Video and design tools for people and AI agents, built for Hermes Agent. One name everywhere: the app, the hermes-studio command, the MCP server and the Hermes plugin.
Repo: https://github.com/PabloTheThinker/hermes-studio
Long video, YouTube, X, Twitch, or a livestream slice in. Captioned 9:16 shorts out. A Canva-style Design page for carousels, story covers, thumbnails and posts, which your AI agent can build and edit too. Nothing is posted for you.
One line. It downloads the latest app from this repo's Releases, checks it against the published SHA-256, and sets everything up. No admin rights, no Python, no FFmpeg, nothing else to install. Run the same line again to update.
Linux (any terminal):
curl -fsSL https://raw.githubusercontent.com/PabloTheThinker/hermes-studio/main/scripts/install.sh | bashWindows (PowerShell):
irm https://raw.githubusercontent.com/PabloTheThinker/hermes-studio/main/scripts/install.ps1 | iexYou get:
- Hermes Studio in your app menu (Start menu and desktop on Windows).
- The
hermes-studiocommand, using the engine inside the app, so it works with no extra setup. - A fixed place your AI apps can start it from:
hermes-studio mcp install claudejust works.
Then:
hermes-studio doctor # everything green?
hermes-studio app # open the app
hermes-studio update # later: get the newest releaseLinux options: … | bash -s -- --version v0.5.1 pins a release, --uninstall removes the app (your clips stay), --help lists the rest. It installs to ~/.local/share/hermes-studio and needs no FUSE.
macOS: not yet. The command line below works today.
It runs on your computer only. The speech model (about 75 MB) downloads the first time you clip; after that it works offline. Nothing is uploaded or posted.
From Releases: Hermes-Studio-Setup-<version>.exe (Windows) or Hermes-Studio-<version>.AppImage (Linux: chmod +x, then open it). SHA256SUMS.txt lists the checksums.
Open the desk (hermes-studio studio) and pick Design. Choose a size (carousel 4:5, story 9:16, square, YouTube thumbnail, X/LinkedIn), start blank or from a template, then:
- drag, resize, rotate and snap to the centre and edges; multi-page with a page strip
- text in five bundled fonts (Archivo, Playfair Display, Caveat, JetBrains Mono, Open Sans; all OFL), colours, alignment, spacing, banners
- shapes, lines, cards, pills and a bottom shade for text over photos
- photos from your computer or your clips: remove background, enhance, looks (black and white, warm, cool, punch, fade, noir), all on your machine
- brand colours, layers, undo/redo, keyboard nudges, autosave
- resize a design into another format; export every page as PNG or JPG
Designs live in clips/designs in your Hermes folder as plain JSON plus images, so an AI agent and you edit the same file. From the command line:
hermes-studio design new --template carousel --size tiktok-carousel # returns an id
hermes-studio design edit <id> '[{"op":"update","page":1,"index":0,"set":{"text":"MY HOOK"}}]'
hermes-studio design render <id> --format jpg # PNG/JPG per page
hermes-studio design resize <id> --size story
hermes-studio photo cutout me.jpg # transparent PNGBackground removal uses U²-Net small (Apache-2.0, 4.5 MB, bundled; runs offline). The editor uses Fabric.js (MIT), bundled.
Pick Edit in the desk to cut a video on a real timeline: import files (proxies, filmstrip thumbnails, waveforms, shot changes and local Whisper words are made in the background), drag, trim, split, slip and roll clips with snapping, add text and markers, find words in the transcript and cut them, remove fillers or long pauses in one step, apply presets (fades, crossfades, title and end cards, music ducking, close the gaps), and export at 1080p, 4K, vertical, or ProRes (draft/standard/high quality, optional fps override) with a progress bar and cancel. The preview plays every track and shows the captions the render will burn. Every change is one entry in the history with Undo; press ? for the keys (J/K/L, arrows, S, M, Ctrl+C/V/D, multi-select).
