Meet Vibe

Vibe is the AI that edits alongside you. You tell it what you want in plain language, and it makes the changes on your timeline. It isn't a separate tool you switch into. It's editing the same composition you are, so anything it does, you can adjust, and anything you set up, it can build on.

Opening the panel

Vibe lives in the editor header. Click the sparkles button (Klipeo agent) next to Export and the panel takes over the timeline area. A layout toggle in the Vibe header moves the panel between the bottom and the right side, whichever suits the edit you're doing.

On a phone, the editor opens straight into a full-screen Vibe screen instead of the desktop timeline, with the same transcript, composer, and media gallery.

Talking to it

Type your request in the composer and press Enter to send. Shift+Enter adds a newline. Press Escape once to arm a clear, and again to start a fresh chat. The composer shows Starting agent… while Klipeo establishes the durable run. Your text and attachments remain in place until the server accepts that run; if startup fails, correct the problem and press Send again without rebuilding the request.

Checkpoints are saved automatically by the harness; they are not tools you need to request. After reconnecting, matching editor receipts can be retried without applying the same edit again.

New editor sessions apply a tool's edit batch and confirm its composition save before reporting the tool as completed. If saving fails, keep the tab open and resolve the save error; a visible local edit is not yet a saved edit. If the editor cannot confirm it, Vibe stops further edits and asks you to reconnect and inspect the timeline; retrying blindly could duplicate changes.

Inserted, overwritten, and appended clips retain their IDs across follow-up edits. Insert shifts unlocked lanes together; overwrite clears only the target lane's covered span. Detaching audio returns a separate audio clip and mutes the video; it does not automatically group the two clips.

A few things you can do while writing:

Voice input uses the same composer on desktop and phone. It stays open while you speak even if a local agent connection changes; sending the finished request is still subject to the selected agent being available.

If opening the microphone takes more than 15 seconds, Vibe stops the attempt and shows a retry message. A take runs for up to three minutes and keeps going through pauses: if the transcription service drops the connection while you think, Vibe reconnects underneath and keeps what you already said. After you finish speaking, realtime transcription has an 8-second deadline; uploaded recordings have a 50-second deadline. You can cancel while waiting. Cancelled or timed-out takes never insert a late transcript into your draft.

Keep your asks concrete. "Trim the dead air and add TikTok-style captions" lands better than "make this good," because Vibe can map it to real operations. There's a longer guide to phrasing in Skills and cost habits in Save on AI cost.

Run mode: Ask, Plan, Auto Approve

Under the composer, next to the + menu, a dropdown picks how Vibe handles permission for a turn. It reads Ask by default.

• Ask — Vibe edits right away but stops and asks before anything destructive: deleting or replacing clips, tracks, assets, captions, effects, and similar. Reversible additions and adjustments proceed on their own. The first time a request needs paid generation (images, video, voice, music, or sound effects), Vibe asks once, and that approval covers the rest of the request.
• Plan — Vibe proposes the steps first and waits for your approval before anything moves (details below).
• Auto Approve (the pill reads Auto) — Vibe makes destructive edits and generates media without a confirmation prompt, and any question or approval it does raise is answered for you: approvals are approved, clarifications are handed back to Vibe to decide. Switching to it while a run is going applies to that run too. Use it when you want the run to finish on its own and you're ready to undo anything you don't like.

In Plan mode Vibe doesn't touch the composition at all. It can look at your footage, transcript, and timeline with read-only tools, read a Skill's instructions, write the steps down, and ask you questions. It finishes by showing a titled plan with a one-line summary and the ordered steps, each labeled with the Skill or tool it will use, and one question: Approve this plan?, with Approve and Revise. Pick Revise, say what's wrong, and it plans again. Nothing runs on your timeline until you approve.

The mode stays where it is between turns, so a plan-first habit sticks until you switch it. It's a separate thing from the checklist Vibe writes for itself mid-run, covered under The plan, and what it remembers below.

What happens after you hit send

Vibe works in the open. You'll see it think, run its tools, and change the timeline as it goes, rather than going quiet and dropping a finished result on you.

Because every change streams in as it happens, you can stop reading the summary and just watch the preview. If something looks wrong, you don't wait for the end. You fix it right there.

Chat text can continue arriving while an editor refresh is pending. Timeline changes still apply in order. If the connection drops, Vibe reconnects to the existing run rather than starting the job again; the header shows Reconnecting… while the connection is being restored. If the transcript has not loaded yet, the chat shows a recovery message instead of the new-conversation suggestions.

A temporary error while recovering a saved run does not cancel that run. Vibe keeps its resume record and tries again when the browser comes back online. If recovery still fails, reload the editor after checking your connection.

Saved run events load from the database before the live event connection starts. A deployment can use a separate event gateway for the live connection. Your existing sign-in authorizes each reconnect; there is no second account to connect. Reloading or closing the stream does not send Stop to the agent. A slow live connection does not hold up that saved history. If live delivery loses an event or disconnects, Klipeo reads the missing committed events from the database before continuing. While replaying from the database, Klipeo retries an unavailable live transport after a five-second cooldown. You do not need to refresh to trigger that retry.

