Oh My Cassette: Chat Your Raw Clips Into a Finished Cut

简体中文 | English

Point Claude Code, Codex, or Hermes at a folder of raw clips.
Describe the video you want. Get back a finished, beat-synced cut.

14 clips in. One prompt. ~13 minutes to a rendered file.
_{No timeline. No editing software. No GPU.}

⚡ Install in 30 seconds

# Claude Code
claude plugin marketplace add Cassette-Editor/oh-my-cassette
claude plugin install oh-my-cassette@cassette-editor

# Codex
codex plugin marketplace add https://github.com/Cassette-Editor/oh-my-cassette.git
codex plugin add oh-my-cassette@cassette-editor

Restart your agent, then say: "Edit the clips in ./footage into a 30-second travel vlog with beat-synced cuts."

Want to try the workflow before installing? Open the public web demo.

Needs Python 3.11–3.13, ffmpeg, and a Cassette account. Full setup — including OpenCode, Hermes, and any other MCP host — is in Quick Start.

🎥 Overview

Oh My Cassette is an open-source AI video editing plugin and local MCP server for Claude Code, Codex, Hermes Agent, and OpenCode. It turns natural-language chat into finished montage videos on Cassette — beat-synced cuts, auto-matched music, subtitles, transitions, and picture-in-picture — with minimal token overhead.

The agent does the parts that make editing tedious: it watches every clip, picks the shots, plans the cut, syncs it to the beat, and renders — while you stay in chat. Because it runs through your agent, you can do all of it from your phone.

🖥️ How it works

Upload your clips, describe the video you want, and the agent edits it.

The cut it produced

_{The screen recording is compressed to keep this page light, so the terminal text looks softer here than it does on your machine; the cut below it is the full-quality render. Prompt to rendered file took 9 minutes 53 seconds of real time, sped up above. A person's face is pixelated in both the screen recording and the cut it produced, for privacy — that blur is not something the agent added.}

_{Token cost: that session used 55K output tokens and 2.33M billed input tokens (2.07M of them cache reads) on Claude Opus 5 — roughly $4 at API list price, measured from the Claude Code transcript of this exact recording. One brief, one turn, start to exported file. The editing itself runs on Cassette, so the agent only pays for the brief and the timeline digests, not for the footage.}

🎬 Case Videos

Every case below was edited end-to-end by an AI agent through Oh My Cassette, from the exact prompt shown — real inputs, real processing times, and the output is what the agent delivered.

→ See all six cases — daily vlog, cinematic short, and game highlights, each with its exact prompt.
_{★ More cases are on the way. Star the project to follow along.}

🚀 Quick Start

Before You Start 🎬

Oh My Cassette connects Codex, Claude, or Hermes to the Cassette Agent. You need:

• Codex, Claude Code, or a working Hermes Agent installation.
• A Cassette account.
• For Hermes only, a configured gateway such as QQ or Telegram.

If you plan to use Hermes and it is not installed yet:

curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash

If the Hermes Agent gateway is not configured yet:

hermes gateway setup

Oh My Cassette currently supports QQ and Telegram gateways.

Requirements

• macOS, Linux, or Windows for the local MCP plugin (Codex, Claude Code, or OpenCode). The Hermes Agent gateway path is macOS/Linux only.
• Codex, Claude Code, or OpenCode for the local MCP plugin, or Hermes Agent with its gateway configured.
• Python 3.11–3.13.
• ffmpeg, required for Hermes gateway normalization and optional API export thumbnails.

Supported video uploads

Cassette currently accepts video source files that match all of the following limits:

MOV, HEVC/H.265, AV1, VP9, ProRes, HDR, 10-bit, variable-frame-rate video, 24/25/50/60 fps, 4K, oversized files, and videos longer than 60 minutes are rejected rather than transcoded in the cloud.

Upload retention

Uploaded videos and plugin-managed previews, job state, and exports share one immutable 24-hour deadline measured from the session's first ingest. After that deadline the plugin removes its local copies and the service removes its media; another session — even on the same host or the same account — cannot reach them. Keep your own copy of every source file, and move any export you want to retain outside the plugin-managed data directory before the deadline.

