
- Source media is sacred (AGENTS.md): never modify, transcode, convert, proxy,
relink, or derive source media unless explicitly asked. Outputs go to sidecars,
scratch, or the analysis project root.
- Frame-first color: inspect Resolve-rendered frames before applying any
grade/look/LUT/CDL/DRX, and preserve a recoverable grade version.
- Guards refuse rather than fabricate on the advanced server — read a
"refused" message before retrying (usually wrong value space, log-encoded
frames, missing media, or a missing optional dep; call the advanced
capabilities tool).
Deeper references
AGENTS.md — canonical brief + the ## Domain Routing index (all platforms).docs/SKILL.md — operating reference for both servers.docs/kernels/ — per-action depth. resolve-advanced/README.md — offline catalog.
