Agentic Video Editor
An agentic video editor is a video editing tool in which an AI agent performs the edit itself, and agentic video editing is what that agent does: it indexes your footage, plans a sequence of operations, executes them against a real edit decision list, renders a preview, looks at the frames it produced, and revises. You state the outcome in plain English instead of operating a timeline. This distinguishes it from an AI-assisted editor, where AI features are buttons you press inside a manual workflow, and from an automated editor, which runs one fixed recipe end to end and cannot be redirected once it starts.
Valmera is one. Upload real footage up to 14 GB or 3 hours, describe the outcome, and the agent does the rest — then exports a full-quality MP4 from your original file. It is also drivable by another agent: the complete toolset is published as an MCP server, so Claude can perform the edit from inside a conversation.
Try the Agent on Your Own Footage
50 free credits on every account, no card required. Upload, type the edit, watch the preview.
Start free →The Loop That Makes It Agentic
The difference between an agentic editor and an automated one is a return path. An automated tool goes from input to output once. An agentic one closes the loop: it produces something, looks at what it produced, and decides whether that satisfies the request. Here is the loop Valmera runs, every turn.
Three things in that picture are load-bearing. Step 1 happens once per upload and every later edit reuses it, which is why a request against a three-hour file is not three hours of work. Step 3 writes to an edit decision list (EDL) rather than to your video, which is what makes every operation reversible and the original untouchable. And step 5 is the one competitors do not have: the agent is handed real frames out of the preview it just rendered, carrying a faint tenths grid, so it catches a caption sitting over a face without you telling it and re-aims by measurement rather than impression.
Step 6 goes back to your original upload. Previews render from a smaller proxy because they have to be fast; the deliverable never does. Worth asking of any tool that shows you a preview.
Why It Edits a List, Not Your Video
Handing an autonomous agent write access to the only copy of a shoot would be an unreasonable thing to ask, so Valmera does not ask it. The agent never touches a pixel. It writes decisions into an edit decision list — an EDL, the same idea a non-linear editor (NLE) has used since tape — and a renderer turns that list into video. Every claim about reversibility on this page follows from that one design choice rather than from a promise.
Four consequences fall straight out of the picture. Cutting is subtraction from a list, so undoing a cut is an insertion into a list — restore_range is not recovering deleted data, because nothing was deleted. Every write produces a new version, so "go back two edits" is addressing a number rather than hoping for an undo stack. The final render replays the list against your original file, which is why export quality does not depend on how many turns it took to get there. And because the list is small structured data rather than video, the agent can read the whole current edit back with get_edl before it changes anything.
The two clocks in the caption are the part that trips people, and they are the reason a request like "cut the silences, then caption it" has an order. Silences are removed in source seconds. Captions are timed against the transcript that survived, in output seconds. Do it the other way round and every caption after the first cut is wrong by the length of everything removed before it. An agent that plans is an agent that gets that order right without being told; a fixed recipe gets it right only if someone hard-coded it.
Assisted, Automated, Agentic
Most of the argument about what counts as an agentic video editor is a two-way argument that needs a third box. There are three shapes on the market, and the middle one is the one routinely mislabelled.
| AI-assisted | Automated | Agentic | |
|---|---|---|---|
| Who sequences the steps | You do, between assists | A fixed recipe, once | The agent, per request |
| What you supply | Clicks, drags, trims | An upload and a preset | A described outcome |
| Redirect it mid-job | You are the one driving | No — re-run and hope | Yes — say what is wrong |
| Inspects its own output | No — you review it | No | Renders a preview, reads the frames |
| Dependent multi-step requests | Sequenced by hand | Only what the recipe covers | Sequenced by the agent |
| What you end up with | A project you built | Candidates you curate | An edit you correct by talking |
| Products that sit here | Descript, CapCut, VEED | Opus Clip, Goldcast recipes | Valmera, Mosaic |
To be fair to each name in that last row. Descript is a transcript-based suite with a genuine track record, and its Underlord assistant suggests and performs edits inside that workflow — the closest of the assisted set to agentic. CapCut and VEED are manual timelines with strong AI assists; you make every cut between them, and both are good at that. Opus Clip is the clearest case of the middle column and a capable specialist: one upload yields a batch of candidate clips, each scored 0–99 for predicted virality, and you curate its picks. Goldcast's Agentic Video Editor inside Content Lab is chat-driven, but its preset "recipes" are the automated half. Mosaic markets itself as an agentic platform and pairs the agent's output with a timeline you finish by hand.
