OpenMontage

What It Does

OpenMontage is an open-source agentic video production system that turns an AI coding assistant (Claude Code, Cursor, Copilot, Windsurf, or Codex) into a video studio. The agent picks one of 12 production pipelines (explainer, documentary montage, cinematic, clip factory, localization dub, and others), then walks a structured flow of research, proposal, script, scene plan, assets, edit, and composition, invoking 121 registered tools across 66 providers. It ships as a repo you clone: Python tooling, a Remotion (React) and HyperFrames (HTML/GSAP) composition runtime, a local FastAPI storyboard board called Backlot, and an instruction layer of 89 agent-skills plus hundreds of production-knowledge markdown files. The target user is a developer or content creator who wants prompt-driven video production with per-stage approval gates and cost governance, not a one-click web video tool.

The Good

The test suite is the strongest GearScope has seen in an agent-skill repo: 1,393 tests pass, 0 fail. The suite (100 test files across tools, contracts, backlot, lib, and qa) runs in about 90 seconds. Beyond unit coverage, tests/contracts/test_agent_instruction_integrity.py validates the markdown instruction files themselves: it asserts that pipeline director documents reference real capabilities such as get_by_capability("music_library") and that they do not mention fictitious functions like handle_explainer_. The repo also runs a session network guard with meta-tests (tests/test_network_guard.py) proving raw socket connects to provider APIs are blocked during tests, so the suite cannot spend money by accident.

The zero-API-key claim is real, verified with three rendered videos. The README promises a free path (Piper TTS offline narration, Remotion composition, FFmpeg post). Running render_demo.py produced code-to-screen.mp4 (25.0s, 3.7 MB), focusflow-pitch.mp4 (22.5s, 4.2 MB), and world-in-numbers.mp4 (23.1s, 4.3 MB), each ffprobe-verified as 1920x1080 H.264 with AAC audio. The registry reported 37 tools available with no keys configured, rising to 38 after installing piper-tts, which unlocks free offline narration.

Every capability claim checked out against the machine-readable registry. The README's own verification command (registry.discover() followed by support_envelope() and provider_menu()) reports 121 tools, each with a 33-field record covering status, runtime (local vs api), provider, stability, and determinism. That confirms the "100+ production tools" claim (121) and the "60+ provider integrations" claim (66 distinct providers, from ffmpeg and piper to fal, Kling, HeyGen, Runway, and Volcengine Ark). The "700+ agent skill and production-knowledge files" claim also holds: 750 unique-content markdown files across the .agents tree, root skills/, and docs, before counting platform duplicates.

The skill layer is deep, valid, and progressively disclosed. The .agents/skills tree follows the Agent Skills open standard with 89 SKILL.md files (89 of 89 with valid name and description frontmatter, descriptions within the 1024-character limit), a median of 138 lines and a maximum of 866, backed by 487 supporting reference files. Pipeline director skills are substantial documents of 400 to 560 lines each (explainer/proposal-director.md is 557 lines), and the whole pipeline-skill layer totals 15,449 lines.

The Bad

The headline agentic loop was not verifiable in review. The flagship showcase videos (a $1.33 animated short, a $4 transformation film) require a live coding-agent session plus paid provider keys for image and video generation. The demos we ran are curated Remotion component renders, not prompt-driven productions. The pipeline machinery behind them (stage directors, approval gates, cost tracker) is covered by tests, but a reviewer without paid keys cannot reproduce the marketed outcome end to end.

No releases, no tags, no changelog at 48K stars. The GitHub API returns zero releases. Version pinning is impossible; consumers ride main and get whatever landed that morning (the repo pushed the day of this review). For a system that orchestrates paid API spend, the absence of a stable version story is a real operational risk.

The README pipeline table lists 11 of 12 pipelines. The repo description says "12 production pipelines" and skills/pipelines/ contains 12 directories, but the README table omits character-animation. The pipeline is documented in body prose, so this is table drift rather than a missing feature, but it is the kind of count mismatch a careful reader trips over.

Setup carries real friction: Python 3.10+, a 170 MB clone, and a slow first render. macOS ships Python 3.9.6, which does not meet the stated requirement; we had to invoke python3.11 explicitly for the venv. The clone with history is 170 MB across 2,103 files, the Remotion composer adds 199 npm packages, and the three zero-key demos took roughly 5 to 7 minutes of render time on an Apple Silicon host.

No documented Hermes or OpenClaw install path. The skills carry metadata.openclaw frontmatter and live in the standard .agents/skills layout that universal installers understand, but the README names only Claude Code, Cursor, Copilot, Windsurf, and Codex. OpenClaw gets a joke section header, not install instructions.

