Repository navigation
Releases
1.6.0: replace_clip carries everything, import_folder works, registry-ready
Install: quit Premiere Pro, then npx -y premiere-mcp@latest setup, reopen Premiere and your AI app, and ask it to run verify_premiere_connection. Agents: see AGENTS.md.
- Claude Desktop: double-click
premiere-mcp-1.6.0.mcpb. - Teammates without a terminal (macOS):
Premiere-MCP-Installer-v1.6.0.zip→ double-clickInstall Premiere MCP.command.
Swapping a shot for a new version now carries the whole clip or changes nothing: trim, in point,
every effect setting by position, keyframes, linked audio laid out as it was (L/J-cuts, other
tracks), edge transitions, and sync sound kept and relinked under a picture-only render; anything
that cannot be reproduced is refused or undone, never reported as done. Exports find Media
Encoder on a standard Creative Cloud install, ffmpeg is found when the client was launched from
the Dock, DeNoise/DeReverb can no longer crash Premiere on nested clips, and the package carries
what the official MCP registry needs to list it.
-
Ready for the official MCP registry.
package.jsoncarriesmcpNameandserver.json
describes the package for registry.modelcontextprotocol.io; a test keeps the two versions in
step. -
replace_clipcarries everything, or changes nothing. It used to copy effect settings by
name throughcopy_effects_between_clips, and Lumetri repeats names (four "Saturation", two
each of Temperature, Tint and Contrast), so one setting overwrote another while the swap
reported success; keyframes were not copied; effects, levels and the layout of the clip's linked
audio or video were dropped, and so were the transitions on its edges. Now the clip and every
clip linked to it are read: track, start, end, source in point, on/off, effect order, every
setting by position (effect, then property index) with keyframes, and the transitions on their
edges. The whole layout is rebuilt first on a temporary copy past the end of the edit, so L- and
J-cuts and audio on another track than its video come back as they were, and compared with the
original. The old clip is touched only when that rehearsal matches exactly, the new media is long
enough (from its real length, read at its real frame rate with drop-frame counted: Premiere lays
a clip in past the end of short media and reads it back as asked), and every span the rebuild
writes is free apart from the old clip's own group. After the swap everything is compared again,
and edge transitions are re-added and read back; if anything differs the swap is undone back to
exactly the state before it. Not carried, and said so: keyframe interpolation (no scripting
read). Refused up front: a transition centred on a cut (QE places them only starting or ending at
the cut), settings that cannot be read on the old clip, and media whose length Premiere does not
report. New live workflow W11 covers it. -
replace_clipcloses the gaps a review of the rebuild found. Each of these could report the
wrong thing:- A swap for other media with the same name and span (a re-render under the same file name)
that then had to be undone stopped at 0 undo steps and said nothing changed, with the new
media still in the edit. The timeline state behindrevert_last_callandwhat_changednow
covers each clip's source item, and a revert that takes 0 steps after a swap is reported as
not undone. - A linked clip from other media (dual-system sound synced to the picture) was rebuilt from the
new media. It now stays exactly where it is and is linked to the new picture. The same goes
for the sync sound under a picture-only VFX render, which used to be refused every time ("the
new media has no audio"); the result lists what was kept and why. New media whose own audio
would land on a kept clip is refused before anything is touched. - New media with more audio channel clips than the old clip used (stereo old, dual mono new)
dropped the extra channel and reported verified success. It is now refused. Dual-mono audio is
laid in from the picture, so clicking one of its channels swaps like clicking the video. - Between the rehearsal and the swap, only the old clip's spans were checked again. A clip
dropped into the slot in between could be overwritten while the swap reported success, and
changed settings were carried from the stale copy. The slot and the settings are now checked
again too. - The source in point of each piece was never compared, and spans were held to 0.02 s, more
than a frame at 50p and 60p. Both are now compared to half a frame of the sequence. - A throw part-way through the rehearsal left temporary pieces past the end of the edit while
reporting nothing changed, and could leave the new item's bin in/out marks moved. Pieces are
now recorded as they land and removed, and the marks are put back however a placement ends,
as tick strings (a seconds number floors to the frame below), and read back. - Dual-mono (or more mono channels) audio that does not sit on the tracks Premiere lays its
channels on could come back with the channels swapped. It is now refused with the tracks
named; audio laid out as Premiere lays it still swaps. - Also: every linked piece is checked for speed changes, not just the clicked one; the group is
checked again right before the swap, so a trim made in between stops it untouched; the
rehearsal is shifted by whole frames (a whole second is not a frame at 23.976 or 29.97); a
spare piece that cannot be removed is reported; one clip refusing selection no longer drops
the rest of the user's selection; and the result says that re-added transitions are back to
their default settings (Dip to Color's colour, Push's direction cannot be read by script). batch_replace_clipsno longer lists a second video clip linked to one already listed.- New unit suite
replace-clip-swap.test.tsruns the generated ExtendScript against a fake of
the measured Premiere behaviour; each case fails on the previous code.
