Agents & MCP · aiclipper.video

Claude Code

claude mcp add --transport http aiclipper https://aiclipper.video/mcp \
  --header "Authorization: Bearer $AICLIPPER_TOKEN"

Cursor

Add to ~/.cursor/mcp.json (or .cursor/mcp.json in a project):

{
  "mcpServers": {
    "aiclipper": {
      "url": "https://aiclipper.video/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

VS Code

Add to .vscode/mcp.json:

{
  "servers": {
    "aiclipper": {
      "type": "http",
      "url": "https://aiclipper.video/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Claude Desktop and other stdio-only clients

Bridge the remote server with mcp-remote in the client’s MCP config:

{
  "mcpServers": {
    "aiclipper": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://aiclipper.video/mcp",
        "--header",
        "Authorization:${AICLIPPER_AUTH}"
      ],
      "env": {
        "AICLIPPER_AUTH": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Tools

Every workspace feature is designed to be available to agents. Tools act only on the account that owns the token.

list_formats read-only

List the video formats (aspect ratios and layouts) clips can be rendered in, with the platforms each one suits.

get_account read-only

Show the signed-in account: remaining AI credits, credit costs, and project limits. Call estimate_credits to price a specific video before create_clips.

export_account_data read-only

Export the signed-in account's own data as one JSON object: profile, projects, clips with scores and descriptions, channels, credit purchases, AI credit usage, agent token names, and connected accounts. Read-only. Excludes video files, full transcript text, and token secrets.

estimate_credits read-only

Estimate the AI credits a project will cost for a video of a given length, and whether the balance covers the start. required_now is transcription (1 credit per started minute) plus clip selection (1 credit); create_clips fails right away when the balance is below it. Translation (1 credit per language) and dubbing (5 credits per dubbed clip) are charged later as clips render. Preview covers are a frame from the video with the headline and cost nothing, so covers is always 0. total is the sum. Clip counts are upper bounds; clip_count "auto" is priced at its upper bound (up to 12 clips, fewer when many formats are rendered). Read-only: nothing is charged.

create_upload

Get a signed URL to upload a local video file (MP4, MOV, WebM, or MKV, up to 2 GB, 15 seconds to 2 hours) straight to storage. PUT the file bytes to url with exactly the returned headers (for example with curl --upload-file), then call create_clips with upload_id. Uploads are the default supported workflow. The URL expires in one hour. An upload that is not turned into a project with create_clips is deleted automatically about a day after it was created.

create_brand_logo_upload

Get a signed PUT URL for a PNG, JPEG, or WebP logo up to 5 MB. Upload bytes with returned headers, then pass upload_id as brand_logo_upload_id to create_clips. Only this account's unused completed logo uploads are accepted.

create_clips

Start a project from an uploaded file (upload_id from create_upload): it transcribes the video, picks the strongest moments, and renders captioned clips in the chosen formats. Processing runs in the background and usually takes a few minutes: poll get_project with the returned project_id until status is completed or failed. Uses AI credits. Credits are checked before any transcription starts: if the balance cannot cover transcription and clip selection, the project fails at once and error_message says how many credits are needed (buy credits, then retry_project). Call estimate_credits first to see the cost.

list_channels read-only

List the connected YouTube channels with their automation state (active or paused), preferences, last check, and last error.

list_projects read-only

List projects newest first using stable signed cursor pagination. Returns projects and next_cursor (null at the end). Pass next_cursor unchanged with the same status filter to read another page.

get_project read-only

Get project status and clips with titles, scores, timestamps, social descriptions, one-hour download links, and read-only publications (provider, status, external_id, url, published_at, error, operation_stage). Only published confirms publication; uncertain requires checking the provider before retrying.

get_clip read-only

Get a clip's score breakdown, rationale, social description, video/caption/cover links, and read-only publications (provider, status pending/published/failed/uncertain, external_id, url, published_at, error, operation_stage). No resumable operation secrets are exposed.

