OpenMontage: Turn Your Coding Agent into a Video Workflow

10 min read · 2 Nov 2026
Last reviewed November 2, 2026

A product video needs an accurate script, useful footage, readable captions and an edit someone can approve. Generating an attractive clip solves one part of that job.

OpenMontage is an open-source AI video framework that gives a coding assistant tools and production workflows for research, scripting, media creation, editing and rendering. The canonical calesthio/OpenMontage repository supports assistants including Claude Code, Cursor and Codex. You direct the production through reviewable intermediate files.

Source review: 2 November 2026. We inspected documentation and source at commit 195eab2, including budget and validation call paths. We did not run a video production or benchmark output quality. The proposed pilot and cost example are illustrative.

What is OpenMontage, and how is it different from Studio?

OpenMontage on GitHub is a developer toolkit you run with a coding assistant. OpenMontage Studio is a separate desktop product from the same team. The official Studio facts page describes a Windows private alpha, with a waitlist for beta access. The GitHub toolkit has its own license.

Studio’s terms dated 30 October 2026 exclude the EU, EEA, UK and Turkey, among other territories including mainland China, Hong Kong and Macau. Those are restrictions on Studio Services, not a territorial clause in the repository’s AGPL license. Check each provider’s eligibility separately.

How does the video production workflow work?

OpenMontage uses your coding assistant to coordinate production. The architecture documentation describes YAML manifests for stage order and Markdown skills for production instructions. The assistant follows those instructions and calls Python tools; there is no separate Python orchestrator.

The animated-explainer manifest illustrates the handoffs below. The review questions are ours; other pipelines use different stages.

From a production brief to an accepted export
Stage	What becomes reviewable	Question for your team
Research and proposal	Sources, audience, concept and proposed production route.	Is this the message the audience needs?
Script and scene plan	Narration, on-screen claims and the planned sequence.	Are the facts correct, and can each scene be made with the approved tools?
Assets	Footage, images, voice, music and their scene assignments.	Do the actual takes match the plan and permitted sources?
Edit and composition	Timing, captions, audio mix and rendered video.	Does the complete video communicate clearly?
Export	Delivery files and production records.	Is the version ready for the intended audience?

Backlot displays production state in a local, read-only board derived from project files. Approvals still pass through the agent workflow. When requesting a revision, you can point to the script or scene that needs changing.

What kinds of videos can OpenMontage make?

OpenMontage supports generated video and editing existing material. Choose a pipeline for the deliverable. A product demonstration needs accurate screen material; a podcast clip needs a faithful excerpt.

Three practical starting points, with our suggested review focus
Deliverable	Repository workflow	Review before accepting
Product walkthrough	The screen-demo pipeline organizes a demonstration around screen material.	Does the recorded interface show the user task clearly? Are labels readable?
Documentary or thematic montage	The documentary-montage pipeline retrieves and arranges real motion footage.	Rights to the footage and whether it supports the narration in context.
Podcast clips	The podcast-repurpose pipeline works from source audio or video to prepare caption-led and quote-led content.	A faithful excerpt with enough context to identify the speaker and understand the statement.

For approved business metrics, our Claude Dashboards and Motion workflow covers data verification and animated reporting. The MoneyPrinterTurbo cost and setup review covers a narrower short-video tool.

How do you install OpenMontage from GitHub?

Inspect the canonical repository’s setup, then run installation and preflight to check which tools are available. The README lists Python 3.10+, FFmpeg, Node.js 18+ and a coding assistant as minimum prerequisites. Treat those as project minimums; choose supported versions compatible with your renderer.

The following commands pin the source reviewed here. The Makefile installs Python and rendering dependencies, attempts optional Piper and HyperFrames setup, and provides the preflight command. Installation needs network access.

git clone https://github.com/calesthio/OpenMontage.git
cd OpenMontage
git checkout 195eab24478dc9363326aec437aefc3701635ddf
make setup
make preflight

Use preflight to scope the brief. The tool registry reports provider availability with dependencies and runtime warnings. A listed integration may still need a binary or model weights. Credentials do not establish model access. Configure the route you intend to use.

To inspect the supplied no-key examples, use make demo-list; make demo renders their Remotion compositions. Those examples use local animation; media-generation providers remain separate. Check the renderer’s license before using it for your organization.

Is OpenMontage free, and what runs locally?

OpenMontage’s source can be obtained without a software purchase, but a finished production has several cost layers. The provider guide distinguishes local execution from cloud APIs and identifies GPU-dependent tools. Local rendering can avoid paid media-generation calls; it does not remove assistant usage, hardware, licensing or editorial work.

A local project directory does not establish an offline workflow. Your coding assistant and selected cloud media providers may receive prompts or supplied material. Decide which assets each service may process before connecting it.

Calculate cost per approved video across the batch, including failed takes:

Cost per approved video =
  (media generation + assistant usage + compute and license allocation
   + human production and review time)
  / number of accepted, publishable videos

For an illustrative batch, assume €24 in media calls, €8 in assistant usage, €8 allocated to compute and licenses, and two hours of human work at €60 per hour. Total cost is €160. Four accepted clips cost €40 each; only two accepted clips raise that to €80 each. With none accepted, there is no meaningful cost-per-approved-video ratio.

These are assumed inputs. Allocate setup and maintenance over the volume they serve; count each retry once. Our agent cost-per-action model explains why the denominator changes an automation decision.

Do OpenMontage’s budget limits stop every paid call?

The reviewed version does not automatically enforce a global spending cap around every paid tool call. Its configuration defaults to mode: warn, a US$10 budget, a 10% reserve and a US$0.50 action-approval threshold. Those settings alone do not block a paid API call.