- A swap for other media with the same name and span (a re-render under the same file name)
-
batch_replace_clipskeeps each clip's trim and effects. It removed every clip that used
the old item and laid the new one in at full length, losing trims, grades and levels and
running over the next shot, and reported the count as success. It now runs a verified
replace_clipper use (video, or audio-only), and a clip whose swap is refused is left exactly
as it was and listed. -
Exports find Media Encoder where Creative Cloud installs it on macOS. The install check only
looked for an .app directly in /Applications, and Creative Cloud puts it one folder down
(/Applications/Adobe Media Encoder 2026/Adobe Media Encoder 2026.app), soexport_sequence
refused with MEDIA_ENCODER_NOT_INSTALLED on a standard install. It now looks inside each
Adobe Media Encoder <year>folder too (ported from upstream 5528abb; the presets half was
ported on 10-21). -
ffmpeg and ffprobe are found outside PATH. An MCP client started from the Dock or Finder
(Claude Desktop, an IDE) gets launchd's minimal PATH, so a Homebrew ffmpeg was reported as not
installed. AfterPREMIERE_MCP_FFMPEG_CMD/PREMIERE_MCP_FFPROBE_CMDand PATH, the server now
tries /opt/homebrew/bin, /usr/local/bin and /opt/local/bin.detect_silenceand the ffmpeg
check behind it ran a bareffmpegand ignored the override; they use the same lookup now. -
DeNoise and DeReverb no longer crash Premiere on nested or multicam clips. On 26.5.2,
adding either to a clip whose source is a sequence crashes Premiere. The 10-29 live sweep hit
it withdenoise_clipon a nested clip, and it reproduced there; multicam clips are sequences
too and are refused the same way as a precaution. Every tool that adds them now checks first
and leaves those clips alone with the reason:denoise_clip,dereverb_clip,
apply_audio_effect_qe(on either track kind),apply_effect,apply_audio_effect,
apply_audio_effect_to_all_clipsandbatch_apply_effect.replace_cliprefuses before it
touches anything when the old clip carries one of them and the new media is a sequence. The
check matches the name in any case and the obsolete "Adaptive Noise Reduction", and refuses
when it cannot read the clip's source. It is not tied to a version: it stays on until a build is
measured safe. -
Batch tools report what they skipped.
apply_audio_effect_to_all_clipsreported
allOk: truewhatever happened (it usedArray.every, which this ExtendScript engine lacks),
andbatch_apply_effectreturned plain success. Both now returnsuccess: falsewith
partial,appliedCount,skippedCountandfailedCountwhen any clip was left out, and
check that each effect was actually added. -
replace_clipandstem_exportwork under the default profile. Their internal steps
(copy_effects_between_clips,export_audio_only) are not incore, so they were refused:
replace_clipreported success with effect values at their defaults, and every stem failed.
Internal steps now run past the profile, as transaction steps do.replace_clipstops before
touching the old clip if any of its effects could not be added to the new media or the values
could not be copied, removes its temporary copy (with all its linked audio), and fails if the
final copy did not land. -
freeze_frameon an image sequence (EXR/DPX/PNG plates) holds the frame at the requested
time, picked from the sequence's own files, instead of always the first frame. Whether media is
a sequence comes from Premiere's own media type ("Video" for numbered stills, "Still Image" for
a single still), not from numbered files beside it, so one photo from a folder of numbered
photos still holds itself. Animated GIFs seek as video does. -
set_qe_component_paramreads the value back. A value that went tonan, or that
...
