I6 · Open-source video pipeline + hosted tier — deep dive

Researched 2026-10-20 from GitHub, skills.sh, MCP registries, npm, Hacker News and primary policy pages. Estimates are labelled.

Bottom line

• Treat I6 as a cheap distribution asset with a small chance of a breakout, not as a revenue track. Of about 10,100 GitHub repos created in 2026 that mention "video" and "skill", 1.9% reached 100 stars and 0.3% reached 1,000 (GitHub search). The median 2026 Show HN post with "video" in the title got 2 points (HN API).
• Don't open-source a generic renderer or shorts generator. HeyGen gives away HyperFrames (67.8k stars, 2.7M npm downloads a month, in Anthropic's plugin directory), which already makes beat-synced lyric videos and sells cloud renders per credit (GitHub). None of 902 "shorts generator" repos created in 2026 passed 1,000 stars.
• Build for agents, sell to people. The 2026 video breakouts are agent tools: video-use (29.4k stars), video-shotcraft (12.6k) and ffmpeg-skill (2k in 8 weeks), and video skills hold several top-50 spots on skills.sh. But only people star and pay; pay-per-call video servers for agents show tiny usage and bill through crypto or Stripe, both closed to you.
• Lead with two pieces nobody else ships. video-dna turns a creator's or format's videos into a reusable style file. shorts-qa is a video checker that uses no AI model. Both run on the user's own agent and keys, so the Claude billing rule never comes up.
• Realistic first 90 days (estimate). Base case: 100–300 stars and $0. Good case (roughly the top 3%): 1,000–2,500 stars and $50–150 a month. The best solo precedent, Postiz, had ~12.8k stars and $700 a month two months in and $14k by month 14 (self-reported — DEV). $10k needs a breakout plus a year.
• Paddle's rules shape the hosted tier. Its policy (updated 13 April 2026) bans face swaps, deepfakes, voice impersonation, "automated social media marketing", downloaders and "stored value", and it restricts human-like faces (Paddle). Sell a monthly plan of render and check minutes with no AI model, at $15–49.
• Replace the 30-day kill test. 1,000 stars in 30 days is roughly a top-0.3–1% result, so use the day-30, 60 and 90 checks below. Your time: about one working day once, near zero after that.

The scenario, week by week

What you'd launch. Two small repos from code you own, each a one-line install into Claude Code, Codex or Cursor:

• video-dna: point your agent at 5–20 videos of one format; get a style file and generator-ready prompts.
• shorts-qa: one command scores any vertical video for pace, dead air, loudness and caption placement, as JSON.

The hosted tier waits for demand. Open the mockup for the README hero and proposed pricing.

gantt
    title I6 launch scenario from 2026-10-21 (estimates)
    dateFormat YYYY-MM-DD
    axisFormat %d %b
    section Build
    Extract and scrub owned code      :b1, 2026-10-21, 7d
    Paddle KYC and domain review      :b2, 2026-10-21, 14d
    section Launch
    Publish video-dna and shorts-qa   :l1, 2026-10-28, 4d
    Listings and launch posts         :l2, 2026-11-01, 3d
    section Learn
    Issues and weekly releases        :r1, 2026-11-04, 16d
    Day 30 signal check               :milestone, m1, 2026-11-20, 0d
    Blocks and lyric-sync             :r2, 2026-11-21, 28d
    Day 60 hosted decision            :milestone, m2, 2026-12-20, 0d
    section Sell
    Hosted beta if earned             :h1, 2026-12-21, 29d
    Day 90 revenue check              :milestone, m3, 2027-01-19, 0d

Week 0 (21–27 Oct): build. The AI extracts the code from MontageRnD, AdFactory and CreativeCenter and strips out Vortex art, the private Kaye material and the non-commercial MMS_FA aligner (CC-BY-NC). It writes the skill files, a command-line tool with machine-readable output, and READMEs with demo GIFs from your own footage. You start Paddle verification and domain review now (5–7 business days) and approve MIT for the skills; the hosted service stays closed.

Week 1 (28 Oct–3 Nov): publish. The repos go public. The AI gets them listed on skills.sh (a skill appears once people install it), Anthropic's plugin directory (315 plugins, none for video DNA or checking), the official MCP registry, Smithery and the big "awesome" lists. You post once to Show HN and 2–3 subreddits. Base case: 10–60 stars this week; a 6-follower GitHub account amplifies nothing. On the HN front page, AI repos gain on average 121 stars in 24 hours and 289 in a week (arXiv).

Weeks 2–4 (4–20 Nov): iterate. A release and a short post every week, Postiz-style ("Every minor feature was a version"). Base case: 0–3 issues, answered by the AI within hours. Day 30: 50–150 stars, 150–500 installs a month. If it catches: 1,000–1,500 stars, like ffmpeg-skill (1,432 by week 4) and openfilm (1,338 by week 3) (star history).

