FrameOS

FrameOS for AI agents

Turn your long videos into short clips by asking the AI agent you already use. Paste a podcast, stream, interview or webinar link into Claude, ChatGPT, Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI or Perplexity, and FrameOS finds the strongest moments, reframes them for your format, adds captions and hands you back a ranked list of clips. You stay in charge of what gets exported, captioned, thumbnailed and posted.

This repo holds the FrameOS plugin, skills and extension packages. They connect your agent to the hosted FrameOS connector (an MCP server at https://frameos.studio/mcp) and teach it how to use FrameOS well: when to check your credits, how long to wait for a render, which caption styles exist, and never to post anything without your go-ahead.

Status, 2026-10-26: the hosted connector is live, and every listed app except Devin can sign in. https://frameos.studio/mcp is up. Claude Code, Claude (web, Desktop, mobile), ChatGPT, Codex, Cursor and VS Code with GitHub Copilot are tested end to end against a real account; Gemini CLI, Perplexity and the Copilot CLI can sign in but are not tested end to end yet (beta). Devin (including Windsurf) is coming soon: it can't sign in yet. See docs/status.md for what is left, and try the whole workflow against the mock server without spending credits.

What you can say to your agent

You don't need to learn any commands. Talk to your agent the way you'd brief an editor:

• "Turn this episode into 5 vertical shorts for TikTok: https://youtube.com/watch?v=..."
• "Make 3 square clips for LinkedIn from this interview, focusing on what we say about pricing."
• "Show me my clips from yesterday's stream, best first."
• "Make the captions on clip 2 bigger, move them up a bit and use the Beasty style. Then give me the MP4."
• "Where in last week's podcast do we talk about hiring? Give me quotes with timestamps."
• "Make 2 YouTube thumbnails for my best clip."
• "Write an Instagram caption for that clip." Then, when you're happy: "Post it to my Instagram."
• "Turn this webinar into a week of content: 5 clips, captions, thumbnails and copy for YouTube and LinkedIn."
• "How many FrameOS credits do I have left, and what did I spend them on this month?"

More ready-to-use prompts, grouped by job, are in docs/workflows.md.

What happens when you ask for clips

1. Your agent checks you're signed in and have credits.
2. It sends the link (or uploads your file, if your agent can run commands) to FrameOS with your settings: how many clips, which format, and what to focus on.
3. A render usually takes 10 to 30 minutes. Your agent tells you roughly how long and checks back on its own, or, in chat apps that can't wait, tells you to ask "check my FrameOS render" later.
4. You get a ranked list: title or hook, length, score, where it sits in the original video, and a preview link.
5. You pick what happens next: restyle captions, export MP4s, make thumbnails, draft copy, or post.

How credits work

• It works on every FrameOS plan. All you need is credits.
• A render costs 1 credit per started minute of source video, and you're only charged when clips are delivered.
• Thumbnails cost 10 credits each.
• Exports, caption changes, copy drafts, collections and posting are free.
• A render only starts when your balance covers the whole video, so you never end up with a half-paid render.
• Up to 3 videos can be processing at once in a workspace. Your agent queues the rest and submits them as renders finish.
• The server enforces both of these rules (live since 2026-10-06), and the skills follow them too.
• Ask your agent "What is my FrameOS credit balance?" any time. Plans and prices live on https://frameos.studio/pricing; your agent won't quote them.

Install

Pick your app. Each guide has the exact steps, how to check it worked, how to sign in again, and known limits. Claude Code, Claude (web, Desktop, mobile), ChatGPT, Codex, Cursor and VS Code are tested end to end; Gemini CLI and Perplexity (since 2026-10-08) and the Copilot CLI (since 2026-10-23) can sign in but are not tested end to end yet; Devin is coming soon (status).

Check it worked. In a new conversation, ask: "What is my FrameOS credit balance?" You should get your workspace name and a credit number. A "connected" badge on its own isn't proof; only a real answer is.