Smoke Test Results

Host-based hands-on testing on macOS (aarch64) with Python 3.11.11, Node 24.13.1, and FFmpeg 8.0.1. The sbx microVM daemon was unavailable, so runs were executed directly on the host with logs captured to the sandbox convention (pitfall: host fallback documented in our methodology).

Run A. Fresh environment, no deps preinstalled

$ python3 --version
❌ Python 3.9.6 (macOS system) is below the required 3.10+; must use an explicit python3.11
$ /Users/openclaw/.local/bin/python3.11 -m venv /tmp/om-venv-fresh
✅ venv created
$ /tmp/om-venv-fresh/bin/pip install -r requirements.txt
✅ clean install of the 16-line requirements file
$ python /tmp/om-fresh-check.py (imports tools.tool_registry, discovers)
✅ 121 tools discovered with zero API keys configured
$ ffmpeg -version && node --version
✅ ffmpeg 8.0.1 and node 24.13.1 present

Pass rate: 4 of 5. The single failure is the host's own system Python, not the repo; with Python 3.11 everything installs and the registry imports on the first try.

Full sandbox log →

Run B. Host with deps preinstalled

$ /tmp/om-venv/bin/pip install piper-tts
✅ installed; zero-key available tool count rises 37 to 38 (offline TTS unlocked)
$ cd remotion-composer && npm install
✅ 199 packages in 12s, no errors
$ python om-smoke-registry3.py (README capability-envelope commands)
✅ 121 tools, 66 distinct providers, 38 available at zero keys, all claims verified
$ python om-structure.py (structural validation)
✅ 89/89 frontmatter valid, 12 pipeline dirs, LICENSE present, SKILL.md bodies 39/138/866 lines min/median/max
$ python om-count.py (700+ file claim check)
✅ 750 unique-content skill and knowledge markdown files (claim holds)

Pass rate: 5 of 5. This is the canonical install-and-structure verdict: every dependency resolves and every headline number in the README is confirmed by the repo's own tooling.

Full sandbox log →

Run C. Functional verification

$ /tmp/om-venv/bin/python -m pytest tests/ -q
✅ 1393 passed, 11 skipped, 0 failed in 90s
$ python render_demo.py code-to-screen
✅ 1920x1080 H.264+AAC mp4, 25.0s, 3.7 MB
$ python render_demo.py focusflow-pitch
✅ 1920x1080 H.264+AAC mp4, 22.5s, 4.2 MB
$ python render_demo.py world-in-numbers
✅ 1920x1080 H.264+AAC mp4, 23.1s, 4.3 MB
$ registry.provider_menu()["image_generation"] with zero keys set (negative case)
✅ graceful degradation: available is an empty list and every unavailable provider carries explicit install_instructions

Functional pass rate: 5 of 5. Positive cases assert on real rendered output (ffprobe stream data, not exit codes); the negative case proves missing keys degrade into per-provider guidance instead of crashes.

Full sandbox log →

What the runs tell you

The executable layer installs cleanly once the Python version is right, the registry and skill structure match every claim the README makes, and the zero-key path produces real video files. What remains untested is the last mile: a full agent-driven production with paid generation providers, which needs a live session and a budget.

Setup Walkthrough

1. Clone the repo: git clone https://github.com/calesthio/OpenMontage.git (170 MB with history, 2,103 files).
2. Run make setup. This creates the venv, installs the 16-line requirements.txt, runs npm install inside remotion-composer (199 packages), installs piper-tts for free offline narration, and warms the HyperFrames npx cache.
3. On macOS, confirm your Python meets the 3.10+ floor before step 2; the system 3.9.6 fails. We used an explicit python3.11 venv.
4. Copy .env.example to .env and add whichever provider keys you have. Every key is optional; more keys unlock more of the 84 key-gated tools.
5. Open the repo in your coding agent and give it a prompt from PROMPT_GALLERY.md, or run make demo to render the three zero-key demo videos immediately.
6. Budget time for the first render: Remotion bundles and renders frames headlessly, roughly 2 minutes per demo on an Apple Silicon machine.

Alternatives

1. remotion-dev/skills - the official Remotion agent-skills pack covers React-based video composition alone, MIT licensed, if you do not need the pipeline and provider-routing machinery.
2. ningzimu/codex-ppt-skill - image-based deck generation for slides rather than video, cheaper per output, when a presentation is the actual deliverable.
3. ConardLi/garden-skills (web-video-presentation) - a template-driven video presentation skill with TTS, far smaller scope, when you want one branded format instead of twelve pipelines.