None of these are bad products. The point of the middle column is that automation is often the right tool, and calling it agentic makes it impossible for a buyer to tell which one they are getting.
Five Questions That Tell You Which One You Have
Marketing copy will not settle the previous section, because every product in all three columns now uses the word. These five questions will, and each one can be answered in an afternoon with a trial account. Run them against Valmera too.
- 01
Can you redirect it mid-job by describing what is wrong?
Automation has one input: the settings you started with. If the only correction available is re-running the whole thing with a different preset, the tool is not planning — it is executing a recipe. An agent takes “the cuts are too tight and the music is loud under the intro” as its next instruction and works from the edit it already has.
- 02
Does it look at what it produced?
This is the one almost nothing passes. Ask whether the system inspects its own render before handing it over. A tool that writes captions and never sees them will happily place one across a face for twenty minutes. Valmera renders a preview and hands the agent real frames out of it, on every turn.
- 03
Will it take several dependent steps in one sentence?
“Cut the silences, then caption it” is a dependency: captions timed against the raw recording drift the moment a cut lands. If the interface makes you do the two operations in separate passes, the sequencing is your job, and sequencing is most of editing.
- 04
Can it prove what it did?
An agent that reports on its own work has an obvious failure mode. Ask what stops it claiming an edit it did not make. Valmera verifies every reply server-side against the edit decisions actually recorded, and when nothing changed the reply says nothing changed.
- 05
Can anything other than its own interface drive it?
If the agent really is the editor, the agent is replaceable and the tool surface is the product. That is testable: does a documented tool registry exist, and can another model call it? Valmera publishes its complete registry as an MCP server, and it is the same one the in-house agent uses.
A tool can fail all five and still be the one you should buy. If what you need is a hundred vertical clips a month, an automated clipping agent beats a deliberative one on the only axis you care about. The questions are for telling the two apart, not for ranking them.
The Five Jobs of a Video Agent
In "It's time for agentic video editing" (Justine Moore, a16z, 21 January 2026), the work a video agent will take on is split into five parts: Process, Orchestrate, Polish, Adapt and Optimize. It is the clearest map of the category published so far. Here is Valmera against it, with the incomplete rows marked incomplete.
| Stage | What the model describes | Valmera today |
|---|---|---|
| Process | Sort the raw footage — “identifying A-roll vs. B-roll, processing multiple camera angles, and comparing takes” | Partial — find_silences and get_shots turn a raw file into a stringout, and get_kept_transcript flags repeated phrases so a retaken line is findable. No multi-cam angle sync. |
| Orchestrate | Coordinate several models in one project — “generate the images, send them to a video model, and stitch the outputs together” | Yes — generate_image, generate_video, generate_sfx, search_stock and fetch_url all land in the same EDL through insert_media and add_overlay. |
| Polish | “Adjust lighting between clips, clean noise out of the audio track, or take out filler words” | Partial — set_color_grade and apply_look for the look, remove_filler_words and cut_silences for the speech, set_master_loudness for level, enhance_video for picture denoise. No audio denoise. |
| Adapt | Repurpose across platforms — “cut a YouTube podcast into short clips with different aspect ratios to post on your X, Instagram, and TikTok accounts” | Partial — auto_reframe, set_frame and add_aspect_shift cover 16:9, 9:16, 1:1 and 4:5, with captions rescaling to the new frame. No dubbing, translation or publishing, and one clip per request. |
| Optimize | Editorial judgement — “craft a few draft versions of a video for you to iterate on”, answering notes like “the opening is too slow” | Partial — render_preview plus look_at is the review half and it runs every turn, and “the opening is too slow” is a normal instruction. One deliverable per request, so drafts to compare are separate requests. |
The pattern in the gaps is consistent: Valmera is strong where a decision can be measured off the footage, and weak where the job needs a second source — another camera, another language — or a batch of variants. Producing several drafts to A/B test against each other means asking several times.
One Request, Traced
Everything above is a claim about sequencing. This is the shape of a single turn on a talking-head recording — one sentence in, and the calls the agent makes before it answers. Arguments are abbreviated and the timings are one illustrative example rather than a benchmark, but every tool named is real and documented on the tool reference.
YOU Cut the dead air and the ums, karaoke captions, chill
music under my voice, make it a Reel.