Install system tools:

# macOS
brew install uv ffmpeg

# Debian/Ubuntu Linux
sudo apt-get update
sudo apt-get install -y ffmpeg
curl -LsSf https://astral.sh/uv/install.sh | sh

# Windows (PowerShell)
winget install Python.Python.3.13
winget install Gyan.FFmpeg

Install

Codex

codex plugin marketplace add https://github.com/Cassette-Editor/oh-my-cassette.git
codex plugin add oh-my-cassette@cassette-editor

Start a new Codex task after installation so plugin discovery runs again. The plugin contributes the host-neutral cassette-video-edit and cassette-model skills plus a local MCP process named cassette. Fresh editing sessions use GPT-5.6 Luna with Extra High thinking; invoke $cassette-model only when you want to inspect or change that session setting.

Claude Code

claude plugin marketplace add Cassette-Editor/oh-my-cassette
claude plugin install oh-my-cassette@cassette-editor

Restart Claude Code after installation. You can verify the installation with:

claude plugin details oh-my-cassette@cassette-editor

Claude Code can keep the plugin up to date on its own once you enable auto-update for the marketplace — see Update.

Fresh editing sessions use GPT-5.6 Luna with Extra High thinking. Invoke /cassette-model when you want to inspect or change the current session's model and thinking level.

OpenCode

OpenCode's plugin manager installs npm packages only, and its plugins cannot contribute MCP servers, so there is no marketplace entry to add. One command instead:

curl -fsSL https://raw.githubusercontent.com/Cassette-Editor/oh-my-cassette/release/scripts/install_opencode.py | python3 -

That downloads the current release, writes the cassette server into ~/.config/opencode/opencode.json (merging with any servers and settings already there), installs the host-neutral skills into ~/.config/opencode/skills/, and installs /cassette-model into ~/.config/opencode/commands/. Restart OpenCode afterwards. Fresh editing sessions use GPT-5.6 Luna with Extra High thinking; the command changes the setting only when you invoke it.

Re-run the same command to update. Only git is not required — the release tarball is fetched with Python's standard library.

Cassette credentials are shared across Codex, Claude Code, OpenCode, and Hermes, so if you have already set up another host there is nothing more to do. Otherwise the installer prints the setup_local_mcp.py command to finish authentication. Jamendo uses the host-specific BYOK setup described below.

The plugin tree lands in ~/.oh-my-cassette (override with OMC_HOME). --dry-run previews the changes. If you prefer a git checkout, clone it and run the installer from there — it registers that tree and leaves it alone, and --sync fast-forwards it to the release channel.

Any other MCP host

The runtime is host-neutral, so any client that launches a local stdio MCP server can use it. Point the client at scripts/run_local_mcp.py (run with python3, or python on Windows) and set CASSETTE_RUNTIME_ADAPTER=mcp. The server ships full workflow guidance in its MCP instructions, and every tool returns a typed phase/next_action so a host without the packaged skill can still drive the flow. For the best experience, also install the cassette-video-edit and cassette-model skills (or equivalent system prompts). Generic clients can call cassette_config or ask in natural language to change the current session's model.

Letting an AI assistant do the install for you — Cline, or any agent with shell access? Point it at llms-install.md: a machine-readable setup guide covering the runtime pre-build, host registration, authentication, media roots, and verification.

