Unflick
unflick
A free, ad-free video player — for humans and AI.
unflick.app · MIT · Cross-platform · CLI + MCP first
unflick plays your video files and your YouTube / Bilibili / Twitch / Vimeo / 26 other streaming-site URLs, in a beautiful native window. Auto-skips YouTube sponsor segments via SponsorBlock. Finds subtitles on OpenSubtitles, or generates them locally with whisper.cpp when none exist.
What makes it different: MCP is a built-in control surface, not a third-party wrapper. Other players can be scripted from the outside; in unflick the CLI, the MCP server and the window share one player instance. Ask Claude to skip to the next chapter and the video you are watching skips — same mpv, same playback, no second hidden process. Every feature ships headless before it gets a button.
Install
One-line install:
# macOS — Homebrew
brew install --cask zhitongblog/tap/unflick
# macOS / Linux
curl -fsSL https://unflick.app/install.sh | bash
# Windows (PowerShell)
irm https://unflick.app/install.ps1 | iexOr pick an installer for your platform from the latest release:
Platform | File | Size |
Windows 10/11 (x64) |
| ~80 MB |
Windows 10/11 (x64), AI edition with bundled Whisper |
| ~150 MB |
macOS 11+ (Apple Silicon + Intel, signed + notarized) |
| ~15 MB |
Linux x86_64 (Debian / Ubuntu) |
| ~10 MB |
Linux x86_64 (Fedora / RHEL / openSUSE) |
| ~10 MB |
Linux x86_64 (distro-agnostic) |
| ~80 MB |
Related MCP server: ffmpeg-mcp-video-editor
Three interfaces, one core
unflick exposes the same playback engine through three surfaces:
GUI — modern player window with libmpv-quality playback, keyboard shortcuts, drag-and-drop, picture-in-picture, true fullscreen, chapters, bookmarks, A-B loop, frame stepping, subtitle timing and styling.
CLI — every feature is also a command.
unflick play <file-or-url>,unflick chapter next,unflick bookmark add --name "the good bit",unflick loop a,unflick subtitle auto,unflick subtitle bilingual on,unflick audio eq preset speech,unflick clip 0 5,unflick library scan. Output is JSON; pipe it tojqand automate. Started with--allow-dev,unflick dev snapshot/click/evaldrive and read the player's own window, which is how unflick is tested on all three platforms.MCP server —
unflick --mcpstarts a Model Context Protocol server over stdio. Add it to Claude Desktop / Cursor / Codex CLI's MCP config and your AI agent gets 101 tools (play, seek, chapter_seek, bookmark_goto, ab_loop, subtitle_delay, subtitle_bilingual, screenshot, clip, sponsor_segments, get_subtitles, equalizer_preset, generate_subtitles, library_search, …) plus live resources.
Claude Code users can install the skills and the MCP server in one step:
/plugin marketplace add zhitongblog/unflick
/plugin install unflickThat brings seven skills (play, navigate by what was said, clip/GIF, subtitles, library, screenshot, SponsorBlock) plus the MCP server — see marketplace/.
All three drive the same player. When the window is open it hosts the control port, so unflick pause from a terminal — or an agent calling pause over MCP — pauses the video on screen rather than some invisible second instance. With no window running, the CLI and MCP fall back to a headless daemon and everything still works.
What an agent can do that a wrapper can't
Because MCP talks to the player rather than at it, an agent gets tools that need the player's own state:
search_transcript/seek_to_text— "skip to where she explains the refund policy". Searches whatever subtitles are open: a loaded file, a sidecar.srt, an embedded track, or the ones you just generated with whisper.generate_chapters/set_chapters— give a file that shipped without chapters a real chapter list. It's not just data back: the chapters appear on the progress bar and respond tochapter_seekand PgUp/PgDn.describe_frame— hand the model the frame that's on screen right now, as an image. Same player, so it sees what you see.
Free, ad-free, local — by charter
unflick will never:
carry banner / pre-roll / "sponsored" ads
collect telemetry of any kind
require an account, login, or cloud connection
gate features behind a paid tier or trial countdown
bundle adware, browser hijackers, or "recommended" software at install time
Everything that matters runs on your machine. Library scans, subtitle transcription, AI rewriting — all local. Your files stay on your filesystem.
The SponsorBlock auto-skip for YouTube is the active form of "ad-free" — actively skipping ads in the content unflick plays, not just refusing to show its own.
