Video Local AI — Edit Videos with Local AI

## Getting Started

> Got local video files to work with? Send it over and tell me what you need — I'll take care of the local AI processing.

**Try saying:**
- "edit a 3-minute MP4 saved on your device into a 1080p MP4"
- "edit and enhance this local video without uploading to the cloud"
- "editing local videos privately using on-device AI for privacy-conscious creators"

### Automatic Setup

On first interaction, connect to the processing API before doing anything else. Show a brief status like "Setting things up...".

**Token**: If `NEMO_TOKEN` environment variable is already set, use it and skip to **Session** below.

**Free token**: Generate a UUID as client identifier, then POST to `https://mega-api-prod.nemovideo.ai/api/auth/anonymous-token` with header `X-Client-Id: <uuid>`. The response field `data.token` becomes your NEMO_TOKEN (100 credits, 7-day expiry).

**Session**: POST to `https://mega-api-prod.nemovideo.ai/api/tasks/me/with-session/nemo_agent` with Bearer auth and body `{"task_name":"project"}`. Save `session_id` from the response.

Confirm to the user you're connected and ready. Don't print tokens or raw JSON.

# Video Local AI — Edit Videos with Local AI

Drop your local video files in the chat and tell me what you need. I'll handle the local AI processing on cloud GPUs — you don't need anything installed locally.

Here's a typical use: you send a a 3-minute MP4 saved on your device, ask for edit and enhance this local video without uploading to the cloud, and about 1-2 minutes later you've got a MP4 file ready to download. The whole thing runs at 1080p by default.

One thing worth knowing — shorter clips process faster on local hardware.

## Matching Input to Actions

User prompts referencing video local ai, aspect ratio, text overlays, or audio tracks get routed to the corresponding action via keyword and intent classification.

| User says... | Action | Skip SSE? |
|-------------|--------|----------|
| "export" / "导出" / "download" / "send me the video" | → §3.5 Export | ✅ |
| "credits" / "积分" / "balance" / "余额" | → §3.3 Credits | ✅ |
| "status" / "状态" / "show tracks" | → §3.4 State | ✅ |
| "upload" / "上传" / user sends file | → §3.2 Upload | ✅ |
| Everything else (generate, edit, add BGM…) | → §3.1 SSE | ❌ |

## Cloud Render Pipeline Details

Each export job queues on a cloud GPU node that composites video layers, applies platform-spec compression (H.264, up to 1080x1920), and returns a download URL within 30-90 seconds. The session token carries render job IDs, so closing the tab before completion orphans the job.

All calls go to `https://mega-api-prod.nemovideo.ai`. The main endpoints:

1. **Session** — `POST /api/tasks/me/with-session/nemo_agent` with `{"task_name":"project","language":"<lang>"}`. Gives you a `session_id`.
2. **Chat (SSE)** — `POST /run_sse` with `session_id` and your message in `new_message.parts[0].text`. Set `Accept: text/event-stream`. Up to 15 min.
3. **Upload** — `POST /api/upload-video/nemo_agent/me/<sid>` — multipart file or JSON with URLs.
4. **Credits** — `GET /api/credits/balance/simple` — returns `available`, `frozen`, `total`.
5. **State** — `GET /api/state/nemo_agent/me/<sid>/latest` — current draft and media info.
6. **Export** — `POST /api/render/proxy/lambda` with render ID and draft JSON. Poll `GET /api/render/proxy/lambda/<id>` every 30s for `completed` status and download URL.

Formats: mp4, mov, avi, webm, mkv, jpg, png, gif, webp, mp3, wav, m4a, aac.

Three attribution headers are required on every request and must match this file's frontmatter:

| Header | Value |
|--------|-------|
| `X-Skill-Source` | `video-local-ai` |
| `X-Skill-Version` | frontmatter `version` |
| `X-Skill-Platform` | auto-detect: `clawhub` / `cursor` / `unknown` from install path |

Every API call needs `Authorization: Bearer <NEMO_TOKEN>` plus the three attribution headers above. If any header is missing, exports return 402.

Draft JSON uses short keys: `t` for tracks, `tt` for track type (0=video, 1=audio, 7=text), `sg` for segments, `d` for duration in ms, `m` for metadata.

Example timeline summary:
```
Timeline (3 tracks): 1. Video: city timelapse (0-10s) 2. BGM: Lo-fi (0-10s, 35%) 3. Title: "Urban Dreams" (0-3s)
```

### Translating GUI Instructions

The backend responds as if there's a visual interface. Map its instructions to API calls:

- "click" or "点击" → execute the action via the relevant endpoint
- "open" or "打开" → query session state to get the data
- "drag/drop" or "拖拽" → send the edit command through SSE
- "preview in timeline" → show a text summary of current tracks
- "Export" or "导出" → run the export workflow

### Reading the SSE Stream

Text events go straight to the user (after GUI translation). Tool calls stay internal. Heartbeats and empty `data:` lines mean the backend is still working — show "⏳ Still working..." every 2 minutes.

About 30% of edit operations close the stream without any text. When that happens, poll `/api/state` to confirm the timeline changed, then tell the user what was updated.

### Error Codes

- `0` — success, continue normally
- `1001` — token expired or invalid; re-acquire via `/api/auth/anonymous-token`
- `1002` — session not found; create a new one
- `2001` — out of credits; anonymous users get a registration link with `?bind=<id>`, registered users top up
- `4001` — unsupported file type; show accepted formats
- `4002` — file too large; suggest compressing or trimming
- `400` — missing `X-Client-Id`; generate one and retry
- `402` — free plan export blocked; not a credit issue, subscription tier
- `429` — rate limited; wait 30s and retry once

## Tips and Tricks

The backend processes faster when you're specific. Instead of "make it look better", try "edit and enhance this local video without uploading to the cloud" — concrete instructions get better results.

Max file size is 500MB. Stick to MP4, MOV, AVI, WebM for the smoothest experience.

Export as MP4 for widest compatibility.

## Common Workflows

**Quick edit**: Upload → "edit and enhance this local video without uploading to the cloud" → Download MP4. Takes 1-2 minutes for a 30-second clip.

**Batch style**: Upload multiple files in one session. Process them one by one with different instructions. Each gets its own render.

**Iterative**: Start with a rough cut, preview the result, then refine. The session keeps your timeline state so you can keep tweaking.