AgentStack
Browse Sign in
Browse Why AgentStack Sell Docs
Sign in
SKILL verified MIT Self-run

Cassette Video Edit

skill-cassette-editor-oh-my-cassette-cassette-video-edit · by Cassette-Editor

Edit project media through the local Oh My Cassette MCP tools in Codex or Claude — a direct multi-turn conversation with the Cassette agent, with timeline previews, guided questions, and explicit export.

— No reviews yet
0 installs
24 views
0.0% view→install

Install

$ agentstack add skill-cassette-editor-oh-my-cassette-cassette-video-edit

✓ scanned · ✓ verified, works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • ✓ Prompt-injection patterns
  • ✓ Secret / credential exfiltration
  • ✓ Dangerous shell & filesystem operations
  • ✓ Untrusted network calls
  • ✓ Known-malicious package signatures

What it can access

  • ✓ Network access No
  • ✓ Filesystem access No
  • ✓ Shell / process execution No
  • ✓ Environment & secrets No
  • ✓ Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

View the full security report →

Verified badge

Passed review? Show it. Paste this badge into your README, it links to the public security report.

AgentStack Verified badge Links to your public security report.
[![AgentStack Verified](https://agentstack.voostack.com/badges/verified.svg)](https://agentstack.voostack.com/security/report/skill-cassette-editor-oh-my-cassette-cassette-video-edit)

Reliability & compatibility

✓ Security review passed
0 installs to date
— no reviews yet
● 2mo ago

Declared compatibility

Claude CodeClaude Desktop

Compatibility is declared by the source manifest. End-to-end runtime verification is coming, see below.

Preview Execution monitoring

We're building live execution health for every listing: tool-call success rate, median latency, uptime, and last-checked timestamps, measured, not self-reported. It isn't live yet, so we don't show numbers we can't stand behind.

How agent discovery & health will work →
Are you the author of Cassette Video Edit? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

Oh My Cassette local workflow

Use this skill when the user asks Codex or Claude to edit, cut, caption, reframe, subtitle, combine, polish, add music to, or export video, image, or audio media through Cassette.

The cassette MCP server is a local stdio child process. It opens no port. It sends media and edit requests directly to the separate Cassette backend. Do not start or depend on the repository's FastAPI web-demo server for this workflow.

Courier doctrine

You are a courier between the user and the Cassette agent, not an editor or a brief writer.

  • Pass the user's editing words to cassette_run_job as message VERBATIM — never rewrite, optimize, summarize, translate, or expand them. The Cassette agent is the creative brain; it reads the session's uploaded media itself.
  • Relay the agent's questions and plans back to the user verbatim too. You add only three things: the timeline delta, the version numbers, and the live editor link.
  • Do not call cassette_make_prompt on the API transport — it is a legacy browser-transport brief builder.
  • Never ask upfront about model, thinking level, optimization, or BGM. Defaults match the web editor. Change model/thinking only when the user asks, via cassette_config.

Safety and identity

  • Treat only files inside the active host project roots or explicitly configured media roots as ingestible. If cassette_ingest_media returns source_path_not_allowed, ask the user to move the file into the project or run the private setup command with --allowed-root.
  • Never copy credentials into chat. If an affected tool returns auth_required, show its exact error.details.setup_command as a private terminal command.
  • Keep the returned session_id and job_id. Sessions are isolated by default. Hand a session or job to another host only when the user deliberately asks for a Codex/Claude handoff.
  • Use only paths and resource links returned in artifacts. Never invent an export path or ask the MCP runtime to expose another local file.

Conversational editing (multi-turn)

One session is one continuous conversation with the Cassette agent on one persistent thread — the agent remembers every previous turn, and the editor_url deep link stays the same for the whole session.

  1. Call cassette_ingest_media once for each source asset. Omit session_id on the first call so the runtime generates one, then reuse the returned value.
  2. Call cassette_list_assets and confirm the intended files are present.
  3. For every editing request, call cassette_run_job with message set to the user's verbatim words, the same session_id, and wait omitted or false for the normal background path. Follow-ups like "make that title bigger" need no context restating — the agent remembers the conversation.
  4. A turn ends succeeded with the edit committed and NOTHING rendered: the envelope carries timeline_delta, quality.timeline_ctl, and a contact-sheet artifact — that is the per-turn preview. Relay the delta and name the versions ("v3→v7: trimmed the intro to 4.0s").
  5. Pass export=true on a turn ONLY when the user expresses finish/export intent. That turn ends review_required; evaluate and call cassette_review_completion (only decision=export renders).
  6. If BGM is explicitly requested, use cassette_match_exact_bgm (concrete title/artist), jamendo_music_matcher (configured mood/genre), or cassette_match_bgm (Free To Use fallback) — then continue the conversation.