get_transcript read-only

Get a project's timestamped transcript, for example to write posts or pick a custom moment.

correct_transcript destructive

Replace a finished project's transcript with ordered timestamped segments and rerender original-audio clips for free. Dubbed clips are unchanged. Requires the retained original source. No AI credits are charged.

reattach_source

Reattach an original video after retention cleanup without creating a new project or charging AI credits. Upload the same original file with create_upload first. Does not start processing; retry the desired edit afterwards. Requires explicit confirmation of rights.

add_format

Render a finished project's selected moments in another format. Reuses the source video and transcript, so it does not re-transcribe.

reselect_moments destructive

Run moment selection again on a finished project's saved transcript with new settings (goal, clip_count, min_score, clip_length). Existing clips remain available until all replacement clips render successfully, then are replaced together. Transcription is not repeated, so only clip selection (1 credit) is charged again, plus any translation and dubbing for target languages. Needs the original video, which is kept 14 days after the project finishes.

retry_project

Queue a failed project again. Completed steps such as download and transcription are reused.

update_clip

Edit a clip's title and social description. Omitted fields are kept; an empty social_description clears the text so the next render restores the automatically generated one.

recut_clip

Re-render a clip with new start and end seconds in the original video, reusing the transcript. The cut is exact, 5 to 120 seconds, and rendering is free. Dubbed copies cannot be recut (recut the original); the original video is needed (kept 14 days after the project finishes).

delete_clip destructive

Permanently delete one clip and its video, captions, and cover. The project and its other clips are kept. Only allowed when the project is not processing.

generate_hook_cover

Create a new preview cover for a clip from a frame of its video with the hook text. Covers are free: mode "frame" is the only cover mode. An old client that sends "ai" gets a frame cover.

generate_social_copy

Rewrite a clip's social description from its transcript with AI (1 credit): a hook first line, a short setup, and hashtags, in the clip's language. The previous text is replaced; edit it afterwards with update_clip.

publish_clip

Publish a ready clip to the user's connected channel: YouTube Shorts (square or vertical clips up to 3 minutes), TikTok (vertical, under 3 minutes), or Instagram Reels (9:16, 5-90 seconds). The channel must be connected first in Settings → Publishing (web only). The clip's social_description is posted with the video; TikTok and Instagram fetch it through the clip's public share link, which is turned on automatically. Publishing the same clip to the same provider again returns its existing publication.

share_clip

Turn a public share link for a ready clip on or off. On, returns a URL anyone can open to watch and download the clip without an account; off, the link stops working immediately but is restored (same URL) when sharing is turned on again.

rate_clip

Record whether a clip worked: up or down, with an optional short note. Ratings stay in the workspace and help tune moment selection.

list_credit_packs read-only

List the one-time AI credit packs for sale (credits, USD price) and the current balance. Packs are not subscriptions.

buy_credits

Create a secure Creem checkout link for a credit pack. The agent cannot pay: give the checkout_url to the user to open and pay in their browser. Credits are added automatically after payment; check get_account for the new balance.

clear_unfinished_uploads destructive

Cancel and remove all your unfinished, unattached extension stream uploads to free upload slots. This may interrupt active uploads in other tabs. Ask the user before clearing. Completed uploads, project attachments, and other users' uploads are preserved. Returns cleared and failed counts; storage failures retain their slots and can be retried. No AI credits are charged.

delete_project destructive

Permanently delete a project and all of its clips and media. Only allowed when the project is not processing.

Skills

The MCP server gives your agent the tools. The aiclipper skill teaches it the workflow: upload the file, start the project, wait for rendering, and report back the clips. Install it once and just ask: “clip this video for Shorts”. It is a plain SKILL.md file, so it works anywhere agents read skills or instructions.

