What is OpenShorts?
OpenShorts, from the GitHub user mutonby, takes a podcast, webinar or stream by URL or upload, transcribes it with faster-whisper, has Gemini 3.1 Flash-Lite pick 3 to 15 moments, reframes each to 9:16 with face tracking or a stacked two-speaker layout, and burns in captions and a hook line. The same app also makes AI-actor marketing videos and YouTube thumbnails, though those are not exposed over MCP.
The MCP server has eight tools and runs three ways: hosted at mcp.openshorts.app with OAuth or an osk_ API key, self-hosted at /mcp on the Docker backend, or over stdio with python mcp_stdio.py. The repository is MIT-licensed, has about 7,720 stars and no tagged releases, and was last pushed on 11 November 2026.
How to install OpenShorts
claude mcp add --transport http openshorts https://mcp.openshorts.app/mcp --header "Authorization: Bearer osk_..."Is OpenShorts worth it?
OpenShorts is one of the few clip generators an agent can drive end to end. The eight tools are process_video, create_upload, get_job_status, list_clips, get_quota, add_subtitles, recut_clip and publish_clip, so a prompt such as clipping a podcast and scheduling the best three to TikTok becomes a short tool sequence. Jobs take minutes, so the agent polls or passes a webhook_url, and completion webhooks can be signed with HMAC-SHA256.
The model calls are modest. In the clip pipeline the only cloud call is the moment picker, which sends the transcript, not the video, to Gemini 3.1 Flash-Lite; the README puts that at under $0.01 per 10-minute video. Setting LLM_BASE_URL moves it to Ollama, LM Studio or vLLM, though the automatic layout picker and silent videos still need Gemini to look at frames. Dubbing uses your ElevenLabs key, posting goes through Upload-Post, whose free tier allows 10 uploads a month, and the AI-actor shorts outside MCP cost about $0.65 to $2 a video in fal.ai calls. Hosted, Gemini is included and you pay by source minute: Starter works out at $0.12 a minute, dearer than Opus Clip's entry plan per minute by the pricing page's own figures.
What the server can do matters more than what it costs. publish_clip posts publicly: Instagram and YouTube publish directly and TikTok lands as a draft. process_video fetches any public URL, behind an SSRF guard, and the transcript of someone else's video flows into clip titles your agent reads, a modest prompt-injection path. The server's instructions tell the agent to confirm rights to the video and to publish only when asked, but those are instructions to a model, not a lock.
The catch: the self-hosted /mcp endpoint takes no key, and the stock docker-compose.yml publishes port 8000 on every interface. If your Gemini or Upload-Post keys sit in .env, which the backend falls back to, anyone who can reach that port can spend them and post to your accounts. Bind it to 127.0.0.1 or put it behind a proxy with auth. The hosted cloud keeps your videos on OpenShorts' servers, and its retention terms are not stated in the README.
Our take: KinoCut is the other open option to weigh. OpenShorts earns its place when the end goal is posting clips on a schedule, as long as whoever runs the Docker stack treats it like any other service holding publishing credentials.
Where it earns its place
A two-person weekly podcast
Get three vertical clips out of every episode without opening an editor
Ask the agent to process the YouTube URL, review list_clips, and schedule the best three with publish_clip. A 60-minute episode uses 60 cloud minutes, so Creator at $29/month covers about five episodes.
A marketing agency running channels for five clients
Turn client webinars into captioned shorts on a posting calendar
Self-host on a GPU box with each client's Upload-Post key passed per call, so nothing sits in .env, and have the agent recut and re-caption clips before scheduling. Costs are the GPU and Gemini calls of under $0.01 per 10 minutes.
A university communications office with privacy rules
Clip lecture recordings without sending video to a third party
Self-host with LLM_BASE_URL pointed at a local Ollama model and AUTO_LAYOUT off, so only local models see the transcript. Publishing stays manual.