1.5.0: one-command setup, safe retries, undo a whole call
Install: quit Premiere Pro, then npx -y premiere-mcp@latest setup, reopen Premiere and your AI app, and ask it to run verify_premiere_connection. Agents: see AGENTS.md.
- Claude Desktop: double-click
premiere-mcp-1.5.0.mcpb. - Teammates without a terminal (macOS):
Premiere-MCP-Installer-v1.5.0.zip→ double-clickInstall Premiere MCP.command.
npm:
premiere-mcp@1.5.0is published and in npm's review for new packages; until it clears,npx premiere-mcpresolves to a placeholder. The.mcpband installer zip below work now.
Premiere MCP that teammates can install with one double-click, that works with no panel open,
and that is safe to drive under timeouts: retries do not apply an edit twice, a whole call can be
undone, hand edits in Premiere are noticed, and replies are sized to fit. New editor workflows
(version up every shot, remove silence, sync by audio, stringout from a transcript) each come
with a live workflow test; all nine pass on Premiere Pro 26.5.2.
Setup in one command
npx -y premiere-mcp setupinstalls the signed Premiere plugins through Adobe's own plugin
installer and adds the server to Claude Desktop, Claude Code, Codex, Cursor and VS Code
(--clientsto choose,--dry-runto preview; each edited config is backed up).doctor
checks Node, the installed bridge, Premiere's heartbeat and each app. The old hand-copy route
(CEP developer mode) is nowsetup --unsigned.- The server now uses the bridge folder the background bridge watches when
PREMIERE_TEMP_DIR
is not set. It used to make a private folder per session, so a server started without the
variable never reached Premiere.PREMIERE_MCP_SESSION_TEMP=1keeps the old behaviour. - The Claude Desktop bundle installs the signed bridge on first launch and leaves an existing
signed install alone (it used to hand-copy the panel and turn on developer mode every launch). - Packages shrink from 56 MB to 1 MB (npm) and 57 MB to 2.6 MB (
.mcpb): design screenshots and
compiled tests are no longer shipped. AGENTS.md/CLAUDE.md: steps for an AI agent setting this up; README and QUICKSTART
rewritten aroundsetup.- Tools that write without a way to read back say so:
set_parametric_eq,set_compressor
andset_limiterreturnverified: false, and fail when no parameter was accepted. stabilize_clipwrote Smoothness into the wrong effect: Premiere inserts Warp Stabilizer ahead
of Lumetri, and the tool took the last effect on the clip. It now finds the stabilizer by its
internal name and reads Smoothness back. The Lumetri tools find their effect the same way.verify_color_gradeandverify_effect_renderedaccepted asequenceIdand used the active
sequence; a source audit that should have caught it only read the first line of a signature.- Code formatted throughout (one formatting-only commit, listed in
.git-blame-ignore-revs), and
formatting is now checked in CI. Stale comments fixed; out-of-date README translations removed.
Added: safe retries and editor workflows
- "Did my last call run?" Every call is journaled;
get_recent_callsshows what ran,
including calls whose reply was lost to a client timeout. ArequestIdon any call makes
a retry return the first result instead of applying the edit twice. - Background jobs and progress.
start_job/get_job/list_jobsrun slow tools
(exports, renders, transcription) without outliving the client's timeout; clients that
send a progress token get a heartbeat every 5 s. - Ask before destroying. Deleting a bin or item that is used on a timeline, and raw
scripts that delete or quit, now returnneedsConfirmationwith where it is used. Clients
that support elicitation get a yes/no question; on yes the call runs withconfirm. where_used: every sequence and time a project item appears.version_up_shots: relinks every versioned shot (SH010_comp_v003and so on) to the
newest render on disk, keeping trims and effects. Dry run first.remove_silence: cuts the pauses out of a talking clip and closes the gaps on every
track, so music and other audio stay in sync. Asks first; takes a recovery checkpoint.sync_by_audio: finds the offset between recordings of the same moment by waveform
(8 kHz FFT cross-correlation, sub-frame) and can build the synced multi-track sequence.look_at_cut: one image strip of the cut, one labelled frame per shot.stringout_from_transcript: paper edit from a Whisper.jsonor.srt; give the lines
in cut order and get a sequence of just those moments. Unfound lines are listed, not guessed.- Switch task menu mid-session.