License
unflick is released under the MIT License — see LICENSE.
unflick bundles, dynamically loads, or invokes the following third-party software at runtime. Each upstream license is enumerated in THIRD-PARTY-LICENSES.md, and the canonical license text for the LGPL/GPL/MPL-licensed components ships in the licenses/ directory of every installer:
libmpv — LGPL-2.1-or-later — dynamically loaded at runtime (source)
ffmpeg (Windows bundle) — GPL-3.0-or-later — invoked as a subprocess only (source)
yt-dlp — Unlicense (public domain) — invoked as a subprocess (source)
whisper.cpp (AI edition) — MIT — invoked as a subprocess (source)
Tauri + wry + mpv-player/mpv + glutin + objc2 + 590+ Rust crates — all MIT, Apache-2.0, ISC, BSD, MPL-2.0, or Unicode-3.0 (full breakdown)
React 18, Zustand, Framer Motion, Tailwind, Vite — MIT
See THIRD-PARTY-LICENSES.md for the complete attribution and compliance notes (subprocess invocation vs. dynamic linking, GPL aggregation, MPL file-level scope, etc.).
Build from source
git clone https://github.com/zhitongblog/unflick.git
cd unflick
pnpm install
pnpm tauri build
# Windows: ./scripts/fetch-windows-deps.sh fetches libmpv / ffmpeg / yt-dlp / whisper
# bash build-both.sh builds both Standard and AI editionsRequires Rust stable + Node 22 + pnpm 10 + the Tauri 2 platform prerequisites (WebView2 on Windows, WKWebView on macOS, webkit2gtk-4.1 on Linux). Linux loads the system libmpv, so apt install libmpv-dev first. macOS ships its own universal libmpv (it is the only way to play DVDs there — Homebrew's mpv is built without libdvdnav): ./scripts/build-mac-libmpv.sh builds it once, and brew install mpv still works as a fallback for everything but discs.
Releases are cut with ./scripts/release.sh <version> — see scripts/README.md for how the three platforms are built and why macOS is the one that is not built in CI.
Acknowledgments
unflick stands on the shoulders of mpv, ffmpeg, yt-dlp, whisper.cpp, Tauri, and a long list of MIT/Apache-licensed Rust and JavaScript libraries. None of this would exist without those projects.
The website at unflick.app and a sibling project (SoloMD) — also free, also MIT, also bundling MCP — are by the same author. If you like the design, that's where it comes from.
Available Tools
114 toolsab_loopA
Control the A-B loop, which repeats a section of the current file. Set point A and point B to start looping; clear to stop. Omit position to use the current playback position.
Inputs: action (required), position.
audio_delayA
Get or set the audio delay in seconds, for fixing lip-sync. Positive values play audio later. Call with no arguments to read the current delay.
Inputs: seconds, relative.
audio_listB
List all audio tracks (embedded)
Inputs: no named parameters.
audio_selectB
Select an audio track by ID (0 to disable audio)
Inputs: id (required).
bookmark_addB
Save a named position in a file, so it can be jumped back to later. Defaults to where playback is right now in the file the user is watching. Bookmarks persist across sessions and are recorded even in incognito mode, since they are asked for explicitly.
Inputs: file, name, position.
bookmark_clearA
Delete every bookmark for the file being watched, or for the file named. Pass all=true to delete them for every file — that scope is never reached by default.
Inputs: all, key, file.
bookmark_gotoA
Jump to a bookmark by id. Seeks if its file is already playing; otherwise opens that file at the bookmarked position, saving a resume point for the outgoing one. A bookmark on a disc is refused, without loading anything, when that disc is not the one in the drive.
Inputs: id (required).
bookmark_listA
Bookmarks for the file being watched, in timeline order. Pass all=true for every file, or file to ask about a specific one. Each entry carries the id that bookmark_goto takes, and the key it is filed under — the path for a file, the disc's own identity for a disc.
Inputs: all, key, file.
bookmark_removeB
Delete one bookmark by id.
Inputs: id (required).
bookmark_renameB
Give a bookmark a name, or drop the one it has by omitting name.
Inputs: id (required), name.
castA
Send what is playing to a television on the network over DLNA. "list" finds renderers; "to" starts a cast (naming a renderer, or omitting it when there is only one) and pauses playback here so it is not heard twice; "stop", "pause", "resume", "seek" and "status" drive it once it is running. Only local files can be cast — the television fetches them from this machine.