Model and thinking level

  • cassette_config(session_id) shows the current choice and the available options; cassette_config(session_id, model=…, thinking_level=…) changes them (accepts a product id like deepseek/deepseek-v4-pro or a label like "DeepSeek V4 Pro").
  • The preference persists for the session and applies from the next turn — the same semantics as switching the model between turns in the web editor.
  • Defaults (DeepSeek V4 Flash, low thinking) match the web editor. Do not ask upfront; change only on user request and confirm in one line.

Typed progress handling

Treat the structured phase and next_action fields as authoritative. Do not decide routing, progress, or completion from keywords in prose.

  • running or exporting: call cassette_job_status with wait_for_change_sec=30.
  • needs_user: present the pending question, then call cassette_answer_question with the same job_id and the user's response. On hosts that support MCP elicitation, cassette_job_status may collect the answer itself and return the already-resumed status; treat the returned phase as authoritative and do not re-answer.
  • review_required (export turns only): evaluate the full edit result and call cassette_review_completion. Rendering begins only when the explicit decision is export; use continue, needs_user, or failed when that is the validated outcome.
  • succeeded: the turn is done, nothing rendered — relay the delta + preview and continue the conversation, or re-run with export=true when asked.
  • exported: present validated artifacts and their MCP resource links.
  • failed, cancelled, or timed_out: report the structured error and the runtime-derived next action. A thread_busy error means a run is already live on this session's thread (often started from the open editor tab) — wait and retry.

The named monitoring budget is CASSETTE_MCP_MONITOR_BUDGET_SEC, defaulting to 1500 seconds. Use 30-second long-polls until a phase changes or that elapsed-time budget is reached. If it is still running when the budget expires, return the live job_id and explain that the edit continues in the background. Do not tight-poll.

API jobs persist private thread and interrupt metadata and can resume after Codex or Claude restarts. Browser-transport jobs can resume only while the same MCP process retains the browser session; after restart, surface browser_session_lost and start a new browser job if the user wants to continue.

Timeline grounding and the live editor

  • Every user-visible statement about project state comes from cassette_timeline, never from memory. Name the version in replies: "Quick edit v42→v43: trimmed the intro to 4.0s."
  • Lane routing: when the ask names specific clips or values and needs at most a handful of operations, read cassette_timeline then use cassette_edit (requires CASSETTE_DIRECT_EDIT=1; pass expected_version from the read; a stale_timeline error means re-read and retry). When it needs watching footage, music sync, or a plan, use cassette_run_job.
  • The session carries ONE stable editor_url — a live view of the real editor (timeline, scrubbing preview, plan-review card; zero render) for the entire conversation. Hand it to the user once at the first turn and again at questions/review; on desktop offer to open it (open on macOS, xdg-open on Linux). Do not repeat it on every status poll.
  • cassette_job_status responses carry timeline_delta (cumulative changes since the turn started) and plan_progress; relay the delta rather than re-describing the timeline.
  • Preview escalation, one step per explicit user ask: text digest → contact sheet (cassette_timeline with contact_sheet=true) → the editor_url live view → full export. Never auto-render.
  • Showing previews in a terminal host (Claude Code, Codex CLI): the timeline digest is text — when the user asks to see the timeline, reprint quality.timeline_ctl verbatim in a fenced code block in your reply, never a paraphrase. For the contact sheet or storyboard sheet, terminal hosts cannot render image results inline — print the sheet's contact_sheet_uri / storyboard_sheet_uri (file://…) on its own line: most terminals make it cmd+clickable, opening the real pixels in the system image viewer. On a remote/SSH host where a local file:// link cannot resolve, give the editor_url deep link instead. Optionally also render an at-a-glance version in the terminal with chafa -f symbols --size x when chafa is available. Preview files are swept after ~30 days (CASSETTE_ARTIFACT_TTL_DAYS); exports are never auto-deleted.
  • Plan review: with CASSETTE_PLAN_REVIEW=user (the MCP default) a job pauses with an edit_plan_review question — the plan itself, with each storyboard beat as a readable cell (no raw links). The envelope's quality also carries storyboard (typed beat cells) and storyboard_sheet (a tiled image of one source frame per planned beat — show it when the host can display images). Relay the plan verbatim with the link; answer via cassette_answer_question with approve, revise , or reject. The user may instead decide in the open editor tab: a resume_not_waiting_for_user error means the tab answered first — just re-check status. Note: typing a fresh message in the open editor tab cancels an in-flight plugin turn (the tab takes over) — that is product behavior, not an error to retry.

Cancellation and handoff

  • Call cassette_cancel_job only when the user asks to stop the edit.
  • For a deliberate host handoff, provide the exact session_id and active job_id; the receiving host should begin with cassette_job_status rather than ingesting or starting a duplicate job.
  • Exported files remain under the shared Oh My Cassette data directory. Prefer the returned resource link or file URI rather than relocating the artifact.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet, be the first.

Versions

  • v0.1.0 Imported from the upstream source.