activate_tools({ bundle: "task-grade" })(or just
"grade" / "grading") loads a whole task menu with real schemas;deactivate_tools({ bundle })
drops it. Past ~50 advertised tools the reply warns and says what to drop. - Undo a whole call.
revert_last_calltakes back the last edit made through the server
however many Premiere steps it took: it undoes until the sequence matches the state recorded
before the call, and if it never does, redoes back to where it started (by state, not by
count). It refuses when the sequence changed after that call. Script-added markers are not on
Premiere's undo stack, so marker-only edits are refused cleanly. - Hand edits are noticed. The next edit after a change made by hand in Premiere says so
(changedOutsideThisServer), andwhat_changedlists clips added, removed, moved, trimmed
or renamed since the server last looked. - Replies sized to fit. Every result is capped (40,000 characters by default;
maxResultChars
per call orPREMIERE_MCP_MAX_RESULT_CHARS) by halving its longest lists, and says what was
cut.list_project_itemspages (200 by default) and filters by name, type and bin. Results
are sent as compact JSON. - Project notes.
add_project_note/get_project_noteskeep conventions (naming, where
renders land, delivery) in a Markdown file next to the.prproj;get_project_info
includes them, so each session starts knowing them. - Declared result shapes (
outputSchema) for the 15 most-called tools. All fields are
optional and extra fields are allowed, so failures still fit; W9 checks live results. scripts/workflow-tests.mjsW5–W9 run these end to end against live Premiere.
The timeout-safety tools are in core and named in the server instructions; under a task
bundle they are reached through call_tool, which keeps every bundle at or under 50 tools.
Live sweep against Premiere Pro 26.5.2: 496 tools pass and 61 fail, up from 449 and 111.
New fixtures (a real multicam, a caption track, an AE-authored MOGRT) made 54 more tools
genuinely testable, and running them found the bugs below. Every effect below was checked
against a render, not only an API readback.
Added
create_multicam_cutbuilds a finished multicam cut: a new sequence of real multicam
clips, each on the angle you choose, still switchable in Premiere. ExtendScript can
neither create a multicam source nor switch angles; an FCP7 XML<multiclip>import can.freeze_framenow works (26.5.2 has no frame-hold call). It extracts the source frame
with ffmpeg and overwrites it as a still fordurationSecs(new, default 2) on the
clip's own track. The held frame is pixel-identical to the source frame.set_clip_pansets pan through the QE clip pan component (value passed as a string).
Hard left measured -91 dB on the right channel.set_clip_gainandnormalize_clip_peaksapply gain through an Amplify effect, since
clip gain itself is read-only from script. +6 dB measured as +6.0 dB.apply_lut/apply_lut_to_clipselect the Lumetri Input LUT dropdown entry whose
render matches ffmpeg'slut3dof the ungraded frame (the slot takes an index; a path
raised "Illegal Parameter type"). An unlisted LUT is refused;installIfMissingcopies
it into the user LUT folder.uxp-probe/andnpm run uxp:install: a minimal UXP panel, installed as a.ccx,
that runspremiereproUXP calls over a private file queue. Groundwork for the calls
only UXP has (work area, sequence in/out, transitions); not yet wired into tools.
Workflow tests, and what they found
npm run test:workflows(scripts/workflow-tests.mjs) drives five real jobs as chains of
tools against the scratch project and checks them by what Premiere renders and holds, not
by tool reports: per-clip version swap, version-up by relink, a review reel of latest
versions with burn-ins, captions from an SRT, and an audio pass (ducking, stems, loudness).
All five pass, 45 of 45 steps.replace_clipkept nothing. It tookpreserveEffectsand ignored it: the trim, grade
and reframe were lost and the full-length new version ran over the next shot. It now lays
the new media into the old slot with the old source range and carries every effect across.