Let your agent do the install. Agents that can run commands (Claude Code, Codex, Gemini CLI and similar) can install FrameOS for you. Paste this:

Install FrameOS by following https://github.com/theSatvik/frameos-plugins/blob/main/docs/install-protocol.md

What it can and can't do

Your agent can:

• Clip a video from a link - YouTube, Vimeo, Twitch, Kick, public Google Drive files or a direct video file URL - or from a file on your computer when your agent can run commands. Sources need to be at least 30 seconds long.
• Make up to 20 clips per video (3 by default). The number is a maximum, so you may get fewer.
• Frame clips as 9:16 (Shorts, Reels, TikTok), 4:5 or 3:4 (feeds), 1:1 (square) or 16:9 (landscape).
• Steer clipping toward a topic. Use the words the speaker actually says ("pricing", "hiring"), because the focus is matched on words; it's a preference, not a filter.
• Restyle captions with 22 caption styles (or none), 16 fonts, size, position and 17 animations, then export a captioned MP4. Download links last about an hour, and your agent fetches fresh ones when needed.
• Search the full transcript of videos processed since the connector launched (older projects may only have per-clip transcripts) and pull quotes with timestamps.
• Make up to 3 thumbnails per request from real frames of your video with text layout (not AI-generated images).
• Draft a title, caption and hashtags for YouTube, Instagram, Facebook, LinkedIn, TikTok or X.
• Post to YouTube (Shorts), a Facebook Page, an Instagram Business account (Reels) or a LinkedIn personal profile that you connected in FrameOS - only after you confirm each post.
• Organise clips into collections, duplicate clips to experiment, and show your credit usage for the last 30 days.

Not available through your agent (use the FrameOS web app at https://frameos.studio/dashboard, or it's not available yet):

• Scheduling posts, share links, censoring or bleeping, dubbing, or free-form timeline edits.
• Choosing the clip length, or rendering a whole video without clipping.
• Cancelling a render that has started.
• Posting to TikTok or X (your agent can still write the copy).
• Connecting social accounts or uploading a brand logo (both are done once, in the web app).
• Renaming or deleting collections, or removing a clip from one.

Clips made on the free plan carry a FrameOS watermark.

What this plugin sends

• Two addresses, nothing else. Once installed, the plugin points your app at the FrameOS connector, https://frameos.studio/mcp, and at FrameOS sign-in, https://clerk.frameos.studio, which you see once when you connect your account. Upload and download links that the connector hands back (for a file you ask your agent to upload, or a finished MP4) point at FrameOS's own file storage and expire.
• No code of its own. The plugin is written instructions (skills and Gemini CLI commands) and connection settings, with no hooks and nothing that runs when you install or use it. Your AI app makes every request. The repo's dev/, scripts/, tests/ and evals/ folders are maintainer tooling (a local mock server that runs only when you start it and calls nothing outside your machine, plus checks and tests); the plugin never runs them. In apps that can run commands, the skills may have your agent check and upload a video file you picked, wait between render checks, save an MP4 you asked for, or run your app's install steps when you ask it to set FrameOS up.
• No tracking. No analytics or telemetry. The only extra detail the plugin adds is its version number, sent as an X-FrameOS-Plugin-Version header in apps that support it.
• Nothing goes public without you. The skills ask you to confirm every post, every time, before anything is published.

What FrameOS does with your videos and account is in the privacy policy and the terms. Questions: support@frameos.studio.

Try it without credits (mock server)

The repo ships a local mock FrameOS server with the same 27 tools and inputs as the real one, and errors in the same format. It needs no account, no sign-in and no credits, so you can watch the whole flow - submit, poll, clips, captions, export, thumbnails, copy and a fake post - without touching a real account.

git clone https://github.com/theSatvik/frameos-plugins

cd frameos-plugins uv run --script dev/mock_server.py # stdio uv run --script dev/mock_server.py --http --port 8790 # Streamable HTTP at http://127.0.0.1:8790/mcp claude --plugin-dir . --mcp-config dev/mock.mcp.json --strict-mcp-config # Claude Code: these skills + the mock only

Mock links point at mock.frameos.invalid, so nothing real is created. dev/README.md has the full walkthrough and the inputs that trigger each error.

Packages in this repo

The repo root is the plugin. One set of skills serves every host; each host reads its own manifest.

| skills/ | The eight skills: frameos-setup, frameos-clip, [frameos-captions](skills/frameos-caption