Hermes

Install through the Hermes plugin manager (recommended):

hermes plugins install Cassette-Editor/oh-my-cassette

The Hermes installer prompts for your Cassette account email and password and saves them to ~/.hermes/.env. Then run the setup finisher — it configures the same stdio MCP server and canonical editing skill used by the other hosts, sets Hermes's tool timeout to 1800 seconds, detects ffmpeg/ffprobe, and lets you pick the Cassette region — and enable the thin gateway plugin:

python3 ~/.hermes/plugins/cassette/scripts/install_plugin.py --setup-only
hermes plugins enable cassette
hermes gateway restart

You can check the install status anytime from the Diagnose section.

git clone https://github.com/Cassette-Editor/oh-my-cassette.git
cd oh-my-cassette
python3 scripts/install_plugin.py

python3 scripts/install_plugin.py --copy --force

python3 scripts/install_plugin.py \
  --skip-plugin-enable \
  --skip-cassette-url \
  --skip-cassette-auth \
  --skip-jamendo-auth

Use with agent clients over local MCP

Codex, Claude Code, OpenCode, and Hermes use the same self-contained runtime. In this README, MCP server means a local child process connected over stdin/stdout: it opens no port. The separate Cassette backend remains the editing engine and continues to handle authentication, media processing, agent runs, project state, and rendering.

First-run authentication

Cassette passwords are generated by the server and emailed to you. You never choose one, and the plugin never invents one.

Missing credentials do not prevent the MCP process from starting. There are two ways to hand the password over.

In the conversation. Paste the password from your Cassette email and ask the agent to sign in; it calls the cassette_login tool, which verifies the account against Cassette before writing anything and then stores it privately. Nothing else is needed — no terminal, no browser. The trade-off is explicit: the password lands in your agent host's transcript on disk and is sent to the model provider for the rest of that conversation. If that is not acceptable to you, use the terminal instead.

In a private terminal. The command keeps the password out of the transcript entirely — it prompts with getpass, verifies the account before writing anything, and stores credentials in the platform-standard Oh My Cassette config directory. Every auth_required envelope carries the exact command for your install; from a git checkout it is:

python3 scripts/setup_local_mcp.py --email you@example.com

Credentials may also come from process environment variables. Environment values take precedence over protected local config, so cassette_login refuses to write a file that would be shadowed. Importing an existing Hermes .env is explicit and optional:

python3 scripts/setup_local_mcp.py --import-hermes

Either route creates config directories with mode 0700 and credential files with mode 0600, rejects symlinks and permissive files, and never persists access or refresh tokens.

Jamendo BYOK setup

Jamendo music matching is strictly bring-your-own-key for every local agent host. Create a read-only application in the Jamendo developer portal and copy its Client ID. A Client Secret is neither needed nor accepted.

In the conversation. Ask the agent to configure Jamendo, paste your Client ID, and it calls cassette_jamendo_setup. The tool verifies a minimal Tracks request before writing anything. The trade-off is the same as chat sign-in: the ID reaches the host transcript and model provider for that conversation.

In a private terminal. Keep the ID out of the conversation entirely:

python3 scripts/setup_local_mcp.py --jamendo

Codex, Claude Code, and OpenCode store it in the protected settings.json; Hermes stores it in ~/.hermes/.env. JAMENDO_CLIENT_ID in the process environment remains highest precedence. A failed validation preserves the previous working value.

BYOK assigns API access and quota to your Jamendo application; it does not grant commercial rights to selected music. Review the returned track URL, license URL, download eligibility, and attribution requirements before publishing or commercial use.

Resetting the password

You cannot pick a replacement — Cassette generates one and emails it to the account address. Ask the agent for a new password and it calls cassette_login with request_new_password and confirm_replace; or from a terminal:

python3 scripts/setup_local_mcp.py --reset-password

Both routes ask you to confirm first, because the request is irreversible: it replaces the account password everywhere, including on your other machines, it is limited to a few attempts an hour, and the replacement happens before the email is sent — so even a delivery failure kills the old password. Check your inbox before retrying.

The reset only mails the new password; paste it back (in the conversation, or at the terminal prompt) to finish signing this machine in. The verified password is stored in credentials.json:

• macOS: ~/Library/Application Support/Oh My Cassette/credentials.json
• Linux: ~/.config/oh-my-cassette/credentials.json (or under XDG_CONFIG_HOME when set)
• Windows: %APPDATA%\Oh My Cassette\credentials.json

To make this machine forget the stored password without touching the account:

python3 scripts/setup_local_mcp.py --logout

Usage: point it at your clips

1. Start your agent in the folder that holds the clips — or in any parent of it. That folder is the trusted media root (CASSETTE_PROJECT_ROOT, set to the host's project directory), and everything beneath it is ingestible, so ~/videos/trip/raw/*.mp4 works when you start in ~/videos/trip. Clips somewhere else? Register that directory once:

   bash

   python3 scripts/setup_local_mcp.py --allowed-root /absolute/path/to/media
   

   Ingesting a file outside every trusted root fails with source_path_not_allowed.

