Can OpenShorts run fully offline without any API key?

Not according to its README on 2026-10-18: moment picking can use a local OpenAI-compatible server, but Gemini is still needed for layout picking and for silent videos. Reelsaw runs the whole pipeline locally with Ollama and a local Whisper, which is what its installer sets up by default.

Can I use Anthropic, OpenAI or Groq with each tool?

Reelsaw: yes, each of the language, transcription and alignment models is a provider:model string, so you can mix vendors or go fully local, and admins can change them at runtime. OpenShorts: an OpenAI-compatible server can replace Gemini for moment picking; the other Gemini-backed steps stay on Gemini.

Which one picks better clips?

Reelsaw does more work to find out: chunking at pauses with overlap so long videos do not lose their middle, rubric scores as a median of several samples, pairwise comparison of the top ten with swapped positions, an optional separate judge model, and a run-relative gate instead of a fixed count. OpenShorts asks Gemini for 3 to 15 moments in one pass. Neither project publishes a head-to-head benchmark, so try both on your own video.

Which is more open: MIT or AGPL-3.0?

MIT lets anyone, including a company, take the code private and sell it as a service. AGPL-3.0 allows everything MIT allows except that: if you run a modified Reelsaw for other people, you must publish your changes. OpenShorts' core is MIT and its cloud directory is under a separate commercial licence; all of Reelsaw is AGPL-3.0.

Do both use an NVIDIA GPU?

Yes, and for the same three jobs. OpenShorts builds one image with a GPU flag: faster-whisper on CUDA, NVENC encoding, and its detection models. Reelsaw adds a Compose override (docker-compose.gpu.yml) that hands the card to the Speaches container (faster-whisper on CUDA with large-v3), to Ollama when it runs in a container, and to the app, whose ffmpeg then encodes with NVENC; its face, text and alignment models stay on the CPU by default because they are small enough that the GPU did not measure faster. Neither runs CUDA inside Docker on a Mac; there, Reelsaw's installer runs Ollama natively on Metal.

Do both have MCP servers?

Yes. Both can be driven from Claude Code, Claude Desktop, Cursor or a script. OpenShorts also has a REST API with webhooks and a pip-installable CLI; Reelsaw's CLI is an operator tool inside the container, and its MCP server can use OAuth so claude.ai connectors sign in through the browser.

Reelsaw vs OpenShorts

Reelsaw vs OpenShorts: fully local, and any model you want

Reelsaw exists for two reasons: the whole pipeline should run on your own hardware with no API key, and every model in it should be swappable for any vendor. OpenShorts is the other open-source clipper, and it gets most of the way there: MIT core, Docker Compose, an MCP server, and an OpenAI-compatible option for moment picking. But its README says Gemini is still required for layout picking and for silent videos, so a key and a Google account stay in the loop.

OpenShorts is wider: dubbing, AI UGC videos, a YouTube studio, publishing, a hosted tier. Reelsaw is deeper on the thing both tools are for, turning long video into good clips: how moments are ranked, how filler cuts are verified, what the editor can override, and how a shared install runs. If the core capability and control over it are what you are choosing on, Reelsaw is the stronger pick; the rest of this page shows why, and where OpenShorts wins instead.

Pricing checked 2026-10-18 on each vendor's public page. Vendors change plans without notice; verify before you buy.

What OpenShorts costs

Both are free to self-host. OpenShorts also sells a hosted version.

OpenShorts plans
Plan	Price	What it includes
Reelsaw	$0	AGPL-3.0, no hosted tier, no keys needed. Docker, about 10 GB disk, 8 GB RAM with a hosted provider, more for a local language model (the installer picks one by RAM)
OpenShorts self-hosted	$0, plus a Gemini key	MIT core; the cloud/ directory is source-available under a commercial licence. Docker, 8 GB+ RAM; about 5 to 8 minutes per 8-minute video on CPU, about a minute with an NVIDIA GPU
OpenShorts Cloud	Free, then from $12/mo	20 minutes a month free with a watermark; 100 minutes without a watermark from $12 a month

What Reelsaw costs for the same job

The software is free under the AGPL-3.0 and there is no hosted plan to upsell you to. What you pay for is the machine it runs on and, only if you choose cloud models, the per-call price of those models.