Inputs: file, action, seconds, renderer.
chapter_listA
List the chapters of the currently playing file: index, title, start time, and which one is playing. Returns an empty list for files without chapters.
Inputs: no named parameters.
chapter_nextA
Jump to the next chapter. Clamps at the last chapter.
Inputs: no named parameters.
chapter_prevA
Jump to the previous chapter. Clamps at the first chapter.
Inputs: no named parameters.
chapter_seekC
Jump to a chapter by 0-based index.
Inputs: index (required).
cleanupA
Find files an older unflick left behind — on Windows, upgrading from v0.9 moved the install directory and stranded roughly half a gigabyte with no uninstall entry. Reports by default; pass apply to delete. The live thumbnail and cover caches share that folder and are never touched.
Inputs: apply.
clear_chaptersA
Remove chapters added by generate_chapters or set_chapters. A file's own built-in chapters are not affected.
Inputs: no named parameters.
clipB
Extract a video clip segment, optionally as GIF. Requires ffmpeg in PATH.
Inputs: end (required), start (required), gif, file, +1 more.
describe_frameA
Return the frame showing right now as an image, so you can see what is on screen. Optionally seek to a position first. The picture is the one the viewer is watching — GUI, CLI and MCP share a single player. Downscaled to keep the response small; subtitles and on-screen controls are not burned in.
Inputs: max_edge, position.
detect_scenesA
Detect visual scene cuts / shot boundaries in the video using ffmpeg's scene-change filter. Useful for videos without dialogue or transcripts (screen recordings, silent footage, music videos). Set to_chapters: true to immediately install the detected scenes as navigable chapters on the playing file.
Inputs: file, threshold, max_scenes, to_chapters.
dev_captureA
A picture of the unflick interface. This is the UI layer, not the decoded video: mpv draws on its own surface, which sits above the webview on Windows and below it on macOS and Linux, so no single grab means the same thing on all three. Use describe_frame for the picture itself — it comes from mpv, at higher fidelity than any compositor grab.
Inputs: max_edge.
dev_clickA
Click an element in the unflick window. Polls for the selector first, so a click issued right after an action does not race the render, then hit-tests the point before dispatching: if something else is on top, the click is refused and the covering element is named rather than reported as a success. A selector matching several elements is refused unless index says which.
Inputs: selector (required), index, timeout_seconds.
dev_evalA
Run JavaScript in the unflick window and get back what it evaluated to. Either an expression or statements with an explicit return. The escape hatch for what the other five verbs do not cover — prefer them where they fit, because they report their own refusals in terms of the page.
Inputs: script (required), timeout_seconds.
dev_snapshotA
The accessibility tree of the unflick window: every visible element with its role, accessible name, state and a CSS selector that feeds straight back into dev_click and dev_text. Reach for this first — it is how you find out what is on screen. Invisible nodes and unnamed wrappers are elided, so what comes back is what a person would see. Needs the GUI running, started with --allow-dev.
Inputs: depth, selector.
dev_textB
Visible text of every element matching a selector, each with its index. Several matches are returned rather than refused — reading is not acting. Use it to check what a list, a menu or a label actually says, which is where 'undefined' shows up.
Inputs: selector (required).
dev_waitB
Wait for an element to appear, or with gone for it to leave the page. The second half is not optional: verifying that a panel closed is impossible without it.
Inputs: selector (required), gone, timeout_seconds.
disc_listA
Optical drives on this machine and whether each holds a DVD or Blu-ray, plus whether this build can play them at all. Play one with play and the drive path ("D:\"), a disc image ("film.iso"), or a folder holding VIDEO_TS / BDMV; a specific title is mpv's own syntax, "dvd://3". Each drive, and each probed path, also reports key — the identity its bookmarks and resume point are filed under, which changes when the disc does — and label, the volume name.
Inputs: path.
download_subtitleA
Download one subtitle chosen from find_subtitles and load it into the player. Saved next to the video when that directory is writable, so it loads automatically next time. Costs one of the user's limited daily OpenSubtitles downloads.
Inputs: file_id (required), file, load, language.
equalizer_getA
Read the 10-band equaliser: per-band gains in dB, the band centre frequencies they correspond to, the preamp, whether loudness normalisation is on, and the filter chain mpv is actually running. Read this before changing a band so you adjust the user's curve instead of replacing it.
