A folder of footage becomes a finished video through dozens of decisions: which take to use, where to cut, how to balance the sound, and what each delivery version needs. Claude and Codex can help with that work, but useful assistance requires more than a conversation about editing.

With Vagon Cloud Computer, you can bring your AI assistant, desktop video editor, media files, and supporting tools into one Windows workspace. Prepare footage, develop an edit plan, perform supported timeline operations, and inspect the exported results without moving between disconnected environments.

The connection matters. MCP integrations expose specific application operations, while supported computer-use tools let an assistant navigate the editor’s interface. Scripts can handle file preparation and technical checks alongside both.

This guide follows a practical workflow from raw media to a reviewed cut, captions, and delivery versions. You can use Claude or Codex, with editorial decisions and final approval remaining yours.

What Can Claude and Codex Actually Do in a Video Workflow?

Claude and Codex can assist at several stages of production, but their access determines what they can accomplish. Preparing an edit plan from a transcript is different from changing clips in an open Premiere Pro sequence.

Think of the workflow in three layers:

Both Claude Code and Codex support MCP connections. The available editing operations, however, come from the connected server. A bridge that reports sequence information does not necessarily support trimming clips, editing captions, or exporting.

Computer use covers a different need: operating the application through its interface. It can help open a sequence or inspect export settings, but seeing the timeline is not equivalent to checking every frame or listening to the complete mix.

You do not need both assistants to begin. Choose one, give it a bounded task, and inspect the result. Neither should receive a fixed role based on assumptions that Claude is inherently more creative or Codex more technical.

For this guide, the target is a 60-second product video, a shorter vertical version, and checked captions. Each stage should produce something concrete you can review before proceeding.

Source: OpenAI Codex Documentation

Build Your Editing Workspace on Vagon

A complete editing workspace needs more than an AI chat window. The video editor, source media, fonts, graphics, supporting tools, and assistant must be available where the work happens. On Vagon, that environment runs on the cloud computer, while your device provides access to its desktop.

Choose the Editor and AI Client

Create a Windows cloud computer and use Application Selection to choose the available applications your project needs. Use Premiere Pro or DaVinci Resolve as the primary editor rather than preparing two parallel workflows.

Add your preferred AI client where available, then complete its sign-in inside the cloud desktop. If you intend to use Claude Code, confirm that specific setup; selecting Claude does not automatically establish that Claude Code is installed.

Source: Claude Code Documentation

Launch the editor and activate any required license. Check the application version before choosing an MCP bridge, since integrations can depend on particular releases. Preinstallation prepares the software, but AI subscriptions, commercial licenses, and application connections remain separate.

Bring the Complete Project with Vagon Files

Use Vagon Files to transfer footage, audio, logos, and other project assets. Transfers can be prepared even while the cloud computer is stopped.

Keep the working project organized:

Preserve original media in 01_Source and direct generated files to the working or export folders. For an existing project, include linked assets and confirm that the editor can locate them. Install required fonts and check plugin dependencies before treating missing graphics or effects as an AI problem.

Match Performance to the Work

Choose resources for the footage and editing workload, not simply for running the assistant. High-resolution media, effects, and demanding codecs can place different demands on CPU, GPU, memory, and storage.

Vagon’s scalable performance lets you change configurations while retaining your applications and files. Save the project and stop active processing before switching.

Finish with a simple check: import a sample clip, play it with audio, save the project, and export a short test. That confirms the production environment works before you introduce automation.

Turn the Brief and Media Folder into an Edit Plan

Before touching the timeline, give the assistant a clear delivery brief. For our example, that means a 60-second landscape product video, a 30-second vertical version, and an English caption file. Specify the audience, required product claims, approved branding, and anything that must not appear.

Then ask for a read-only media inventory. With access to tools such as ffprobe, the assistant can report technical properties without changing the files.

The report should distinguish facts from decisions:

Technical metadata does not describe what happens in a shot. A filename such as product_closeup_03.mp4 is not proof that the clip contains the intended detail. If the assistant has not inspected the footage through an appropriate tool, it should leave the content description unverified.

Source: Adobe Premiere Help

Build a Timecoded Selection List

Use a transcript generated from the actual recording to propose dialogue selections. In Premiere, Text-Based Editing provides a transcript-based route to working with source dialogue.

Ask the assistant to return each proposed selection with its source clip, in/out points, exact quoted wording, and editorial purpose. Keep source timecodes separate from proposed timeline positions, and confirm the timecode format.

Review those selections against the recording. Removing a sentence can change meaning even when the remaining words sound fluent.

The output of this stage is an approved assembly plan, not a finished cut. It should identify the opening, core demonstration, supporting shots, and closing message, with unresolved choices clearly marked.

Source: Adobe Premiere Help

Connect the Assistant to Your Editing Application

With the media inventory and edit plan ready, choose how the assistant will reach the editor. File access lets it prepare reports and scripts, but does not automatically expose the open project or timeline.