Fully local: Ollama for the language model and Speaches (Whisper) for transcription. The installer picks a model by GPU memory or RAM (qwen2.5:14b with 24 GB of RAM or a 12 GB GPU, qwen2.5:7b from 12 GB, gemma3:4b below that), reuses one Ollama already has, and puts both on an NVIDIA GPU when there is one. Nothing leaves the box and there is no per-minute cost at all.

With cloud models: set OpenAI, Anthropic, Google, Groq, Deepgram, AssemblyAI or any OpenAI-compatible endpoint as a provider:model string and you pay that vendor's metered price for transcription minutes and a few thousand tokens of selection per video. There is no Reelsaw margin on top.

OpenShorts vs Reelsaw, feature by feature

OpenShorts and Reelsaw compared
	Reelsaw	OpenShorts
Runs fully local, no API keys	Yes, and it is the installer's default: Ollama for the language model, a local Whisper for transcription, on-device alignment and face detection. Nothing leaves the machine	Partly. Moment picking can use a local OpenAI-compatible server, but Gemini is still needed for layout picking and silent videos, so a Google API key stays required
Model providers	Any, for each of the three models separately: OpenAI, Anthropic, Google, Groq, Deepgram, AssemblyAI, Vercel AI Gateway, Ollama, vLLM, LM Studio or any OpenAI-compatible server. No vendor is baked in; admins switch at runtime	Gemini by default for picking, layout, scripts and thumbnails; any OpenAI-compatible server for moment picking only. Transcription is faster-whisper (or Parakeet) in-process; dubbing needs ElevenLabs
Entry price	$0. AGPL-3.0 software; you supply the machine and, if you want cloud models, the API keys	$0 self-hosted (plus a Gemini key); hosted from $12/mo
Free tier	Everything, with no plan and no meter	Self-hosted: everything. Cloud: 20 minutes a month with a watermark
Watermark	Never	None self-hosted; on the free cloud plan
Open source	Yes, AGPL-3.0, pure TypeScript	MIT core; cloud/ directory under a commercial licence. Python and FastAPI, React frontend
Self-hostable	Yes: Docker Compose, one-command installer for macOS, Linux and Windows	Yes, Docker Compose; optional CUDA override
AI moment detection	Yes: map, pool, rank with pairwise comparison; any language model, optional separate judge model	Gemini picks 3 to 15 moments per video in one pass
Face-tracked 9:16 reframing	Yes: face, split, stack and fit layouts per segment; 1.2 MB UltraFace detector on the CPU	Yes: MediaPipe and YOLOv8; track, general, split and screencast modes, one layout picked per video by Gemini
Word-level captions	Yes: karaoke highlight, AI-picked key words, boxed or outlined, drag positions on a live preview	Yes: faster-whisper word-level subtitles, hook overlays, AI FFmpeg effects
Filler and false-start removal	Yes: jump cuts located by forced alignment, each one reviewable	Not listed
AI dubbing / translation	No	Yes: ElevenLabs, 30+ languages, your own key
Direct social publishing	No. Export zip with paste-ready text, a YouTube Data API body and a TikTok post body per clip	Yes: TikTok, Instagram Reels and YouTube Shorts, with scheduling; YouTube title and thumbnail studio
Agent and API access	MCP server (Streamable HTTP, OAuth optional) and an operator CLI	MCP server, REST API with webhooks, CLI, agent skill
Usage cap	None	None self-hosted; metered on Cloud
Accounts and teams	Multi-user with admin approval, per-user quotas and usage	Hosted accounts with API keys; self-hosted dashboard

What OpenShorts does better

Breadth beyond clipping: voice dubbing, AI UGC videos with lip-synced actors, a YouTube title and thumbnail studio, and one-click publishing.
A hosted option for people who do not want to self-host, with a GPU behind it.
An optional Parakeet transcription backend (NVIDIA NeMo), about twice as fast as Whisper on the same card.
A larger community: about 6.8k GitHub stars when checked.