AGENT find_silences(min_len=0.35) 84 pauses, 3m12s total
get_words(0, 1140) word-level timings
cut_silences() 19m04s -> 15m52s
remove_filler_words() 61 removed -> 15m31s
add_captions(preset="karaoke") timed against the KEPT
transcript, not the raw one
list_music_library(mood="chill") candidate tracks
add_music(duck=true) fitted to 15m31s
set_frame("9:16", mode="crop")
auto_reframe() samples kept frames, aims
the crop at the speaker
render_preview() -> v4
look_at(output_times=[0.4, 62.0, 310.0, 900.0])
^ the agent reads its own frames here, and finds this:
set_caption_style(position="middle")
render_preview() -> v5
AGENT "Cut 3m12s of pauses and 61 filler words — 19m04s down
to 15m31s. Captions were sitting over your chin once
the crop went vertical, so I moved them up. Music
ducks under speech."Read the order rather than the count. get_words runs before any cut because no tool in Valmera accepts a timestamp the model invented. Captions come after the cuts, for the two-clock reason above. Music is fitted last of the audio work because its length is a function of an edit that did not exist when the request was typed. Reframing comes after captions so they can rescale into the new aspect ratio. None of that order was in the request. A fixed recipe can encode one such order; it cannot derive the order for a request it was not written for.
Then the last three lines, which are the ones that matter. The agent rendered, looked, found a caption over a chin that nothing in the transcript or the timeline could have predicted, fixed it, and rendered again — before telling you anything. That is one turn, not a support ticket. And the reply describing it was checked server-side against the decisions actually recorded, so "3m12s of pauses, 19m04s down to 15m31s" is a read of the EDL rather than a sentence the model composed.
What Makes Valmera's Editing Agentic
It plans multi-step edits from one sentence
"Cut the dead air, add karaoke captions, put chill music under my voice and reframe it for Reels" is four operations with dependencies between them. The cuts change the timeline the captions must be timed against, and the music has to be fitted to a program length nothing knows until the cuts are done. The agent sequences that itself.
It cannot lie about what it did
Every reply is verified server-side against the edit decisions actually recorded. The agent cannot claim a change it did not make, and when nothing changed it says so. This matters more in an agentic tool than an assisted one: you are trusting a report rather than watching your own hands.
Nothing it does is destructive
The agent edits a decision list, not pixels. Your original file is never modified, every operation creates a new version, and anything cut can be restored by asking. Otherwise, handing an autonomous agent the only copy of a shoot would be an unreasonable thing to ask.
Which Decisions the Agent Owns
"Autonomous" is the same category named by how much the tool decides alone, and asked that way the question has no answer — an agent that decides everything is reckless and one that decides nothing is a toolbar. It becomes answerable when you sort the decisions by whether they can be measured off the footage.
MECHANICAL — the agent owns these
Where the silences are. Where a word starts and ends. Which phrases repeat because a line was retaken. Where the shot genuinely changes. What loudness the mix should hit. Every one of these has a right answer measurable off the file, so asking you about them wastes the turn. Valmera decides them and tells you what it did.
CRAFT — the agent proposes, you correct
How tight the cuts feel, when to punch in, which track fits, how big the captions want to be, whether the grade is doing too much. These have better and worse answers rather than right ones, and they depend on taste you have not stated yet. The agent commits to a defensible default, shows you the preview, and treats “looser” or “different track” as the correction it was waiting for.
INTENT — not the agent’s to make
What the video is arguing. Which of two takes is the honest one. Whether a clip should exist at all. An agent cannot recover these from the footage because they are not in the footage. Valmera has one tool for this, ask_user, and it is reserved for exactly the calls no measurement can settle — an editor who asks about everything is as useless as one who asks about nothing.
The gradient is also why autopilot is the wrong word for this. Autopilot implies you set a destination and stop participating. Here you keep the intent tier and hand over the other two, which is the division of labour a good human editor already works under. The autonomous video editor page works the same argument from the autonomy side.
How to Edit a Video with an AI Agent
- 1Upload your footageUp to 14 GB or 3 hours per video, in MP4, MOV, MKV or WebM. Valmera indexes it once — word-level transcript with speaker labels, silences, shot boundaries and labeled frame tiles — with visible progress. Every later edit reuses that index.
- 2Describe the outcome, not the stepsType what the finished video should be: "cut the dead air and the ums, add karaoke captions, put chill music under my voice, and reframe it for Reels". The agent sequences the dependent operations itself.