Before you commit
- Setup
- Hosted: add https://mcp.openshorts.app/mcp as a connector in claude.ai or ChatGPT (OAuth), or use claude mcp add --transport http with an osk_ key. Self-hosted: git clone, docker compose up --build, then point the client at http://localhost:8000/mcp, or run python mcp_stdio.py over stdio.
- Integrations
- Eight tools: process_video, create_upload, get_job_status, list_clips, get_quota, add_subtitles, recut_clip, publish_clip. Read-only: get_job_status, list_clips, get_quota. REST API with OpenAPI docs at /docs and signed completion webhooks. Calls Gemini, optionally ElevenLabs, Upload-Post and S3.
- Data and privacy
- Self-hosted, video stays on your server and only the transcript goes to Gemini, or nowhere with a local LLM; frame-based layout picking sends frames to Gemini. The self-hosted /mcp endpoint has no authentication. publish_clip posts to connected TikTok, Instagram and YouTube accounts. Hosted, video is processed and stored on OpenShorts' servers; retention terms are not published in the README.
- Cost at scale
- Cloud bills by source-video length, not by clip, with MCP and API calls at no extra fee: Starter $12/month for 100 minutes, Creator $29 for 300, Pro $59 for 750, yearly about 17% less. Self-hosted, Gemini costs under $0.01 per 10-minute video and an 8-minute video takes 5 to 8 minutes on CPU or about a minute on an NVIDIA GPU.
- Limits
- AI-actor shorts and YouTube Studio are not exposed over MCP. Ollama's 4,096-token default context truncates silently; the README asks for 16,384. Small 3B local models do not return valid JSON reliably. The self-hosted endpoint works only while your machine is on.
What’s good
- Eight tools from URL to scheduled post
- Transcript-only cloud call; can run on a local LLM
- Hosted plan uses OAuth 2.1 or revocable osk_ keys
- Free, MIT, no watermark when self-hosted
What’s not
- Self-hosted /mcp has no auth; compose exposes port 8000
- publish_clip posts publicly to social accounts
- Layout picker and silent videos still need Gemini
- No tagged releases; Docker and 8 GB RAM to self-host
Best for
Podcasters, small video teams and agencies who want an agent to cut long recordings into captioned shorts and queue them for posting.
OpenShorts alternatives
OpenShorts: common questions
- Is OpenShorts worth it?
- A capable, cheap way for an agent to cut long videos into vertical clips, let down by a self-hosted endpoint that asks for no key at all. We rate it 3.7 out of 5.
- How much does OpenShorts cost?
- Self-hosted: free (MIT), you pay your own API keys. Cloud: Free 20 source minutes/month with watermark; Starter $12/month (100 min); Creator $29/month (300 min); Pro $59/month (750 min); extra 60 min for $9.
- What are the downsides of OpenShorts?
- Self-hosted /mcp has no auth; compose exposes port 8000; publish_clip posts publicly to social accounts; layout picker and silent videos still need Gemini; no tagged releases; Docker and 8 GB RAM to self-host.
- Who is OpenShorts best for?
- Podcasters, small video teams and agencies who want an agent to cut long recordings into captioned shorts and queue them for posting.
- How much does OpenShorts cost for a team or in production?
- Cloud bills by source-video length, not by clip, with MCP and API calls at no extra fee: Starter $12/month for 100 minutes, Creator $29 for 300, Pro $59 for 750, yearly about 17% less. Self-hosted, Gemini costs under $0.01 per 10-minute video and an 8-minute video takes 5 to 8 minutes on CPU or about a minute on an NVIDIA GPU.
- Is OpenShorts safe for company data?
- Self-hosted, video stays on your server and only the transcript goes to Gemini, or nowhere with a local LLM; frame-based layout picking sends frames to Gemini. The self-hosted /mcp endpoint has no authentication. publish_clip posts to connected TikTok, Instagram and YouTube accounts. Hosted, video is processed and stored on OpenShorts' servers; retention terms are not published in the README.
- What are the best OpenShorts alternatives?
- The closest alternatives we have reviewed are Kinocut, OpenScreen and OpenEdit.
Answers checked Nov 11, 2026.