Claude Code

mkdir -p ~/.claude/skills/aiclipper
curl -fsSL https://aiclipper.video/skills/aiclipper/SKILL.md \
  -o ~/.claude/skills/aiclipper/SKILL.md

Use .claude/skills/aiclipper/ inside a project to share it with your team. It loads when you ask to clip a video.

Cursor, Codex, and other agents

Copy the file below into your agent’s rules or instructions file (for example .cursor/rules/aiclipper.mdc or AGENTS.md), or point the agent at https://aiclipper.video/skills/aiclipper/SKILL.md.

aiclipper/SKILL.md

Download

---
name: aiclipper
description: Turn the user's own long video (a local file or their YouTube video) into short captioned social clips with aiclipper.video. Use when the user asks to clip, cut highlights from, or repurpose a video for TikTok, YouTube Shorts, Instagram Reels, LinkedIn, or X.
---

# aiclipper.video

Creates short, captioned clips from long videos through the `aiclipper` MCP server
(https://aiclipper.video/agents). The AI finds the strongest moments, reframes them for each
feed, burns in captions, and writes social copy. Processing takes a few minutes and uses the
user's AI credits.

## Before you start

- The `aiclipper` MCP server must be connected (endpoint `https://aiclipper.video/mcp`, header
  `Authorization: Bearer <agent token>`). If its tools are missing, tell the user to create a
  token in Settings → Agent access and follow https://aiclipper.video/agents.
- Only process videos the user owns or has permission to use. `create_clips` requires
  `rights_confirmed: true`: ask the user to confirm ownership or permission first, and never set it
  on their behalf without asking.
- Call `get_account` first if you are unsure the user has credits. New projects need credits.
  If they run out, `list_credit_packs` shows the one-time packs and `buy_credits` returns a
  checkout link the user opens and pays in their browser (you cannot pay). Credits arrive after
  payment.
- Before creating a project, call `estimate_credits` with the video's length in minutes (and the
  formats, `target_languages`, or `cover_mode` you plan to use). `required_now` is what the project
  needs to start (transcription plus clip selection) and `enough` says whether the balance covers it.
  Tell the user the `total` before you start. If `enough` is false, `create_clips` would fail at once,
  so buy credits first.

## Clip a local video file (preferred)

Upload original files. YouTube URL ingestion, channel automation, and the browser extension are disabled by default; use those paths only when tools/list advertises their inputs or tools.

1. Work out the file facts:
   - size in bytes: `wc -c < FILE`
   - MIME type from the extension: `.mp4` → `video/mp4`, `.mov` → `video/quicktime`,
     `.webm` → `video/webm`, `.mkv` → `video/x-matroska`
   - Base64 MD5: `openssl md5 -binary FILE | base64`
   - Limits: 2 GB, 15 seconds to 2 hours, must contain audio.
   - Source resolution is retained; YouTube downloads and extension SABR imports select up to 2160p (4K). Exports keep existing format dimensions (up to 1080p), with H.264 CRF 18. A 4K source improves crop detail; it does not enable 4K exports.
2. Call `create_upload` with `filename`, `byte_size`, `content_type`, `md5_base64`.
3. PUT the file to the returned `url` with exactly the returned `headers` (Content-Type,
   Content-MD5, Content-Disposition) and the file bytes as the body. The `example` field is a
   ready-made curl command. The upload must return HTTP 200. The URL expires in one hour.
4. **The upload is not a project yet.** Call `create_clips` with `upload_id` (do not also pass
   `source_url`) and `rights_confirmed: true` to start processing. An upload that is not turned into a project with create_clips is deleted automatically about a day after it was created.

For separate streams already uploaded by the browser extension, pass the completed video stream's
ID as `upload_id` and the completed audio stream's ID as `audio_upload_id`. Both uploads must belong
to the signed-in user, be unused, and total no more than 2 GB. `audio_upload_id` requires `upload_id`;
do not pair audio with `source_url`. Ordinary local video files need no separate audio upload.