2. Say what you want, in one message, naming the folder. No upload step to run yourself — the agent ingests the files it needs.

   Edit the clips in ./footage into a 30-second travel vlog with beat-synced cuts. Add the title "KOTA KINABALU" at the start and end, and keep the rhythm light.

3. Keep going in the same conversation. Each turn commits the edit and returns a timeline digest plus a contact-sheet JPEG saved locally with a clickable link — nothing renders. Hermes labels this as the thumbnail and uses the same saved file; no editor deep link is exposed. "Make the intro shorter", "swap the music for something calmer", and "undo that" all continue the same session.

4. Say "export" when you're happy. That's the only thing that starts a render; the finished file lands in cassette/exports/<job_id>/.

Supported inputs are video, image, and audio files (.mp4, .mov, .jpg, .png, .mp3, .wav, and friends). Mixed folders are fine — send the footage and the music track together.

Guided editing flow

What the plugin does under the hood on each of those turns:

1. Ask your agent client to edit one or more media files in the current project.
2. The plugin ingests only media inside the active project or another explicitly trusted root. It canonicalizes paths and rejects traversal and symlink escapes.
3. Describe the edit, answer any guided choices, and start the job. cassette_run_job is the wait: the host calls it exactly once for that user turn while the runtime streams progress notifications.
4. When that call settles, follow its typed phase and next_action; do not start a status-poll loop or retry the edit in the same user turn. cassette_job_status is reserved for a deliberately detached or interrupted call.
5. If Cassette needs a real user decision, answer it with the returned job ID. On hosts that support MCP elicitation, cassette_job_status collects the answer inline and returns the already-resumed status; other hosts use the cassette_answer_question round-trip. API jobs persist their private continuation metadata across host restarts.
6. When editing completes, review the result. Rendering starts only after an explicit export decision.
7. The result contains validated absolute paths, file URIs, MIME type, size, and an MCP resource link for each exported artifact. Large media bytes are never embedded in the tool response.

When a background job reaches a terminal state (finished, needs input, failed, or cancelled), the MCP runtime posts a best-effort local desktop notification — osascript on macOS, notify-send on Linux — so you learn a long render is done even after the monitor budget hands the job back. Set CASSETTE_MCP_NOTIFY=0 to disable it.

Sessions are isolated by a cryptographically random session ID. Codex, Claude Code, OpenCode, and Hermes share host-neutral storage, so you can deliberately hand a session or job ID from one host to another; nothing is shared implicitly.

Additional trusted media directories can be registered during setup:

python3 scripts/setup_local_mcp.py --allowed-root /absolute/path/to/media

Exports stay under the shared Oh My Cassette data directory at cassette/exports/<job_id>/. Only files contained in that job-specific directory can be returned.

How the plugin reaches Cassette

One way: direct calls to the separate Cassette backend. Authentication retries once after a 401, access tokens are kept in memory only, and continuation metadata is persisted, so a paused job resumes after the agent client restarts.

There is no browser to install, drive, or keep alive. The Playwright transport that used to sit behind CASSETTE_TRANSPORT=browser has been removed; a leftover setting is reported once on stderr and ignored.

MCP tools

The local MCP runtime exposes the same 16 tool names as Hermes:

Every tool returns a structured envelope with ok, typed data or error, session_id, job_id, the current phase, and a runtime-derived next_action.