Inputs: no named parameters.
equalizer_presetA
Apply a named equaliser curve and switch the equaliser on. Use equalizer_presets to see what exists and what each is for. 'speech' lifts dialogue out of a loud mix and 'night' tames explosions while keeping voices — those two answer most requests about not being able to hear what people are saying.
Inputs: name (required).
equalizer_presetsA
List the available equaliser presets with a description of what each is for and the curve it applies.
Inputs: no named parameters.
equalizer_resetA
Remove every audio filter: flat curve, no preamp, no normalisation.
Inputs: no named parameters.
equalizer_setA
Adjust the equaliser. Set one band with band (0-9, low to high) plus gain in dB, or the whole curve at once with bands. enabled switches the equaliser in and out without discarding the curve, normalize evens out quiet dialogue against loud action, and preamp makes headroom so boosted bands don't clip. Gains are clamped to ±12 dB. Every change rebuilds mpv's audio filter chain, so set a whole curve in one call rather than nine.
Inputs: band, gain, bands, preamp, +2 more.
file_infoA
Get media file metadata (duration, resolution, codecs) without affecting current playback
Inputs: file (required).
filter_listA
List current video filter values (brightness, contrast, saturation, gamma, hue)
Inputs: no named parameters.
filter_resetA
Reset all video filters to 0 (neutral)
Inputs: no named parameters.
filter_setC
Set a video filter (brightness | contrast | saturation | gamma | hue) to a value in -100..100
Inputs: name (required), value (required).
find_subtitlesA
Search OpenSubtitles and return the candidates without downloading anything. Use this when the user should choose, or to check what exists before spending a download. Each result carries a file_id for download_subtitle, plus moviehash_match — true means it is synced to this exact file. Costs no download quota. Requires the user's own OpenSubtitles API key in the opensubtitles_api_key setting.
Inputs: file, hash, query, languages.
frame_back_stepA
Step back exactly one frame. Pauses playback.
Inputs: no named parameters.
frame_contact_sheetA
Sample a grid of frames (cols × rows) evenly across a time range of the video and return a single tiled JPEG image plus the exact timestamp of every cell. Ideal for surveying what visually happens across a video or section in one multimodal tool call instead of calling describe_frame a dozen times.
Inputs: end, cols, file, rows, +2 more.
frame_stepA
Step forward exactly one frame. Pauses playback, as frame stepping is for inspecting a still.
Inputs: no named parameters.
generate_chaptersA
Give a file that has no chapters a set of them, derived from the pauses in its transcript. The result becomes real navigation: it shows up in chapter_list, marks the progress bar, and responds to chapter_seek. This is a rough heuristic — if you have read the transcript, set_chapters will give better titles and boundaries.
Inputs: count.
generate_subtitlesA
Generate subtitles for a video using whisper. Mode 'local' uses bundled or supplied whisper.cpp; mode 'api' uses OpenAI. Returns the path of the generated .srt file.
Inputs: video (required), mode, model, api_key, +2 more.
get_positionB
Get saved playback position for a file
Inputs: path (required).
get_statusB
Get current playback status including state, file, position, duration, volume, and speed. file is the path the user gave; key is what resume points and bookmarks are filed under (the same string, except for a mounted disc, where it identifies the disc rather than the drive) and label is a disc's volume name.
Inputs: no named parameters.
get_subtitlesA
Find and load subtitles for the playing video from OpenSubtitles, in one step. Picks the best match: a subtitle synced against this exact file if one exists (checked by file hash, so the timing is correct), otherwise the most-downloaded subtitle for the title. Prefer this over generate_subtitles when the video is a released film or show — a human-made subtitle is more accurate and arrives in a second rather than minutes. Requires the user's own OpenSubtitles API key in the opensubtitles_api_key setting; it fails with instructions if unset. Costs one of the user's limited daily downloads, so do not call it repeatedly.
Inputs: file, load, query, languages.
incognitoA
Read or set incognito mode. While on, nothing is written to the play history — including plays started from the CLI or by you. Call with no arguments to read it.
Inputs: enabled.
keybind_listA
List every keyboard shortcut: action id, label, current key, default, and whether the user has changed it. Keys read as Mod+Alt+Shift+key, where Mod is Ctrl on Windows/Linux and Cmd on macOS.
Inputs: no named parameters.
keybind_resetA
Restore a shortcut to its default. Omit action to reset every binding at once.