Use MCP for Supported Timeline Operations

An application-specific MCP connection typically follows this path:

Claude or Codex → MCP server → application bridge → video editor

The installation depends on the project you choose. For example, PremiereProMCP documents a local server and UXP panel, with Premiere 26.0 or newer among its requirements. It is a community integration, not an Adobe-supplied connection.

Run the editor, bridge, and AI client inside the same Vagon desktop for this setup. Follow the server’s instructions for your chosen client rather than copying a configuration intended for a different transport or application.

For Resolve, check the installed release, edition, and scripting requirements against the selected integration. Do not assume that a bridge documented for Studio works identically with the free edition.

Start with a read-only request:

Compare the response with the editor. A connected server is only useful if it can identify the correct project and expose the operations your plan requires.

Enable Computer Use for Interface Tasks

Set up Claude computer use or Codex computer use separately, checking account availability and application permissions. For this workflow, the desktop client should operate inside Vagon so it reaches the cloud editor.

On Windows, computer use operates in the foreground. Keep the target application visible and avoid competing mouse or keyboard input while the assistant works. Stop the task before taking over.

Use structured tools for supported, repeatable edits and computer use for tasks that require the interface. Neither connection removes the need to inspect the resulting sequence.

Build a Rough Cut Without Losing Editorial Control

The first assembly should turn your approved selections into an editable sequence, not commit every creative decision at once. Keep the source project intact and work in a clearly named copy, such as Product_Landscape_v01.

Approve the Selection Before Editing

Ask the assistant to present the proposed order before applying changes. Each entry should identify a source clip, source in/out points, intended timeline position, and purpose.

For our product video, the structure might be:

This is a proposed structure, not a reason to force every video into identical timing. Watch the selected footage and listen to the dialogue before approving the assembly.

If a transcript selection begins halfway through a thought, adjust it. If the requested demonstration shot is missing, flag the gap instead of substituting unrelated footage without explanation.

Apply the Plan to a New Sequence

Once the plan is approved, give the assistant a narrow execution request:

Use the connected server’s documented operations. A bridge may support importing media, placing clips, and trimming ranges while lacking another operation your plan needs. The Premiere MCP example documents several of those timeline functions, but its coverage should not be generalized to every integration.

If a required action is unavailable, complete it with the editor’s own tools or use an appropriately reviewed script. Do not let the assistant silently replace an unsupported edit with a different result.

Work in small batches. Assemble the dialogue first, inspect it, then add approved supporting shots. This makes an incorrect source range or placement easier to locate than a single instruction that changes the entire timeline.

Check the Cut in the Actual Editor

After each batch, compare the assistant’s change report with the sequence. Confirm the correct clips, edit points, track placement, and approximate duration.

Computer use can help navigate to named sequences, inspect markers, or capture particular timeline positions. It can also bring a questionable transition into view. Those actions support review, but screenshots do not establish smooth motion, clean sound, or correct synchronization.

Source: Adobe Premiere Help

Play the rough cut yourself with audio. Check for clipped words, repeated phrases, awkward pauses, unintended gaps, and B-roll that obscures an important demonstration. Confirm that shortening the dialogue has not changed the speaker’s meaning.

Record corrections against specific timeline positions rather than asking the assistant to “make it flow better.” For example: preserve the complete sentence at one cut, extend a supporting shot over another, or remove a repeated claim.

The rough cut is ready for the next stage when its story and timing are approved. Captions, graphics, and delivery versions can then build on a stable sequence instead of following edits that are still changing.

Create Captions and a Vertical Version from the Approved Cut

Once the main cut is approved, prepare the supporting deliverables from that sequence. Keep the landscape master intact so caption corrections and vertical adjustments do not unexpectedly change it.

Prepare Captions After the Main Edit

Generate captions from the edited dialogue, not from an untouched source transcript. Removed sentences and rearranged clips change timing. Adobe’s Text-Based Editing documentation distinguishes transcript editing from the captions workflow and recommends generating captions from the final edited sequence.

Claude or Codex can help review an exported caption file against approved terminology:

This is text review, not proof of synchronization. Play the sequence to confirm that captions follow the speech, remain readable, and do not cover important product details. Check speaker names, numbers, and technical terms against the recording.

Choose the required delivery type deliberately: a separate caption file, burned-in text, or supported embedded captions. These are different outputs, not interchangeable export settings.

Source: Adobe Premiere Help

Reframe for Vertical Delivery

Duplicate the approved sequence for the 30-second vertical version. Establish its dimensions and duration, then decide which message survives the shorter format. Simply cropping the 60-second master does not produce a coherent shorter edit.

Ask the assistant to propose the reduced selection before changing clips. Preserve a complete opening, useful demonstration, and clear closing rather than speeding up the entire video to fit.

Source: Adobe Premiere Help

For framing, inspect representative points throughout each shot:

Where supported, structured tools can apply approved position or scale changes. Computer use can help navigate to checkpoints and inspect the composition. Neither method should assume that one correctly framed screenshot represents the whole moving shot.

Finish by playing the vertical sequence from beginning to end. Its captions must match its own revised dialogue and timing, so generate or retime them after the shorter edit is settled. Save the vertical project changes and caption file as clearly labelled versions.