Weeks 5–8 (21 Nov–18 Dec): buy more tickets. Publish HyperFrames add-on blocks and lyric-sync. Four or five small repos beat polishing one. A "Run it in the cloud" waitlist button on every README measures hosted demand for free.

Weeks 9–12 (19 Dec–19 Jan): hosted beta, only if earned. Base case: no hosted tier; the repos stay as tooling for I3, I5 and I10 and as public proof of your work. Good case: a Paddle beta at $15 and $49 with 4–10 paying by day 90 ($50–150 a month), close to Postiz's early ~$55 a month per 1,000 stars.

Months 4–12: revenue follows automation users, not stars. Postiz went from $700 to $6.5k a month between months 2 and 10 while stars grew only ~65% (12.8k to 21k), after shipping an n8n node and a public API (star history). Your equivalent is an agent API key and an MCP server. Estimate: $10k a month needs 15,000–25,000 stars' worth of reach and a year.

Funnel, cumulative to day 90 (all estimates)

What comparable repos show

Stars from the GitHub API on 2026-10-20; early pace from OSS Insight, which matches GitHub within 3% for September–October launches.

• Most launches stall. Of 2026 repos mentioning "video" and "skill", 6.7% reached 10 stars, 1.9% 100, 0.32% 1,000 and 0.04% 10,000. "Shorts generator": 902 launches, none at 1,000.
• Companies hold the top and give the engine away. HeyGen, Browser Use, VEED (open-edit) and Diffusion Studio charge for cloud runs or AI credits. Staying engine-only is risky: Revideo's team moved to a commercial editor (Midrender).
• No open-source video repo publishes revenue. The closest proof is outside video: Postiz, and Plausible with 21,000+ paying subscribers on 29.6k stars after seven years (Plausible).
• Treat stars as upper bounds. Fake stars surged in 2024, mostly in AI repos (arXiv); some 2026 skill repos show thousands of stars and single-digit watchers.

Agents or people?

The evidence

• Attention moved to agent tools. HyperFrames skills each get ~100,000–160,000 installs a week on skills.sh; Remotion's best-practices skill has 647K in total (skills.sh). 2026 shorts generators for people stalled.
• Agents find tools by install rank. The most-installed skill, Vercel's find-skills (3.9M), tells agents to "Check the Leaderboard First", which ranks by installs (SKILL.md). Popular tools keep getting more popular.
• Skills work everywhere. Agent Skills became an open standard on 18 Dec 2025 (Claude blog), supported by about 45 agent apps including Claude Code, Codex, Cursor and Copilot; agents read only each skill's name and description until a task matches (agentskills.io).
• MCP directories are crowded, and video usage there is tiny. Smithery lists 20,786 servers. Of 150 "video" servers, 24 show any use, and the top one has 4,719 uses (Smithery).
• Agent pay-per-call exists but runs on rails closed to you. Weftly charges $0.50–2 per job in USDC or through Stripe (Smithery). Only 3 x402 sellers grossed over $10k in 30 days (x402scan).
• People pay; automation users stay longest (Postiz, via n8n and its API).
• READMEs now address agents. "Install this skill for me: [repo link]" (video-shotcraft); llms.txt on Remotion, HyperFrames and Firecrawl. ChatGPT refers ~10% of Vercel's new sign-ups (Vercel).

Recommendation: build for agents, sell to the person who runs the agent. Your primary customers are people who already use Claude Code, Codex or Cursor and need short videos: app founders, marketers, developer advocates, small studios. They star 2026 video skills, their agents do the setup, and their own keys pay for the AI. Non-technical creators are a different business (a polished web app marketed on TikTok against OpusClip and CapCut) that belongs in I3 or I10. Add a no-agent web form only if the waitlist asks.

What "optimizing for agents" means in practice

• A skill file whose description names its trigger tasks and keywords; a one-line install; a Claude Code plugin.
• A CLI where every command takes --json, returns typed errors and describes itself (contract --json, like ffmpeg-skill).
• Deterministic output: same inputs and seed, same video.
• --dry-run showing minutes and money before anything paid runs, and receipts so retries never pay twice (VEED's open-edit "spends nothing without your yes").
• Small text files instead of frames (video-use packs a take into ~12 KB).
• Users' own keys via environment variables, llms.txt, and an MCP server built from the same contract.

flowchart TD
    P["Person with a video job"] --> W["Searches web, X, YouTube, HN"]
    W --> R["Lands on GitHub repo"]
    R --> C["Clones it or gives up at setup"]
    C --> H["Hosted tier page"]
    A["Person asks their coding agent"] --> F["Agent runs find-skills"]
    F --> L["skills.sh, plugin directory, MCP registry"]
    L --> N["One-line install"]
    N --> J["Agent runs CLI with JSON output"]
    J --> D["Dry run shows minutes and cost"]
    D --> H
    H --> PAY["Person pays via Paddle"]