Inputs: action.
keybind_setA
Rebind a keyboard shortcut. Fails if the key is already taken by another action, naming the conflict rather than silently stealing it.
Inputs: key (required), action (required).
library_listB
List all media files in the library
Inputs: no named parameters.
library_removeC
Remove a media entry from the library by ID
Inputs: id (required).
library_scanC
Scan a directory for video files and add them to the library
Inputs: dir (required).
library_searchC
Search the media library by title or path
Inputs: query (required).
library_transcript_searchA
Search spoken dialogue across every video in the media library that has a sidecar (.srt/.vtt) or cached transcript. Returns each matching cue with its video path, title, and timestamp so you can immediately open that video at that line with play (passing file and seek).
Inputs: query (required), limit.
load_subtitleC
Load an external subtitle file
Inputs: file (required).
mouse_listA
List mouse bindings: wheel up/down, click, double click, middle click, and the four right-drag gestures, each with the action it runs.
Inputs: no named parameters.
mouse_resetB
Restore a mouse trigger to its default. Omit trigger to reset all of them.
Inputs: trigger.
mouse_setA
Point a mouse trigger at an action. Unlike keys, two triggers may share an action — wheel-up and drag-up both raising volume is fine, since the inputs are distinct. Pass "none" as the action to disable a trigger.
Inputs: action (required), trigger (required).
now_playingA
What is playing, described the way a person would: title, artist, album, and whether there is any picture (an embedded cover is not video). Use this rather than get_status when the question is 'what is this', and get_status when it is 'where are we'.
Inputs: cover.
pauseA
Pause playback. Acts on the player the user is watching — when the unflick window is open, this pauses the video on screen.
Inputs: no named parameters.
pitch_correctionA
Read or set whether changing playback speed keeps the original pitch. On by default, which is what makes 1.5x speech listenable rather than chipmunk-like. Turn it off only when the pitch shift is the point, such as checking a musical tempo change.
Inputs: enabled.
playA
Play a video file or URL. URLs from supported streaming sites (YouTube, Bilibili, Twitch, Vimeo, etc.) are extracted via the bundled yt-dlp before playback.
Inputs: file (required), seek, proxy, speed, +1 more.
playlist_addC
Add a file to the playlist
Inputs: file (required).
playlist_clearA
Clear all entries from the playlist
Inputs: no named parameters.
playlist_listA
List all playlist entries with index, path, and current track indicator
Inputs: no named parameters.
playlist_loadA
Load an .m3u, .m3u8, or .pls playlist file into the player's queue. Relative paths inside the playlist are resolved against the playlist file's directory.
Inputs: path (required), append.
playlist_nextA
Advance to and play the next track in the playlist
Inputs: no named parameters.
playlist_playB
Play a specific playlist entry by index
Inputs: index (required).
playlist_prevA
Go back to and play the previous track in the playlist
Inputs: no named parameters.
playlist_removeC
Remove a playlist entry by index
Inputs: index (required).
playlist_repeatA
Get or set the playlist repeat mode. 'off' stops after the last entry, 'one' loops the current file, 'all' wraps around. Call with no arguments to read the current mode.
Inputs: mode.
playlist_saveB
Save the current playlist to an .m3u, .m3u8, or .pls file.
Inputs: path (required).
playlist_shuffleA
Get or set playlist shuffle. Enabling it reorders playback without renumbering entries, and keeps the current track playing. Call with no arguments to read the current setting.
Inputs: enabled.
recent_clearA
Forget the play history. Scanned library metadata is kept — only the played-at times and counts are cleared.
Inputs: no named parameters.
recent_filesA
Recently played files, newest first, with how many times each was played and when. Covers files opened directly, not just ones scanned into the library.
Inputs: limit.
resumeA
Resume playback after a pause. From end-of-file this rewinds to the start rather than doing nothing.
Inputs: no named parameters.
save_positionB
Save playback position for a file (for resume playback)
Inputs: path (required), position (required).
screenshotC
Take a screenshot of the current video frame
Inputs: output.
script_addA
Install a .lua or .js mpv user script into the config scripts directory, enable it, and load it into the running player immediately.
Inputs: path (required).
script_listA
List installed mpv user scripts (.lua and .js) in the config scripts folder, with each script's path, language, description, and whether it is enabled.
Inputs: no named parameters.
script_removeB
Delete an installed mpv user script from the config scripts directory.