Handle Repetitive Production Tasks with Small AI Jobs

Beyond assembling the cut, Claude and Codex can help with production tasks that are easy to describe but tedious to repeat. Keep each request tied to a defined input and a reviewable output.

Turn Client Feedback into Actionable Notes

Give the assistant timecoded feedback and identify the exact review version. Ask it to separate editorial changes, graphics corrections, and questions needing clarification.

For example, “the logo appears too early at 00:18” can become a proposed task attached to that position. Do not apply markers from an older export to a revised timeline without checking the reference. If the connected MCP server supports markers, the approved notes can be added through those tools; otherwise, retain a separate revision list.

Source: Adobe Premiere Help

Check a Batch of Exports Against the Brief

Ask the assistant to inspect exported files with ffprobe and compare their properties with a delivery manifest.

This catches technical inconsistencies, not weak pacing, incorrect branding, or unreadable captions.

Prepare Language and Title Variants

Use approved copy to create a version checklist covering titles, captions, filenames, and destinations. Have translations reviewed before inserting them, and inspect longer text in the actual composition.

One assistant can perform all three jobs. If you switch between Claude and Codex, pass along the approved files and current project version rather than allowing overlapping edits.

Export, Inspect, and Package the Deliverables

An approved timeline is not yet a delivery file. Before exporting, match each sequence to the brief: the landscape master, shorter vertical version, and their corresponding captions.

Use a compact delivery manifest to make those requirements explicit:

Set codec, audio, and other technical requirements from the client’s specifications rather than asking the assistant to choose an unexplained “best quality” preset. In Premiere, inspect the relevant export settings, including whether captions are burned in, exported separately, or embedded in a supported format. Check whether proxy or preview options are appropriate for the final delivery.

If the connected tools support export, have the assistant report the selected sequence, destination, and settings before starting. Otherwise, configure the export in the editor. Adding a job to a queue does not establish that rendering finished successfully.

Source: Adobe Premiere Help

Check the Actual Output

Use the batch-checking process from the previous section to compare completed files with the manifest. Then open and play the exported videos, not just the timeline previews.

Listen for missing audio, clipped dialogue, and synchronization problems. Inspect framing, titles, caption timing, and the opening and closing frames. A technically valid file can still contain the wrong sequence or an outdated logo.

Keep playback-stream limitations separate from problems in the exported media. If a suspected glitch appears during remote review, inspect the downloaded output before concluding that the render is damaged.

Package a Usable Handoff

Place approved exports and matching caption files in clearly labelled delivery folders. Include the editable project and required linked assets when the handoff calls for them, subject to asset licensing.

Use Vagon Files to transfer the finished package. Confirm the delivered copies open correctly, preserve the working project, and stop the cloud computer after active exports and transfers have completed.

Source: Adobe Premiere Help

Final Thoughts

Claude and Codex become useful editing partners when their instructions lead to changes you can inspect in the actual project. Use them to prepare media, propose selections, perform supported operations, and check delivery files, while keeping editorial approval in your hands.

Vagon Cloud Computer brings the assistant, desktop editor, project assets, and processing resources into one workspace. You can develop the workflow from preparation through export without moving the production environment onto your everyday device.

Start with one assistant and one short project. Establish a reliable connection, preserve your originals, and expand the automation only after the saved project and exported results pass review.

FAQs

Can Claude and Codex edit videos directly?

They can perform supported operations through an application-specific MCP integration or approved computer-use tools. Their capabilities depend on the client, editor, permissions, and available tools. A normal chat does not automatically have access to your timeline.

Do I need both assistants?

No. Start with the assistant you can configure and supervise most easily. If you use both, work sequentially and pass along the current project version, approved edit plan, and outstanding changes. Avoid simultaneous control of the same editor.

Can I start without an MCP server?

Yes. With appropriate file and tool access, an assistant can prepare media inventories, review transcripts, draft commands, and check exported files. You can apply its approved edit plan manually. MCP is needed only when your chosen workflow relies on operations exposed by that server.

Is Claude Code the same as Claude Desktop?

No. Claude Desktop is the application containing several working modes, including a Code tab. Claude Code also runs through other clients, such as its terminal interface. Features, configuration, and computer-use access should be checked for the specific client you use.

Can both assistants control the editor at the same time?

Avoid it. Overlapping actions can change the active sequence, selection, or dialog while another operation is running. Stop one assistant before handing control to the other, and confirm the application’s current state.

Are editing software licenses and AI subscriptions included with Vagon?

Do not assume so. Preinstalled applications and license entitlements are separate. Complete the required software activation and AI sign-in, and check applicable subscriptions and remote-use terms. MCP bridges may require additional configuration.

Does a faster cloud GPU make the assistant itself faster?

Not necessarily. More suitable hardware can improve supported decoding, effects, playback, and export tasks. It does not directly accelerate a remotely hosted AI model. Benefits depend on the application, codec, workload, and selected configuration.

Can AI verify the final video without human review?

It can assist with technical checks and targeted inspection, but those checks do not establish editorial quality. Watch and listen to the actual exports, verify captions and claims, and approve the deliverables yourself.