You can send a correction while Vibe works: press Enter, or the send button that appears beside Stop once you type. The draft stays in the composer until the server accepts it. Then the message appears in the chat, dimmed, with Queued · Vibe reads this at its next step. When Vibe reads it, the mark closes into a check with Read by Vibe, and Vibe's answer to it starts as a new reply below your message. Until then, Edit takes the message back into your draft, added after anything you have already typed. Text-only corrections with unchanged settings can enter at the next safe model step, after the current tool call finishes. Attachments, changed plan settings, and approval answers remain separate turns. Sending a correction does not stop the current job or bypass an approval. Queued corrections are restored when you reopen the active run. If a request times out, retry the unchanged draft to check the same submission. A newly selected preview target must wait for a separate turn because corrections use the active run's editor context.

While a reply is being written, its text, tool groups, generated media, and delivery cards appear in the order they happen. You can expand or collapse each tool group while the agent works. The current group stays open between calls unless you collapse it yourself; earlier groups collapse when the reply moves on unless you opened them manually. Once the reply finishes, its work folds into one Worked for 1m 5s · 12 steps summary that starts collapsed: the tool groups and the notes Vibe wrote along the way sit inside it, and the answer, generated media, and delivery cards stay in view below it. Open the summary to see every step in order. This also applies when you reopen a saved chat. Older replies retain their previous layout. Failed-step counts remain visible in the collapsed summary; unfinished steps are not marked as successful when a stream stops. Text and tool-start updates do not wait for editor refreshes. One shimmering status row remains visible throughout the active run, including between tool calls. Its label follows the actual phase: Catching up, Reading your timeline, Thinking it through, Writing the reply while the answer streams in, or what the current step is doing, such as Generating a video shot or Looking at the media. Every step in a tool group reads the same way, in plain words while it runs and in the past tense once it finished. It does not cycle through invented editing or rendering activity. Reduced-motion settings keep the label readable without shimmer.

Pending image, video, and audio work uses the same grid-reveal loading card. Each active media tool keeps its own card, including while another result has already arrived. Transcription, media import, export, and delivery steps use the same treatment while their tool is running; a completed or failed tool removes its pending card.

Text corrections can also be consumed at the next step of an approval-authorized run. A correction does not inherit the earlier approval: newly requested effects still pass the normal action-specific permission checks. A provider call already in progress finishes before the next step can consume your correction.

Long cloud conversations can create a model-generated context handover at a safe step boundary. Klipeo saves the handover before replacing the model context and retains active skill references and the files attached to the current request. A handover never happens before the model has seen your request. This does not remove the visible chat history.

Sharing a chat

Select Share in the agent header, choose invited people or anyone with the link, then select Create share link. Chats are private until you explicitly create a link. Invited readers must sign in with a verified email matching the invitation; sending email invitations is not automatic.

The read-only page shows user and assistant messages and tool status in pages of 200 messages. Select Older messages to continue through history. It includes future messages added to this chat, but does not grant editor access or expose private media and raw tool inputs or outputs. Public readers can copy the visible text. Revoke link disables future access; it cannot erase copies someone already saved. Creating a new link after revocation does not reactivate the old link.

Sharing requires its database migration to be deployed. If access cannot be loaded or a change cannot be confirmed, close and reopen the dialog to check the saved state before retrying. Chat-share pages and their bearer links are excluded from client analytics and replay.

Continuing an edit

If Vibe needs something from you mid-turn, it asks. The question shows up in the composer as a small wizard: a text box, a single-choice list, or a multi-select, depending on what it needs. Answer it and the turn carries on.

Anything Vibe generates during a turn (images, video, audio) shows up inline as a media card in the transcript, and collects in the media gallery so you can find it again later without scrolling back.

Vibe can also rework a clip that is already in your media bin: change its look or light, continue it past its last frame, or swap a person in it for the one in your photos. Each result is a new clip in the bin, so the original stays as it was.

Vibe looks at its own work. After a visual edit it captures stills from the preview and looks at them before saying the edit worked, and it looks at every picture or clip it generates before placing it, so a warped hand or a wrong product gets caught and redone. Ask what is in a clip or photo in your media bin and it looks at that too. The stills go to the model for that turn only; they are not saved in the chat.

The Outputs card distinguishes downloadable files from render submissions. A render receipt confirms that the request was submitted, not that a finished video is ready. The card checks the render job and shows queued, rendering, uploading, completed, failed, or cancelled status. A completed job with an available download gets a Download video action. If checking fails, use Check again; this only refreshes the status and does not submit another render.

Several agents at once

A turn can run several tools in the same round, and when it does, each one shows as a named production role directly on the timeline at the point where it is working. This keeps the transcript stable while two edits land in different places at the same time. The pointers clear when the turn ends, and one that fails stays put with the error on it.