More things to publish

Where the AI compute runs

Next steps

flowchart TD
    A["Week 1 - scrub and extract owned code"] --> B["Week 2 - publish video-dna and shorts-qa"]
    B --> C["Day 30 - signal check"]
    C -->|"under 50 stars and under 100 installs"| K["Stop promoting, keep as in-house tooling"]
    C -->|"otherwise"| D["Weeks 5 to 8 - blocks and lyric-sync"]
    D --> E["Day 60 - hosted demand check"]
    E -->|"500 stars or 30 on waitlist"| F["Paddle hosted beta"]
    E -->|"below that"| K
    F --> G["Day 90 - 10 paying or 150 USD a month"]
    G -->|"yes"| H["Grow hosted tier and agent API"]
    G -->|"no"| K

I do:

• Extract and scrub video-dna and shorts-qa; write skill files, the CLI, an MCP server and llms.txt.
• Make demo GIFs from your footage; write READMEs with an "Install this skill for me" line.
• Submit listings (plugin directory, MCP registry, Smithery, awesome lists) and the HyperFrames blocks.
• Ship weekly releases, answer issues, and track stars, installs and the waitlist on a dashboard.
• Build the hosted render-and-check service and Paddle checkout only after day 60 says go.

You do once:

• Paddle verification plus one domain with pricing, terms, refund and privacy pages (start in week 1).
• Approve the licences (MIT skills; hosted code private) and the publishing account.
• Post once to Show HN and 2–3 subreddits.
• Confirm you own the AdvertorialTikTok engine before I6.5.

You do recurring:

• A 10-minute go/kill decision at each checkpoint.
• Moving Paddle payouts into your FOP account.

Go/kill thresholds

• Day 30 (20 Nov): under 50 stars and under 100 installs across both repos → stop promoting, keep as in-house tooling. 150 stars, 500 installs or 10 issues or pull requests from strangers → push on.
• Day 60 (20 Dec): open the hosted beta only at 500 stars or 30 on the waitlist.
• Day 90 (19 Jan 2027): keep the hosted tier only at 10 paying or $150 a month.
• Month 6 (Apr 2027): best repo under 3,000 stars → stop counting I6 toward $10k a month.

Risks and open questions

• Incumbents give it away. HeyGen, VEED or Browser Use could add DNA or checking any time; your edge is niche focus and speed, not a moat.
• Skills are easy to copy. Keep the hosted service and the curated style library closed.
• Inflated signals. One advertising skill pack shows ~15,000 installs per skill on 1 GitHub star. Judge by issues, repeat use and paying customers.
• Paddle policy. Are prepaid render minutes "stored value"? Is a render tool "automated social media marketing"? Ask during domain review; price as monthly plans meanwhile.
• Remotion licence. Remotion-based tools pay $0.01 a render ($100 monthly minimum) once a company passes 3 people (Remotion). Your renderer avoids Remotion; keep it that way.
• Music terms. Free-plan Suno outputs are non-commercial, and Suno bans using outputs to build competing services (Suno). Never resell Suno-via-Kie generation.
• Supply-chain trust. Skills run code on users' machines. Sign releases, skip post-install scripts, watch for scam forks (NarratoAI saw a fake token).
• Open question: does any open-source video tool earn real hosted revenue? None publishes figures; HyperFrames' cloud credits are the test to watch.

Sources

1. GitHub search, 2026 repos by stars
2. HN API, 2026 Show HN video posts
3. HN launch effect on stars (arXiv)
4. Fake stars on GitHub (arXiv)
5. HyperFrames repo
6. HyperFrames cloud render
7. HyperFrames music-to-video
8. npm, hyperframes
9. OpenMontage
10. video-use
11. video-shotcraft
12. ffmpeg-skill
13. openfilm
14. MoneyPrinterTurbo
15. AI-Youtube-Shorts-Generator
16. short-video-maker
17. YumCut
18. Diffusion Studio
19. VEED open-edit
20. Postiz at $14.2k a month
21. Postiz at $700 a month
22. Postiz star history
23. Plausible
24. skills.sh leaderboard
25. Remotion skills on skills.sh
26. find-skills skill
27. Agent Skills standard
28. Claude blog, Agent Skills
29. Claude plugin directory
30. Smithery, video servers
31. Weftly
32. x402scan
33. Vercel, AI referrals
34. Paddle acceptable use
35. Remotion licence
36. Suno terms
37. MMS_FA licence
38. Revideo, Midrender
39. cangjie-skill
40. drama-skills