Local preview links and plan review

The runtime returns no editor deep link. A …?projectSessionId=<id>&chatSessionId=<uuid> URL is a bearer capability: the backend binds no owner to a scratch session, so the only checks on that route are "signed in" and "knows the id" — any authenticated account that sees the link can open the project and run edits on the thread. Tool output ends up in chat transcripts, logs and screen recordings, so the runtime no longer emits one, and the skills instruct the agent not to construct one. Previews are the timeline digest, the contact sheet, and the export.

This narrows exposure rather than closing it: the route still resolves for anyone who reconstructs the URL. Binding a session to its owner has to happen server-side.

Two concurrency semantics worth knowing: a plugin turn never cancels a run started from the open editor tab (it fails typed as thread_busy instead — wait and retry), while typing a fresh message in the tab DOES cancel an in-flight plugin turn (the tab takes over; existing product behavior).

Behavior change (0.4.14): the agent receives the user's message verbatim (no brief wrapper), sessions are multi-turn on one thread, and a turn ends with the edit committed but nothing rendered — the envelope carries timeline_delta, quality.timeline_ctl, and a contact-sheet preview instead; pass export=true on the turn where the user asks to finish. A fresh session starts immediately with GPT-5.6 Luna and Extra High thinking; model selection is opt-in through $cassette-model in Codex, /cassette-model in Claude Code/OpenCode, /cassette_model in Hermes, or an explicit natural-language request. The saved session preference applies from the next turn.

Behavior change (0.4.0): on MCP hosts, edit_plan_review now surfaces as a real question by default (CASSETTE_PLAN_REVIEW=user) instead of being silently auto-approved — answer with approve, revise <feedback>, or reject, in chat or in the open editor tab (first answer wins). Set CASSETTE_UNATTENDED=1 to restore the previous fully headless behavior. Status envelopes additionally carry timeline_delta (what changed) and plan_progress, fed by the run's SSE event stream (CASSETTE_API_STREAM=0 disables).

Ready to Use with Hermes 📼

Now you can pick up your phone and DM your agent! Don't forget to keep your agent alive and network connected.

In QQ or Telegram:

1. Send one or more video, image, or audio files.
2. Wait for the saved-material acknowledgement.
3. Send the edit instruction in the same conversation or prefix it with /edit. The session starts with GPT-5.6 Luna and Extra High thinking; send /cassette_model only when you want to change them. Your words go to the Cassette agent verbatim; optimization and BGM remain explicit via /refine and /music.
4. Each turn ends with the edit saved (timeline delta + contact-sheet preview, no render). Depending on the client, the preview is delivered as an image or a labeled local thumbnail link. Keep editing in the same conversation. Say "export" when you want the video, and the plugin renders and delivers it through the gateway when supported.

• Assets and video state are preserved within the same conversation session. You can send additional messages in the same session to further modify the edited video results.

• Use /new or /reset to start a fresh Hermes session and clear the live Cassette session and your assets for that conversation.

• QQ is set to Chinese and Telegram is set to English by default, you can set language by command /cassette language zh/en manually.

Update

The runtime checks the release channel once a day and, when a newer version exists, tells your agent — which mentions it once and offers to run the command below for you. Set CASSETTE_UPDATE_CHECK=0 to turn that check off.

Claude Code: automatic updates

Claude Code checks for marketplace and plugin updates after your session starts, with a random delay of up to ten minutes, so the session you are in keeps the version it launched with — you are prompted to run /reload-plugins, or the new version loads next launch. It is off by default for third-party marketplaces, so turn it on once — scripts/setup_local_mcp.py offers to do this during setup (skip with --no-auto-update), or do it yourself in /plugin → Marketplaces → cassette-editor → Enable auto-update, or in ~/.claude/settings.json:

{
  "extraKnownMarketplaces": {
    "cassette-editor": {
      "source": { "source": "github", "repo": "Cassette-Editor/oh-my-cassette" },
      "autoUpdate": true
    }
  }
}

Declaring it in settings.json wins over the /plugin toggle — Claude Code syncs the declared value into its marketplace state and then points you back at the settings file to change it.