- 3Judge the preview and refine by talkingThe agent renders a preview and inspects the frames. React in plain English — "looser cuts", "bigger captions", "different track" — and each refinement is another full pass. Export a source-quality MP4 from your original file when it is right.
Indexing a long upload takes a while and shows progress throughout; the analysis is done once and reused by every edit after it.
What Happens With a 3-Hour Recording
Valmera accepts a single upload up to 14 GB or 3 hours, in MP4, MOV, MKV or WebM. It is worth asking any tool in this category for its ceiling, because almost none publish one. Goldcast's help documentation is the honourable exception, and it is instructive: it states that the Agentic Video Editor in Content Lab is currently limited to videos of about five minutes, and that a subscription covers 15 individual videos per year, with an add-on for more. Both are real constraints that a feature list would never surface, and both are the kind of number that decides whether a tool can touch your actual footage.
Length is the hardest axis in agentic editing, for a mechanical reason: a three-hour recording is roughly thirty thousand spoken words, and no model reads that off the pixels on request. It reads the index. So Valmera pays that cost once, up front — word-level timestamps with speaker labels, every silence measured, shot boundaries detected, and labeled frame tiles so the agent has actually seen the footage rather than been told about it. Indexing shows progress throughout; every request afterwards reads the index and is fast.
What that buys on a three-hour multi-guest podcast: removing every pause is one call rather than a scrub through the stringout, because silence detection already ran; filler word removal covers the whole file in the same pass, and cuts land on word boundaries so nothing sounds clipped; "find where he talks about the pricing objection" is a transcript search returning real timestamps rather than a guess; and cutting that answer down to a 9:16 aspect ratio for Shorts, TikTok or Reels is one more sentence, with captions rescaled and the crop aimed at whoever is speaking. Speaker labels are what make the last part work: with two people in the room, "keep only the guest's answers" is a cut by speaker, not a cut by guesswork about who is on camera.
Two honest limits at that length. Valmera has no multi-cam sync, so a three-camera shoot has to be one cut file before it arrives — the EDL addresses one main video, not a bin of angles. And Valmera produces one deliverable per request, so turning a long recording into ten Shorts is ten directed requests rather than one batch — which is a genuine strength of the automated clippers and a genuine weakness here.
The Agent's Toolset, Enumerated
108 tools: 97 editing, plus 11 session tools for projects, uploads, indexing, rendering, export and download. None of them touch your footage — they either read the index or write to an edit decision list. These names are the live registry, not a summary of it; what each does and refuses is on the tool reference page.
Cutting and transcript
The keep list decides what survives. Cuts snap to word boundaries, and anything cut can be put back.
cut_silences · remove_filler_words · cut_range · cut_output_range · restore_range · keep_segments · beat_align_cuts · reset_edit
Captions and on-screen text
Captions are the spoken words. Text elements are designed graphics. Separate systems, separate tools.
add_captions · set_caption_style · set_caption_mutes · set_caption_fixes · add_text · add_kinetic_text · add_text_behind · remove_text · add_title_card · add_color_screen · add_corrupt_screen
Framing and motion
Zooms and speed are what make a talking head watchable. Framing decides which platform the result belongs on.
set_frame · auto_reframe · add_aspect_shift · remove_aspect_shift · add_zoom · remove_zoom · add_zoom_path · remove_zoom_path · punch_in_on_emphasis · set_speed · remove_speed · add_freeze_frame · set_transitions · set_fades
Audio and music
Music, voiceover, sound effects and the speaker's own audio are four layers with independent gain.
list_music_library · add_music · swap_music · set_music_fit · remove_music · extract_audio · list_sfx_library · add_sfx · move_sfx · remove_sfx · generate_sfx · sound_design_pass · add_voiceover · remove_voiceover · set_audio_gain · set_volume · set_master_loudness
Colour and look
Grades apply to the whole program; stylize effects are windowed. The two compose.
set_color_grade · set_grade_custom · add_stylize · remove_stylize · enhance_video · apply_look
Inserts and B-roll
Inserts interrupt the program; overlays sit on top of it. Both accept anything in the project — uploaded, fetched or generated.
