Skip to main content
skill CX Codex

gemini-omni — Claude Code & Codex skill: video generation and editing

A skill for Claude Code and Codex that generates short clips via gemini-omni-1.1-flash and edits them through follow-up prompts without re-generating. Each generation returns an interaction_id so the next call changes only what you specify.

AGPL-3.0 ★ 78,316 updated 2026-10-24

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add calesthio/OpenMontage --skill gemini-omni -a claude-code
Open on GitHub

What gemini-omni does

The skill runs on Claude Code and Codex through the gemini_omni_video tool, which calls the Gemini Interactions API directly with no gateway in between. Output is 3–10 second clips at 720p with synthesized audio, on-screen text, and scenes timed via bracketed ranges like [0-3s].

Within the OpenMontage toolkit, the differentiator is stateful editing. A generation call returns interaction_id; passing it as previous_interaction_id on the next call patches the clip in place. Keep edit prompts short and surgical — a descriptive scene prompt rewrites the whole clip. Routes through fal.ai or Runway do not carry Google's interaction state, so conversational editing only works through the direct gemini_omni_video call.

The skill reads GEMINI_API_KEY or GOOGLE_API_KEY from the project environment, the same key used by google_imagen and google_tts. Output video is billed at $17.50 per million tokens, roughly $0.10 per second; when usage data is unavailable, the adapter falls back to a maximum 10-second estimate. Reference images go in reference_image_paths and are bound to roles in the prompt via <IMAGE_REF_0>, <IMAGE_REF_1>, etc.; <FIRST_FRAME> pins an image as the opening frame. No seed support means generations are not reproducible.

What it helps with

  • edit a generated clip in place without re-running the full generation
  • generate a clip with synthesized audio and timed on-screen text
  • bind a reference image to a character or style using prompt tags
  • schedule scene changes across a single prompt with timecodes
  • restyle or remove an object from an existing uploaded video

Example prompts for the agent

  • Generate a 5-second clip: tabby cat on a sunny windowsill, tail twitching slowly, soft background music
  • Remove the phone from her hand, keep everything else — pass previous_interaction_id from the last response
  • Start from <FIRST_FRAME> using hero.jpg: woman walking, [0-3s] she pauses, [3-6s] turns around

How to install gemini-omni

1Skills CLI (Vercel)

npx skills add calesthio/OpenMontage --skill gemini-omni -a claude-code

2Manually: put the files in place

~/.claude/skills/gemini-omni/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/gemini-omni/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about gemini-omni

How do I install gemini-omni in Claude Code?

Run npx skills add calesthio/OpenMontage --skill gemini-omni -a claude-code. The skill is then available at ~/.claude/skills/gemini-omni/SKILL.md.

What does the gemini-omni skill do?

It generates 3–10 second 720p video clips through gemini-omni-1.1-flash and lets you edit them conversationally. Each response carries an interaction_id that the next call uses to patch only the specified part of the clip.

Does gemini-omni need its own API key?

No separate key — it uses GEMINI_API_KEY or GOOGLE_API_KEY already in the project environment, the same credentials as google_imagen and google_tts. Access requires a paid Gemini API tier.

Similar skills

All skills for AI and ML →

From the same set101

The whole set →
hyperframes-cli hyperframes-cli guides the HyperFrames video build loop, from composition validation and timeline inspection to AWS Lambda rendering. CX Codex +1 ★ 78,316repo stars synthetic-screen-recording The synthetic-screen-recording skill guides synthetic terminal demos in Remotion with TerminalScene. Use it when the command sequence and output can be authored ahead of time. CX Codex +1 ★ 78,316repo stars svg-path-draw svg-path-draw — a rule that gets the agent to trace an SVG outline stroke by stroke, then fade in a wordmark once the drawing settles. CX Codex +1 ★ 78,316repo stars hacker-flip-3d hacker-flip-3d supplies guidelines for creating a 3D character flip animation with randomized glyph decoding. Helps build terminal and airport-style text reveals. CX Codex +1 ★ 78,316repo stars discrete-text-sequence discrete-text-sequence — a rule for GSAP scenes where the terminal line swaps whole states instead of typing evenly: typos, backspaces, pasted chunks and holds. CX Codex +1 ★ 78,316repo stars animation-groups The animation-groups rule organizes ManimGL animations for parallel, sequential, or staggered playback. Use it when several scene elements need coordinated timing. CX Codex +1 ★ 78,316repo stars transform-animations transform-animations — a rule that teaches Claude Code and Codex how to transform objects in ManimGL — from basic shape morphing to complex function-based transformations. Use it when writing Python animations of mathematical objects. CX Codex +1 ★ 78,316repo stars embedding embedding instructs your coding agent to debug ManimGL scenes interactively via IPython, letting you test object positions and animations without restarting the render. CX Codex +1 ★ 78,316repo stars

Project sponsors

Companies that keep this analytics open