Inputs: name (required).
script_toggleA
Enable or disable an installed mpv user script by filename (e.g. "autocrop.lua") or stem ("autocrop"). Enabling loads the script immediately; disabling takes effect on the next launch.
Inputs: name (required), enabled (required).
search_transcriptA
Search the currently playing file's subtitles for a phrase and return every match with its timestamp and speaker (if known). Works with loaded subtitle files, sidecar .srt/.vtt/.ass/.ssa files, embedded text tracks, and subtitles you generated with generate_subtitles. Use this to answer questions about what is said in a video, or to find the moment to jump to. Case-insensitive substring match, not regex.
Inputs: query (required), file, limit, speaker.
seekC
Seek to a position in seconds
Inputs: seconds (required).
seek_to_textA
Jump playback to where a phrase is spoken. Finds the phrase in the subtitles (optionally filtered by speaker) and seeks just before that line, so the sentence is heard from its start. This is the 'skip to the part where they talk about X' tool.
Inputs: query (required), file, speaker, occurrence.
sessionA
What the user was last watching and how far in — and, with action "restore", reopen it there. unflick writes this down every few seconds while playing, so it survives a closed window or a crash. Use it when someone asks to get back to what they were watching without naming the file. "clear" forgets it.
Inputs: action.
set_chaptersA
Set the chapter list for a file that has none, from your own reading of its content. Times are in seconds; they are sorted and clamped to the file automatically. These become real chapters — navigable with chapter_seek and drawn on the progress bar. Fails if the file already ships with its own chapters.
Inputs: chapters (required).
set_speedA
Get or set the playback speed of the window the user is watching. Call with no arguments to read the current rate. Pitch is corrected by default, so speech stays natural — see the pitch CLI command to turn that off.
Inputs: rate, relative.
settings_getB
Read all settings, or a single key if provided
Inputs: key.
settings_pathA
Get the absolute path of the settings.json file
Inputs: no named parameters.
settings_setC
Set a single settings key to the given JSON value
Inputs: key (required), value (required).
settings_unsetC
Remove a single settings key
Inputs: key (required).
set_volumeC
Set volume level
Inputs: level (required).
shutdownA
Shut down the unflick daemon. Declined when the control port is held by the GUI — close the window instead.
Inputs: no named parameters.
sponsor_segmentsA
Fetch SponsorBlock skip segments for a YouTube URL. Returns the configured categories plus a list of {start, end, category, action_type, uuid} segments. Empty list when none exist (404). Errors only on network/parse failure or non-YouTube URLs.
Inputs: url (required).
startupA
The last launch's startup timeline, phase by phase, in milliseconds from process start — process init, video pipeline, window shown, the launch file opening, React mounting, the control port. Use it to say where a slow start went rather than that it was slow.
Inputs: no named parameters.
stopA
Stop playback and unload the file. The resume point is saved unless the file was watched to the end.
Inputs: no named parameters.
subtitle_bilingualA
Show two subtitle tracks at once — typically the original plus a translation. Call with no arguments to read the current state. primary is the bottom line and the one subtitle_select refers to; secondary is drawn just above it. Each accepts a track id from subtitle_list or a path to a subtitle file, which is loaded if it is not already. With no tracks named, the original and the translation are picked automatically. Selecting a single track with subtitle_select turns this off.
Inputs: layout, enabled, primary, secondary.
subtitle_delayA
Get or set the subtitle delay in seconds. Positive values show subtitles later. Call with no arguments to read the current delay. Useful for correcting AI-generated subtitles, which often drift a few hundred milliseconds from the dialogue.
Inputs: seconds, relative.
subtitle_listA
List all subtitle tracks (embedded and external)
Inputs: no named parameters.
subtitle_selectA
Select a subtitle track by ID (0 to disable subtitles). Choosing a single track turns bilingual mode off.
Inputs: id (required).
subtitle_style_getA
Read subtitle appearance: scale, vertical position, colour, border size, bold.
Inputs: no named parameters.
subtitle_style_setA
Set one subtitle appearance property. Returns the full style after the change.
Inputs: name (required), value (required).
transcript_exportA
Export the transcript to a formatted string or file in txt, srt, vtt, or json format, optionally filtered to a single speaker.
Inputs: file, format, output, speaker.
transcript_getA
Return the full transcript of the currently playing file as timed cues (including speaker labels when present), plus where it came from. Use it to read or summarise a video, or as input for writing your own chapter list with set_chapters. Long videos produce a lot of text — prefer search_transcript or transcript_summary when you only need a passage or overview.