Your AI app edits the same timeline over MCP. In Propose mode (the default) each agent edit waits in the sidebar until you Apply or Skip it, and Preview shows what it would change first; Ask lets agents only look, Auto lets them edit directly. Before/after frames show what each edit did, and a draft lets an agent try a whole idea on a copy you keep or discard. Agents read the edit in plain words (timeline_outline) and can try edits without making them (timeline_check). In a browser, paste the Edit page code that hermes-studio studio prints; the desktop app connects by itself.
If you'd rather install it as a Python package: you need Python 3.11+ and FFmpeg with libass (the app above has both built in).
uv tool install git+https://github.com/PabloTheThinker/hermes-studio # or: pipx install git+https://…
# gives you the `hermes-studio` command
hermes-studio doctor # checks FFmpeg, captions, speech, linksEveryday use:
hermes-studio run talk.mp4 # 3 hook-first shorts
hermes-studio run https://youtu.be/ID -n 5 # 5 shorts from a link (live streams too)
hermes-studio run talk.mp4 --mode captions # caption the whole video
hermes-studio run stream.mp4 --layout split # facecam on top, gameplay below
hermes-studio run talk.mp4 --prompt "the pricing part" --filter cinematic
hermes-studio list # your runs, newest first
hermes-studio show <id> # clips, scores, files
hermes-studio open <id> # open the folder
hermes-studio studio # the desk in your browser
hermes-studio render <project> --preset 1080p --json # export an Edit timeline (also 4k, vertical, prores)hermes-studio --help and hermes-studio <command> --help list everything. Clips are saved to your library (clips/library inside your Hermes folder; hermes-studio doctor shows where); -o DIR also copies them to DIR. Typos get a "did you mean".
Coming from the old hermesclip name? Update with the same install line, then re-run hermes-studio mcp install <app> so your AI app uses the new name. The old hermesclip command still works and tells you the new name. Old HERMESCLIP_* settings still work.
- Add
--jsonto any command: stdout is exactly one JSON object; progress goes to stderr as JSON lines. - Exit codes:
0ok ·1the job failed ·2bad input ·3missing dependency ·4not found. Errors carrycodeand ahint. hermes-studio run … --detach --jsonreturns an id straight away; pollhermes-studio show <id> --jsonuntilstatusiscompletedorfailed.- Each clip in a result has
path,title,score,start,endandseconds.
hermes-studio mcp install claude # Claude Code
hermes-studio mcp install grok # Grok CLI
hermes-studio mcp install codex # Codex CLI
hermes-studio mcp install cursor # Cursor (~/.cursor/mcp.json)
hermes-studio mcp install hermes # Hermes Agent (config.yaml)
hermes-studio mcp config # the JSON block for anything elseThen ask it: "make 3 shorts from ~/Videos/talk.mp4". Tools: run, show, list, restyle, edit, probe, recommend, copy, name, tools, doctor, plus design_new, design_list, design_show, design_edit, design_render, design_resize and photo. Ask: "make me a 5-slide carousel about my launch". Standard MCP stdio; long runs report progress, or pass detach: true and poll show. The desktop app's Settings → Use with your AI shows the same lines with a copy button.
git clone https://github.com/PabloTheThinker/hermes-studio.git
cd hermes-studio
python3 -m venv .venv
.venv/bin/pip install -e '.[reframe]'
.venv/bin/hermes-studio doctorCopy hermes_plugin/hermes-studio/ into $HERMES_HOME/plugins/hermes-studio/ and add hermes-studio to plugins.enabled (or use hermes-studio mcp install hermes). Hermes skill: copy skills/hermes-studio/ into $HERMES_HOME/skills/hermes-studio/. Tools appear in the next session. Does not post.
https://pablothethinker.github.io/hermes-studio/
Inspired by and partly ported from BridgeClip by BridgeMind (MIT). The clip scoring rubric and the framing modes started there. See NOTICE.
Built for Hermes Agent by Nous Research.
MIT. See LICENSE. Third-party notices are in NOTICE.
The desk is local-only and hardened against browser attacks. See SECURITY.md, and report problems privately.