Because the setup prompt only runs at setup, an install that predates it — or one where it was declined — would never hear about the toggle again. So on Claude Code the runtime also reads that setting at startup and, while auto-update is off, asks your agent to mention it at most once per session and point you at the toggle. Your agent will not edit your Claude configuration itself. CASSETTE_UPDATE_CHECK=0 silences this along with the release check.

Claude Code's DISABLE_AUTOUPDATER turns off all automatic updates including plugins; pair it with FORCE_AUTOUPDATE_PLUGINS=1 to keep plugin updates while managing Claude Code itself manually.

Codex

Codex refreshes configured git marketplace snapshots on its own, but installed plugins are cached per version, so one command applies the new one:

codex plugin marketplace upgrade cassette-editor   # only needed if the snapshot is stale
codex plugin add oh-my-cassette@cassette-editor

OpenCode

Re-run the install command — it is the same command for installs and updates:

curl -fsSL https://raw.githubusercontent.com/Cassette-Editor/oh-my-cassette/release/scripts/install_opencode.py | python3 -

The local launcher updates its locked, plugin-managed virtual environment automatically on the next start after any of these.

Hermes

If installed through the Hermes plugin manager:

hermes plugins update cassette
hermes gateway restart

If installed from a git checkout (symlink install), update the checkout:

git pull --ff-only
hermes gateway restart

If the plugin was installed with --copy, reinstall the copied plugin after pulling:

git pull --ff-only
python3 scripts/install_plugin.py --copy --force
hermes gateway restart

See CHANGELOG.md for what changed in each release.

rm ~/.hermes/plugins/cassette        # removes only the symlink, not your checkout
hermes plugins install Cassette-Editor/oh-my-cassette
hermes gateway restart

❓ FAQ

Which AI agents does Oh My Cassette work with?

Claude Code, Codex, OpenCode, and Hermes Agent are supported out of the box, and any other MCP host can connect to the local cassette MCP server. Sessions live in a host-agnostic data directory, so you can start an edit in one host and continue it in another.

Does it edit videos locally or in the cloud?

The plugin runs locally beside your agent and handles media ingestion, edit planning, and job supervision. The actual editing and rendering happen on Cassette, so you don't need a GPU or any editing software installed.

Do I need a Cassette account?

Yes. The installer asks for your Cassette account email and password on first run and stores them locally; the plugin authenticates with Cassette on your behalf.

What kinds of videos can it make?

Montage and story edits — vlogs, travel videos, music-driven shorts, cooking tutorials, product commercials, and game highlights — with beat-synced cuts, subtitles, picture-in-picture, transitions, and auto-matched background music. See Case Videos for real examples with prompts and processing times.

Can I see what the agent is doing mid-edit?

Yes. After each editing turn you get a timeline digest and a contact-sheet JPEG saved locally. Terminal clients present a clickable local thumbnail link, and supported gateway clients can deliver the preview as an image. Before the edit runs, the agent can also surface the plan as a storyboard sheet (one source frame per planned beat) for review. The plugin intentionally does not expose an editor deep link.

Can I try it without installing anything?

No. Install the plugin in a supported agent host and sign in with a Cassette account.

Is Oh My Cassette free and open source?

This plugin is free and open source under the MIT license — all of it, including the MCP server and the skill.

Rendering runs on Cassette, a separate hosted service that requires an account. See Cassette's pricing for what an account costs.

🔨 Development & Troubleshooting

Setup for contributors, the full configuration reference, transport internals, the diagnostic scripts, and answers to common runtime problems all live in docs/development.md.

Quick diagnostic — run this first when something misbehaves:

python3 scripts/diagnose_local_mcp.py   # Codex / Claude / OpenCode
python3 scripts/diagnose_install.py     # Hermes

Both report bootstrap, config, transport, and media-root state without printing credentials.

💬 Community

• Discussions — ask questions, share the cuts you made, and see what's planned.
• Discord — chat with contributors and other users.
• Issues — bug reports and feature requests. Issues labelled good first issue are a good place to start.
• CONTRIBUTING.md — how to propose a change.

Made something you like with it? Post it in Discussions — we feature the best cuts in the showcase.

License

MIT. See LICENSE.