🛠️ OpenMontage Turned Claude Into a Video Studio

The government pulled Fable 5. That story is still moving, I covered the arc two issues ago, but the thing that actually changed my week was a GitHub repo with 8,358 stars that dropped while everyone was watching the policy drama.

OpenMontage describes itself as the world's first open-source agentic video production system: 12 pipelines, 52 tools, 500+ agent skills. You point your AI coding assistant at it and it becomes a full video production studio. That claim is big enough to earn a raised eyebrow, so I dug in. The architecture holds up. The skill count is real. This is not a wrapper on top of a wrapper, it's a production graph with real tool calls.

Pair that with palmier-pro, a macOS video editor built specifically for AI workflows, 4,814 stars, and you have a week where the open-source video stack for operators quietly leapfrogged a lot of expensive subscriptions.

Meanwhile, jeremylongshore/claude-code-plugins-plus-skills landed with 2,404 stars and a number I had to read twice: 425 plugins, 2,810 skills, and 200 agents for Claude Code, with a proper CLI package manager. That is not a curated list. That is an ecosystem.

The through-line this week: the Claude Code skill layer is compounding faster than anyone expected, the video production stack just went open-source, and the policy ceiling on the most capable models is lower than it was a month ago. Three things shifted. All three matter.

Let's get into it.

[Skill] jeremylongshore/claude-code-plugins-plus-skills, 2,404 stars. 425 plugins, 2,810 skills, and 200 agents for Claude Code, with a ccpi CLI package manager and an open marketplace at tonsofskills.com. This is the closest thing Claude Code has to an npm ecosystem right now.

[Skill] nicobailon/visual-explainer, 8,824 stars. An agent skill that generates rich HTML pages and slide decks for diagrams, diff reviews, plan audits, data tables, and project recaps. Stop pasting markdown walls into docs; let Claude build the visual.

[Skill] op7418/guizang-ppt-skill, 18,303 stars. An AI-agent skill for generating polished HTML slide decks with editorial magazine and Swiss layouts, image prompts, social covers, and a WebGL presentation runtime. The most-starred skill in this week's shortlist for a reason.

[Skill] JimLiu/baoyu-design, 1,709 stars. Runs Claude Design locally as an agent skill inside Cursor or Claude Code. Produces polished UI mockups, prototypes, decks, and wireframes as self-contained HTML, no claude.ai/design required.

[Repo] calesthio/OpenMontage, 8,358 stars. World's first open-source agentic video production system: 12 pipelines, 52 tools, 500+ agent skills. Turns your AI coding assistant into a full video studio.

[Repo] palmier-io/palmier-pro, 4,814 stars. A macOS video editor built specifically for AI workflows. Pairs directly with OpenMontage for a full local production stack.

[Repo] andrepimenta/claude-code-chat, 1,057 stars. A polished VS Code chat interface for Claude Code. Not the terminal; a proper chat UI, inside your editor.

[Repo] livekit/agents, 11,075 stars. The go-to framework for building real-time voice AI agents. If you are wiring speech into an agent pipeline, this is where you start.

[Repo] Wan-Video/Wan2.1, 16,300 stars. Open large-scale video generative models. The open-weight answer to Sora, and it's further along than most operators realize.

[Repo] fishaudio/fish-speech, 30,880 stars. State-of-the-art open-source TTS. Fast, high-quality, self-hostable. The default pick for any voice layer you don't want to pay per-character for.

[MCP] Sendmux/sendmux-sdk, Official monorepo of SDKs, CLI, and MCP servers for Sendmux email APIs across TypeScript, Python, Go, PHP, Rust, and Ruby. Wire transactional email into a Claude agent without building your own API glue, the MCP server handles auth and dispatch.

Want to run a setup like this yourself? aigent-OS ($49) →

- Fable 5 stays pulled. The US government's ban on Anthropic's most advanced model remains in effect, and the operator question is now practical: which tasks were you routing to Fable 5 that need a reroute? The policy fallout is still live, TechCrunch lays out who fills the gap.

- MCP's real value is auth isolation. Simon Willison quoted Sean Lynch this week on why MCP beats skills and CLI for serious agent work: it isolates the auth flow outside the agent's context window entirely. That single architectural property changes the security posture of anything you wire up. The full quote is short and worth reading.

- AI sprawl is now a named problem. Business Insider ran a piece on what happens when teams vibe-code their way into a hundred overlapping tools with no coherent stack. The diagnosis is accurate. A fragmented tool layer is a cost center with a complexity tax. This is why operators who build coherent stacks are compounding; everyone else is whack-a-mole.

- The Pragmatic Engineer on the Fable ban. Gergely's breakdown covers the engineering and policy implications in more depth than anyone else this week. If you're building on Claude and didn't read this, read it before Monday is over.

This week: Custom slash commands. One command for every workflow you repeat.

Most operators run the same 4-5 multi-step workflows inside Claude Code repeatedly, kick off a PR review, scaffold a new module, run a preflight check before deploy. Typing out the same instructions every session is friction that compounds. Custom slash commands kill it.

1. Create a .claude/commands/ directory at the project root (or ~/.claude/commands/ for user-global commands). Add a markdown file named after your command, e.g. preflight.md. 2. Write the instructions in that file exactly as you'd type them to Claude. Use $ARGUMENTS anywhere you want the user's input injected: Run a preflight check on $ARGUMENTS before deploy, lint, type-check, confirm no secrets in diff. 3. In-session, invoke it with /preflight src/api/auth.ts. Claude picks up the full instructions from the file, injects your argument, and runs.

You'll know it worked when a workflow you used to type out in 80 words runs in one command, and when a new team member uses it correctly without you explaining anything.

Claude Code docs: slash-commands

Move: Pipe WhisperX into a Claude summarizer for async meeting intel.

Most operators are sitting on hours of recorded calls, demos, and standups with no indexed knowledge inside them. This move turns any audio file into searchable, queryable context in under 15 minutes.

1. Clone WhisperX and run whisperx your_audio.mp3 --output_format json, you get a JSON transcript with word-level timestamps and speaker diarization. 2. Pipe the JSON output into a Claude prompt: Given this transcript JSON, extract: action items with owner and deadline, key decisions made, open questions unresolved, and a 3-sentence summary. 3. Write the output to a markdown file named by meeting date and topic. Commit it to your repo's /docs/meetings/ directory so it's searchable with standard tools.

You'll know it worked when you grep a question from three weeks ago and get the exact meeting, speaker, and timestamp in under five seconds, without watching a single recording.

This week's featured pick is aigent-OS, a way to run Claude Code enhanced. Stock Claude Code writes the code. aigent-OS turns it into an operator: a caddy that surfaces the right skill the moment your task needs it, a learning loop that turns every repeated mistake into a guardrail so the same wrong turn is far less likely to cost you twice, and a routed crew of specialists you own. Same model, operator-grade output. At $49 founding it earns its cost the first week. See this week's featured drop in the store.

The open-source video stack, a 2,810-skill Claude Code library, and a government ban still reshaping which models you can route to, that is the week in one breath. The operators who adapt the stack first when the rules shift are the ones who compound. You're already reading.

See you Tuesday.