It never slides a clip past other clips: an intermediate version that did left Premiere's
track order scrambled, and everything after the moved clip rendered black.replace_clip_media(relink, the usual version-up) verifies the new path and renames the
bin item and the timeline clips that still carried the old file name, so the timeline no
longer reads "v001" while playing v002.import_media/import_foldercreate a named destination bin (ora/b/cpath) that
does not exist yet, instead of refusing. Ambiguous names and bad ids still refuse.- ExtendScript-only syntax errors are now caught in tests. An unescaped
/inside a regex
character class, and an ES3 reserved word (short) as a variable name, each broke a whole
generated script in Premiere while the acorn-based syntax test stayed green. The test now
rejects both, with self-tests proving it does.
Installing for a teammat...
v1.4.0
Premiere Pro 26.5.2 changed scripting behaviour under the server, and many tools broke silently. This release fixes them, each checked frame by frame against a clip whose frames encode their own number, read back from the rendered program output. Live coverage went from 34.9% to 86.4% of tools executed against a real Premiere host.
Security
- ffmpeg/ffprobe no longer run through a shell. About thirty tools interpolated caller-supplied paths into shell commands, so a crafted
mediaPathcould execute commands. All now pass argument arrays.
Fixed
- Frame capture:
capture_program_output,capture_frame,export_frameandexport_thumbnailreturned frame 0 whatever time was asked for. They are now frame-exact. - Editing: ripple trim, extend edit, J/L cuts, trim, nudge, roll, slide, slip, ripple delete, extract and split are rebuilt on measured edit semantics. Linked audio (including dual-mono pairs) stays in sync, media handles are checked, and a ripple that would collide with another track is refused.
- Targeting: QE tools hit the wrong sequence (QE orders sequences differently) or the wrong clip (QE counts gaps as items). 36 tools now honour
sequenceId, and a missing id is an error instead of a silent retarget. - Transitions and crossfades now land at the cut at the requested length.
- Export dialogs: AAF and FCP XML report dialogs no longer freeze the bridge. On macOS they are dismissed by name.
- Engine: restored the array methods missing from the 26.5.2 script engine, which about 20 tools needed.
- Windows: tools that defaulted their output to a POSIX
/tmppath now use the platform's temp directory. Windows CI passes. - Reverse, colour space, markers, gap tools, start timecode, OMF/AAF settings exports,
duplicate_clip,add_to_timelineaudio clean-up, andcopy_markers_between_sequences(it could hang Premiere) are fixed too. Several tools that reported success without doing the work now either do it or say they can't.
Added
npm run canary: re-measures the 17 scripting behaviours the tools depend on, in about 15 seconds. Run it after every Premiere update.set_scale_to_frame_size_batch.
Known limits on 26.5.2
Clip gain, effect enable/disable, custom LUT files, the work area, and QE speed changes cannot be scripted reliably. The tools now say so. Speed tools undo a wrong result and point to the frame-exact bake_speed_change.
Full details are in CHANGELOG.md. The CEP panel is unchanged from 1.3.1 (bundle 1.1.6).
v1.3.1
Fixes from a full review of the 1.3.0 work, plus two fixes ported from upstream. Replaces v1.3.0.
Fixed
move_clip_to_trackno longer overwrites or loses clips. It refuses an occupied destination unlessoverwrite:true, places and trims a copy before clearing anything, and removes the source last. A native move is read back, not trusted. The fallback re-creates the clip, so it now requiresallowLossyMove:true.- Encoder presets are found.
get_encoder_presetsalso scans~/Documents/Adobe/Adobe Media Encoder/<ver>/Presets, where Media Encoder saves user presets. project_file_editcan't write a broken project. Replacements apply in one pass; markup changes are refused unless allowed; the output never clobbers an existing file;wroteis accurate on every failure path;.prprojpaths only.- MCP
initializereports the right version (it said 1.2.0).
Live harnesses
- Mutating sweeps no longer touch Premiere's global Auto Save settings, the clipboard, or open new projects.
- Sweeps re-check the scratch project as they go; recovery never saves or quits anything else, and never presses Escape or quits Premiere when another client is using it or when health is unknown.
Install
Not published to npm. Install from source:
git clone https://github.com/skrfilms/premiere-mcp.git
cd premiere-mcp
npm install && npm run build && npm run setup:mac
Full detail in CHANGELOG.md.