insert_media · set_insert_window · move_insert · remove_insert · add_overlay · move_overlay · remove_overlay · add_screen_takeover · remove_screen_takeover · fetch_url · search_stock · add_stock_media · generate_image · generate_video
Repair and censoring
Two jobs that get confused. Erasing removes something so it is gone; censoring puts a visible mark over it on purpose.
find_burned_text · erase_burned_text · erase_region · remove_erase · blur_region · remove_blur
Screen recordings
A headless browser records a page, or drives a product and records itself using it, with real click timings to cut against.
record_website · record_website_demo · showcase_demo · enhance_cursor · remove_cursor_enhance · set_screen_frame · remove_screen_frame
The agent's senses
Where real timings come from — no tool accepts a timestamp the agent invented — and where it checks its own work. Nothing here writes to the edit.
get_video_info · get_transcript · get_words · get_kept_transcript · search_transcript · get_shots · find_silences · get_audio_analysis · suggest_emphasis · list_assets · get_edl · look_at · look_at_asset · render_preview · read_skill · ask_user
Session and project (MCP only)
The context the studio UI gives Valmera's own agent for free, and a headless model has to be handed.
list_projects · open_project · create_project · project_state · upload_start · upload_finish · index_status · export_final · wait_for_job · download_url · watch_video
A tool whose backing service is unconfigured is hidden from the registry rather than exposed and failing on call, so a live session can receive fewer than 110. Two of the names above are hidden right now, and naming them is more useful than the caveat: list_sfx_library and sound_design_pass place sounds from a bundled effects pack that has been withdrawn, so they disappear from the registry until it comes back. Generating a sound effect from a description still works — that is generate_sfx, and it is unaffected. If you want to check any of this rather than take it on trust, call tools/list against the MCP endpoint; what it returns is the registry, and this page is a mirror of it.
Some of these have no manual equivalent in the studio at all. B-roll search, screen-recording capture and burned-text erasure exist because the agent needed them, not because a panel was designed for them — which is the practical difference between building an agent into an editor and building an editor for an agent.
Drive It from Claude
The distinction between an agentic editor and an assisted one has a clean technical consequence: if the agent is the editor, the agent can be swapped. Valmera publishes its complete tool registry as a remote Model Context Protocol server over streamable HTTP, with OAuth 2.1, dynamic client registration and PKCE — so you add it as a connector in the Claude app, sign in, and edit video from the conversation. No token to copy. The setup guide covers the Claude app and Claude Code.
The tools are not re-declared for MCP. The engine publishes one registry and the connector serves it verbatim, so the model in your Claude session gets exactly what the in-house agent has — same names, same schemas, same refusals, and no second list to drift. Editing the same project from the studio and from MCP at once is refused in both directions rather than silently interleaved.
If your working environment is already a chat window, an editor your assistant can call is a different product from one you have to leave the conversation to open. We keep an honest comparison of every video editing MCP server, including where another one is the better answer.
Where an Agentic Editor Is the Wrong Tool
A tool that will not name its edges is asking you to find them yourself. Here are Valmera's.
- You want video generated from a prompt. Valmera edits footage you upload. It can splice in short generated clips and stills, but there is no whole-video text-to-video. That is a different category.
- You need frame-level authority over everything. A traditional NLE — Premiere, DaVinci Resolve, Final Cut — still wins on multi-cam sync, node-based colour, keyframed compositing, motion tracking and per-cut transition choice. Valmera applies one transition style per video, hard cuts only: no true crossfade or dissolve.
- You need a subtitle file. Captions are burned into the picture — no SRT or VTT import or export, and no chapter metadata.
- You need audio restoration. No denoise or "studio sound", no per-speaker leveling, no separating music out of an already-baked track. Loudness mastering and per-layer gain are there; repair is not.
- You need volume. One deliverable per request. For one recording into a batch of ranked clips in a single pass, an automated clipper is the right shape and Valmera is not.
- You need a team. No seats, collaboration, share links, brand kits, or direct publishing to YouTube or TikTok — export is a file you download and post yourself.
- Also worth knowing. No custom font uploads (12 families are bundled), no motion-tracked overlays, no native mobile app (mobile browsers work), and English is the best-tested transcription path.
The agent knows these edges too: out-of-scope requests get refused rather than faked, by the same honesty layer that stops it claiming an edit it did not make. The export documentation covers what a finished file contains on each plan.
Frequently Asked Questions
Edit Your First Video Free
50 credits on signup, no card. Or connect Valmera to Claude and edit from the chat you are already in.
Start free →