The CostTracker implementation can reject an excessive reservation in cap mode. However, our search found tracker instances in tests, but no production Python caller constructing one. The shared tool execution wrapper adds caching and activity events, without a budget gate. Issue #723 reports this integration gap and remained open when checked.

An agent or integrator can explicitly use the tracker. Before running a batch, verify that each paid operation reserves an estimate before execution and records its actual cost afterward. Test rejection with a mocked provider so the rejected attempt cannot incur a charge. Add provider-side spending limits where available. Changing warn to cap in YAML alone is insufficient evidence that this works.

What do the quality gates check?

OpenMontage contains useful validation tools and approval checks, but their coverage depends on the production path. A checkpoint records workflow state. Render checks catch technical problems, while a reviewer must still judge whether the finished video is accurate and communicates clearly.

The composition validator checks missing files and invalid cut durations. It also flags narration that extends beyond the video. The agent must invoke it as prescribed. Separately, the standard video render path checks the planned renderer and selected visual requirements; other rendering entry points follow different paths.

The checkpoint writer rejects completion of a gated stage without a recorded human-approval flag. That flag is supplied by the caller. The agent guide describes self-review as advisory and allows progression with warnings after two review rounds. Assign a reviewer to watch the whole export before release.

Can you use OpenMontage commercially?

OpenMontage uses AGPLv3, which permits commercial use subject to its obligations. Internal production and distributing or operating modified software for remote users raise different questions. Section 13 covers source offers for modified network-interactive versions. A rendered video does not automatically inherit AGPL merely because the software produced it.

Check the rest of the stack too. Remotion has separate license terms, including company eligibility rules. Media providers, models, music, footage and voices also need terms appropriate to your intended use. Keep asset provenance with the project; a reference video’s public URL does not grant reuse rights.

A proposed pilot: one 60-second product update

Start with one product task shown in owned screen recordings, using approved claims. The 60-second scope is our choice for this example. Set acceptance criteria and the export format for your audience before generation. Approve a representative sample before completing the film.

Adapt this brief after checking available providers:

Prepare a 60-second product-update video for existing customers.
Use only our supplied screen recordings, logo and approved release notes.
Show one user task from start to finish. Keep interface text exact.
Use a 16:9 master with readable captions and a text transcript.
Do not invent screens, metrics, testimonials or feature behavior.
Propose the script, scene plan, renderer and provider route first.
Show a representative 10-second sample for approval before full production.
Ask before any paid call or change of provider.
Keep asset sources, approved versions and actual costs with the project.
Export a review copy and the files needed to revise it.

The prompt requests spending approval. Verify the execution controls too.

Proposed acceptance checks for the product-update pilot
Check	Pass condition
Product truth	Every visible feature and factual claim matches the supplied recording or approved notes.
Comprehension	A reviewer can identify the user task and next action without the production brief.
Media quality	Watch the whole export with sound. Captions remain readable; narration is complete; transitions contain no unintended gaps.
Rights and routing	Every asset has a recorded source and permitted use; the run uses the agreed providers.
Cost	Record paid attempts, assistant usage, rendering allocation and human time, including rejected takes.
Revision	Change one approved feature label in a test copy. Update every affected scene and retain an identifiable final version.

Run the same acceptance checks against your current production method. Compare total effort per accepted video, including correction time. If generation gets faster but review takes longer, narrow the scope before adding volume.

When would we choose OpenMontage?

We would pilot OpenMontage for recurring video work with a technical owner who can maintain the workflow and review each release. The team can refine the saved production steps over successive runs. For a one-off edit, the team’s existing editor may be sufficient.

Once accepted output becomes repeatable, expand to another format or language. Wavect’s AI enablement service can help implement the workflow and evaluate it against your team’s acceptance criteria.

OpenMontage FAQ

Is OpenMontage a new AI video model?

OpenMontage coordinates media tools through production instructions. Your coding assistant supplies the reasoning, while the selected providers create or process media. Judge a production by the selected model and renderer, using representative source assets. Check the result and its cost on that combination; the framework name cannot establish either.

Do I need a GPU to use OpenMontage?

The requirement depends on the chosen tools. Local video-generation models can need substantial GPU resources; cloud generation shifts that work to a provider, while editing and rendering have their own requirements. Run preflight on your intended setup and try a representative sample before buying hardware. A successful simple animation does not establish capacity for a different model.

Can I start from a reference video?

The project supports reference-driven production for compatible workflows: analyze the reference and develop a differentiated concept from its structure or pacing. Check the selected pipeline’s support rather than assuming every pipeline accepts the same inputs. Specify which creative aspects matter, and establish permission separately before reusing footage, music, a recognizable person or a voice.

Can I continue a production after stopping the agent?

OpenMontage records stage state and artifacts that can help a later session continue the work. Resume behavior still depends on the selected workflow and any provider jobs already in progress. Inspect those records before retrying a paid generation. Resuming a conversation and avoiding a duplicate paid request are separate checks for your pilot.

Can I edit the result after it has rendered?

The final MP4 is an export. Revisions depend on retaining the scripts, source assets, scene plan and renderer project that produced it. Ask for those files in the brief and test one controlled change before committing to the workflow. Do not assume every production route exports a timeline for your preferred desktop editor.

How should I review videos in several languages?

Approve the factual source script first, then review each language’s narration, terminology, captions and timing. Translated speech can change scene length, and a technically valid subtitle can still misstate a feature. Assign someone who understands the target language to acceptance. Account for that review in the cost per approved language version.