Generated shots don't wait for each other. When a request needs several clips or pictures, Vibe starts up to four at once and they render side by side, so a three-shot ad takes about as long as its slowest shot. In article mode, where each result also lands on the moodboard, generations still run one at a time.

For bigger jobs Vibe hands work to worker agents. When each shot needs its own make, check, and fix loop, it gives up to four workers one job each. They work at the same time, look at every result, redo what is wrong, and report back which clip or picture to use; Vibe then places the results on the timeline itself. Workers make new media and read your project but never edit the timeline. Each worker's steps show in the reply as their own group under the job's name, such as Shot 2 · waves · 4 steps, and its cursor on the timeline carries the same name. The groups stay the same when you reopen the chat. They generate under the same approval as the rest of the request, so in Ask mode the one generation approval covers them too. A correction you send while workers run reaches Vibe after each worker's current step: every worker stops there and reports what it finished and what is left, so you don't wait for the whole job. A correction that needs its own turn, such as one with an attachment or changed plan settings, lets the workers finish.

The plan, and what it remembers

This is the checklist Vibe keeps for itself during a run, not the Plan run mode above. For work with several meaningful stages, Vibe writes a checklist and ticks the steps off only when tool results confirm them. The latest checklist is stored with the chat, so it survives a reload without stacking stale earlier versions.

When you state a lasting preference, such as a house style or an audience, Vibe saves it with a remember step into your AI memory in Settings, and every future chat sees it. Say "forget that" and it removes the line. Inside a project, facts about the project go to the project's Memory tab instead.

Checkpoints and rewind

Before every cloud turn edits the composition, Vibe saves a checkpoint. The newest 20 per composition are kept, and each shows as a step in the transcript. Ask Vibe to rewind to one and it replaces the composition with that saved state as one undoable edit. In Manual mode a rewind needs your approval, like any other destructive change.

Approvals

When a tool would change the timeline, publish somewhere, or spend credits, Vibe shows an approval card in the composer that says what the tool will do. Answer it and the run carries on.

Generation asks once per request. When Vibe first needs to generate an image, a clip, a voiceover, music, or sound effects, the card asks whether it may generate media for this request. Approve it and Vibe makes every shot, line, and cue the request needs without asking again for each one. The approval never covers replacing or deleting your own media, and a new message starts a new request. Swapping a person into a clip asks every time, in Auto mode too, because it puts a real person's likeness into your video. If you close the editor while a run is still working, reopening it reattaches to the run and any edit that was in flight lands.

On the Klipeo cloud engine it always works this way, and there's nothing to configure. On a local engine the Vibe header adds a two-state control:

Auto only takes effect once you've confirmed that engine in the connect dialog, and it's confirmed per project. Switching between Manual and Auto clears the confirmation, so the next local run asks you again before it starts. See Engines and models.

Chats and history

Preview frames captured by Vibe are saved privately with stable, authenticated image links, so their chat cards can load after a refresh. Older captures whose image data was not retained show a message asking you to inspect the preview again. Shared text-only chat links do not grant access to these private images. Capture waits for decoded video frames. A load or capture failure is reported instead of being treated as proof that the requested visual change succeeded.

Sessions are saved on the server, so a chat survives a reload. The Vibe header has a recent-chats dropdown where you can load an old chat, rename it inline, or delete it, and a New button to start a clean one.

What it can actually do

Vibe drives the same editing operations you have by hand, grouped into areas like captions, cutting and trimming, audio cleanup, color, transitions, zooms, voiceover, and full end-to-end edits. The curated recipes for the common jobs are listed in Skills. If you can describe an edit in terms of those, Vibe can usually do it.

Talking footage gets the same Smart Cut as the editor's panel. Ask Vibe to tighten a clip, cut the ums and long pauses, or keep the best take of each line, and it picks the passes that fit, cuts what it is sure of in one undoable edit, and tells you what it left in.

Recordings made in Klipeo carry their cursor data to Vibe, so it can zoom where you clicked and cut the dead waits from what you actually did rather than guessing from the picture.

When something goes wrong

If a step fails, Vibe says so in the chat instead of pretending it worked. You'll see which tool had trouble. Raw provider errors are not displayed in the step list. Usually you can rephrase and try again, or do that one piece by hand and let Vibe carry on with the rest.

When the model returns a tool call it can't finish, such as broken arguments or a tool that isn't loaded yet, Vibe asks it to correct the call and carries on, up to twice per request. If Gemini stops a step under its content policy, the chat names the policy reason (for example SAFETY) instead of ending silently. Busy or rate-limited model calls are retried after 1, 2, and 4 seconds.

Where Vibe runs

By default Vibe runs on Klipeo's cloud, so there's nothing to install. Every cloud turn runs on Gemini 3.8 Flash with high thinking effort, so there is no model to pick. You can also point Vibe at a coding CLI on your own machine (Claude Code, Codex, Gemini CLI, or OpenCode), which is free to run and lets you choose that CLI's model. That's all in Engines and models.