Editing with vedit

vedit is a timeline editor. The project is one JSON document; the render is a pure function of it (one ffmpeg filter_complex). Every change goes through the same operations whether it comes from the web UI, the in-app chat or this MCP server.

Set up

Needs the vedit MCP server (claude mcp add vedit -- uvx vedit-mcp, already wired if you installed this as a plugin) and ffmpeg. If ffmpeg is missing: install_ffmpeg. Call open_ui if the user wants to watch the edit live in the browser; both share one project.

Workflow

project_create -> import_media -> preview_media (look at the footage first)
-> plan_edit (pick the material) -> add_clips -> apply_style -> your own refinements
-> preview_grid / preview_frame (LOOK) -> review_edit -> render_video (only if asked)

• Times are in seconds. Animatable parameters take {"kf":[{"t":0,"v":0},{"t":2,"v":1}]}.
• project_info shows the timeline with ids. Call it when you do not already know the state.
• Many cuts: add_clips (atomic, one call). Many other changes: batch (atomic, one undo). One call per cut is slow, floods undo history and leaves a half-built timeline if it fails.
• preview_frame returns the real image. Look at it instead of assuming.
• Cutting to music: music_beats gives BPM, bar offset and the energy profile. Do not guess the tempo.

Rules that stop a video looking flat

1. The music must move: automate its gain with keyframes, pick the section with music_beats.
2. Not all hard cuts. A cut is the default, not the only choice; every transition needs a reason.
3. Never show the same shot twice. If the footage is not enough, make the video shorter.
4. After ten seconds the viewer needs new information; cutting faster on the same images does not help.
5. "Cinematic" is a choice, not a filter: one strong moment held for three seconds beats ten half-seconds.

review_edit counts the choices in the timeline and flags cut-only edits. Until it stops warning, the edit is not finished.

Motion graphics: use blocks first

list_blocks then add_block puts a ready, tuned animation on a track above the footage. 52 blocks in six families: titles (lower_third, title, mask_reveal_title, glitch_title, word_cycle, countdown, logo_shine, chapter_card, end_card), text (caption_pill, kinetic_words, typewriter, chat_messages, notification_toast, location_pin, ticker, quote...), data (counter, bar_chart, donut_chart, line_chart, timeline, rating_stars, price_tag, checklist, stat_card, steps, comparison...), emphasis (highlight_box, circle_highlight, arrow_pointer, marker_highlight, subscribe_button, alert_banner), transitions (color_wipe, flash_frame, iris_wipe, blinds_wipe) and style (letterbox, vignette, film_grain, light_leak, scanlines, viewfinder). One call, no CSS; list_blocks(block=...) gives the parameters. Only when no block fits write HTML with add_html (transparent background; the page clock is the clip clock, so preview and render match). Particles: add_particles. Highlighting a moving subject: track_mask, highlight_subject.

Camera moves

camera_move(clip, move) (list_moves): push_in, pull_out, slow_push, pan_left, pan_right, tilt_up, tilt_down, drift, punch_in, zoom_burst, whip_pan, impact_shake, beat_pulse (needs the BPM from music_beats), handheld, sway, dutch_tilt. The edges stay covered, one undo. Pick one for what the shot says; a different move on every clip is noise.

Plan before building

For more than a handful of clips, or anything that generates images or video (it costs), call propose_plan first: scenes with what, seconds, source and an estimated cost. The user sees times and totals and approves or asks for changes; only then build. An approved plan is saved next to the project.

Narrated videos (voice first)

Write the script, generate the voice first (generate_voice), then edit on the measured times of the voice (speech_times), never on estimated ones. One html clip per chapter, no set_speed on the voice, duck the music under it (duck_music).

Match a reference video

learn_style measures how a video is edited (cut rhythm, transitions, camera moves, voice-over and where cuts fall on its pauses, palette, colour cards) and saves a style. It returns two contact sheets: look at them and add what cannot be measured (characters, on-screen text, graphics) with create_style(style=<id>, guide=...), rewriting the whole guide. Then apply_style after the clips are in.

Generating media

generate_media (image or video, from a prompt, optionally starting from an image) needs a service set up by the user (generation_services says which is ready). It costs the user's credits, so it asks the user; never loop on it.

Never do on your own

• Export (render_video) and publish (publish_video): only when the user asks. The editor asks them to confirm.
• Generate media, install a skill: they ask for confirmation too.
• When a choice is the user's (style, length, format, which version to keep): ask_user with 2-4 options, the recommended one first, instead of deciding at random.

If something is wrong, awkward or missing in the tool itself, report_issue and keep working.