Which one should you pick?

Pick OpenShorts if

You want dubbing, UGC generation or direct publishing from the same tool.
You might use the hosted version some months and self-host on others.
You prefer a Python stack to extend.

Pick Reelsaw if

No API key, ever: the one-command installer sets up Ollama and a local Whisper, and alignment and face detection run on-device. OpenShorts cannot drop Gemini.
Any provider for any model, separately: transcription from Deepgram and selection from Anthropic, or everything from one OpenAI-compatible server, switched at runtime by an admin without a restart.
Selection you can audit: every candidate is scored as the median of several samples, the top picks are compared pairwise with positions swapped to cancel judge bias, a separate judge model can do the scoring, and the rubric breakdown is one click away. OpenShorts asks Gemini for 3 to 15 moments in one pass.
Filler, repetition and false-start cuts located by forced alignment and checked by listening to the seam, each one reviewable. OpenShorts does not list this.
Framing decided per segment, not per video: face, split, stack and fit can change inside one clip at scene cuts, with active-speaker following, and any section can be overridden in the editor.
A real editor: multi-lane timeline, frame-accurate trims, layout sections, transitions, intros and outros, undo.
Built for more than one person: accounts, admin approval, per-user quotas, usage, and pg-boss jobs that resume after a crash.
AGPL-3.0, so a hosted fork has to share its changes.

Common questions

Can OpenShorts run fully offline without any API key?: Not according to its README on 2026-10-18: moment picking can use a local OpenAI-compatible server, but Gemini is still needed for layout picking and for silent videos. Reelsaw runs the whole pipeline locally with Ollama and a local Whisper, which is what its installer sets up by default.
Can I use Anthropic, OpenAI or Groq with each tool?: Reelsaw: yes, each of the language, transcription and alignment models is a provider:model string, so you can mix vendors or go fully local, and admins can change them at runtime. OpenShorts: an OpenAI-compatible server can replace Gemini for moment picking; the other Gemini-backed steps stay on Gemini.
Which one picks better clips?: Reelsaw does more work to find out: chunking at pauses with overlap so long videos do not lose their middle, rubric scores as a median of several samples, pairwise comparison of the top ten with swapped positions, an optional separate judge model, and a run-relative gate instead of a fixed count. OpenShorts asks Gemini for 3 to 15 moments in one pass. Neither project publishes a head-to-head benchmark, so try both on your own video.
Which is more open: MIT or AGPL-3.0?: MIT lets anyone, including a company, take the code private and sell it as a service. AGPL-3.0 allows everything MIT allows except that: if you run a modified Reelsaw for other people, you must publish your changes. OpenShorts' core is MIT and its cloud directory is under a separate commercial licence; all of Reelsaw is AGPL-3.0.
Do both use an NVIDIA GPU?: Yes, and for the same three jobs. OpenShorts builds one image with a GPU flag: faster-whisper on CUDA, NVENC encoding, and its detection models. Reelsaw adds a Compose override (docker-compose.gpu.yml) that hands the card to the Speaches container (faster-whisper on CUDA with large-v3), to Ollama when it runs in a container, and to the app, whose ffmpeg then encodes with NVENC; its face, text and alignment models stay on the CPU by default because they are small enough that the GPU did not measure faster. Neither runs CUDA inside Docker on a Mac; there, Reelsaw's installer runs Ollama natively on Metal.
Do both have MCP servers?: Yes. Both can be driven from Claude Code, Claude Desktop, Cursor or a script. OpenShorts also has a REST API with webhooks and a pip-installable CLI; Reelsaw's CLI is an operator tool inside the container, and its MCP server can use OAuth so claude.ai connectors sign in through the browser.

Related comparisons

Sources

Last updated 2026-10-18.