Inputs: file, speaker.
transcript_rename_speakerA
Rename a speaker label across the transcript (e.g. from "Speaker 1" to "Host" or "Alice"). The rename is persisted in cache so subsequent transcript_get, search_transcript, and transcript_summary calls reflect it.
Inputs: to (required), from (required), file.
transcript_speakersA
List all speakers in the transcript with cue count, word count, speaking duration, and first/last timestamp. Pass detect: true to run lightweight acoustic + conversational turn diarization when subtitles do not already carry speaker labels.
Inputs: file, count, detect.
transcript_summaryA
Return a compact, structured overview of a video's transcript: cue count, word count, speaking duration, top keywords, and timeline segments with representative lead lines. Much smaller than transcript_get on a long video, so reach for this first when asked what a video is about or how it is structured.
Inputs: file, segments.
translate_subtitlesA
Translate an SRT file to another language using OpenAI. Returns the path of the translated .srt file.
Inputs: srt (required), api_key (required), target_lang (required), output_dir.
video_transform_getA
Read how the picture is fitted to the window: aspect override, rotation, zoom multiplier, pan-scan and deinterlace. Distinct from filter_list, which is colour rather than geometry.
Inputs: no named parameters.
video_transform_resetA
Restore the file's own geometry — no aspect override, no rotation, no zoom or pan-scan, deinterlace off.
Inputs: no named parameters.
video_transform_setB
Fix a squashed aspect ratio, straighten a video recorded sideways, crop black bars, or deinterlace broadcast footage.
Inputs: name (required), value (required).
window_modeA
Get or set the shape of the player window the user is looking at. normal is the full player, pip is a small always-on-top video window, music is the compact audio layout with cover art and tags. Call with no arguments to read the current mode. Only works while the unflick GUI is running — the headless daemon has no window and says so.
Inputs: mode.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 114 tools
Most tools have clearly distinct purposes and descriptions explicitly distinguish pairs like get_status vs now_playing and filter_list vs video_transform_get. However, the surface is large enough that several clusters overlap: transcript_get/search_transcript/transcript_summary/library_transcript_search, subtitle listing/select/bilingual, and playback commands play/resume/stop/pause can be confused in context. Descriptions help, but the volume of near-neighbors makes misselection plausible.
Mostly snake_case, but the set mixes verb_noun (get_status, set_volume, load_subtitle) and noun_verb (chapter_list, subtitle_list, settings_get, bookmark_add, equalizer_reset) conventions without a clear rule. It is readable, but not predictable enough for an agent to infer names reliably.
114 tools is an extreme mismatch for a media player MCP server, far beyond the typical 3-15 well-scoped range. Even when individual capabilities are useful, the surface is too large to navigate efficiently and fragments related operations across many separate tools.
The domain is covered exhaustively: playback control, playlists, subtitles, transcripts, chapters, bookmarks, settings, library, equalizer, video transforms, casting, discs, and developer inspection. There are no obvious lifecycle gaps for a media player; if anything, the coverage is overcomplete.
Maintenance
Related MCP Connectors
MCP server for Clipkit — gives AI agents a video toolbox via the Clipkit schema.
Open-source AI video editor: cut, caption, clip and animate real Revline projects from your agent.
Your YouTube library, in the AI tools you already use: Claude, ChatGPT, Cursor, any MCP client.
AI video editor for agents and humans: timeline, captions, color, audio and generation as MCP tools.
Related MCP Servers
- AlicenseAqualityCmaintenanceLet AI agents watch videos: local transcripts, speaker labels, scenes, chapters and exact-moment search from any video URL or file. Fully local, no API keys.43AGPL 3.0
- AlicenseAqualityDmaintenanceGive any MCP client a real video editor — 32 typed tools over ffmpeg, Whisper and MediaPipe, plus an optional local UI with a drag-and-drop timeline.38MIT
- AlicenseCqualityBmaintenanceOpen-source local-first AI media workbench with Canvas, video editor, Assets, and MCP control.38348Apache 2.0
- AlicenseNot gradedqualityBmaintenanceAI-native timeline editor for short-form video. Your agent cuts, captions, adds motion graphics and renders against a real, editable timeline over MCP — not a render queue. Human-in-the-loop by design: it calls request_review and waits while you supervise in a browser. Local-first and AGPL; your footage never leaves your machine.2AGPL 3.0