Server YouTube URL/channel imports are disabled in the current deployment. Use local uploads or completed extension stream uploads; `extension.enabled` is independent of the server `youtube.ingestion_enabled` gate.

## Clip a YouTube video the user owns

Only when `create_clips` advertises `source_url`, pass a single owned video URL and
`rights_confirmed: true`. Default deployments accept uploads only. Never download someone else's
video without permission. If URL ingestion fails, report the error and request the original file;
do not repeatedly retry a blocked link or promise the extension can bypass restrictions.

## Choose formats

Call `list_formats` for details. Pass `formats` as an array of ids:

| id | Ratio | Best for |
|---|---|---|
| `punch` (default) | 9:16 full-bleed crop | TikTok, Shorts, Reels |
| `studio` | 9:16 centered fit | podcasts and interviews with several people |
| `square` | 1:1 | LinkedIn, X, Instagram feed |
| `portrait` | 4:5 | Instagram and Facebook feeds |
| `cinema` | 16:9 | YouTube, X, Facebook Watch |

Other useful `create_clips` options: `clip_count` (an integer 1-5, or `"auto"`; default 3), `clip_length`
(15-120 seconds, the maximum length, default 30), `goal` (`virality`, `education`, `storytelling`), `caption_style`
(`pop` karaoke highlight, `clean` sentence case, or `impact` one big word at a time — ask the user which look they
want; colors and size follow via `caption_color`, `caption_highlight_color`, and `caption_size`),
`cover_mode` (`frame` is free and
makes the cover from a frame of the video with your headline; `none` skips it), `target_languages` for dubbed versions.

Branding: pass `watermark_text` (up to 60 characters), hex `watermark_color`, and
`watermark_position` (use the advertised enum). For logos, call `create_brand_logo_upload` with
filename, byte_size, content_type, and md5_base64 (PNG/JPEG/WebP, at most 5 MB), PUT bytes using
returned headers, then pass its `upload_id` as `brand_logo_upload_id` to `create_clips`. The logo
must belong to this account, be fully uploaded, and not already attached elsewhere.

`clip_count: "auto"` lets the AI decide how many clips the video yields: it picks every strong, self-contained
moment and sizes each clip to its natural start and end, up to `clip_length`. Auto stops at 12 clips per project
(fewer when many formats are rendered) and can return fewer when the video has few strong moments. Use it when
the user asks for "all the good moments" or does not name a number. Clips are scored by AI editorial estimates,
not predicted views, so say that rather than promising a clip will go viral. `estimate_credits` prices Auto at its
upper bound and says "up to".

## Wait for the result

1. `create_clips` returns a `project_id`. Poll `get_project` every 20-30 seconds until `status`
   is `completed` or `failed`. Statuses go queued → downloading → transcribing → analyzing →
   rendering → completed.
2. If `failed`, report `error_message` exactly. For a failed project, `retry_project` re-queues it
   and reuses finished steps.
3. When completed, give the user each clip's `title`, `score`, `format`, `social_description`,
   and `video_url`. Download links expire after one hour: call `get_project` or `get_clip` again
   for fresh ones.

## Publish, share, and refine

- `publish_clip` uploads a ready clip to the user's connected channel: YouTube Shorts (square or
  vertical, under 3 minutes), TikTok (vertical, under 3 minutes), or Instagram Reels (9:16, 5-90
  seconds) — pass the matching `provider`. The user must connect the channel in Settings →
  Publishing first; if it is not connected, tell them to. TikTok and Instagram fetch the video
  through the clip's public share link, which is turned on automatically. Read `publications` in
  `get_clip` or `get_project` to follow pending, published, failed, or uncertain status. Only
  published confirms success; uncertain requires checking the provider before retrying, to avoid
  duplicate posts. Read-only fields include provider, status, external_id, url, published_at,
  error, and operation_stage. Resumable credentials and operation IDs are never exposed.
- `share_clip` turns a public link (`/s/<token>`) on or off for a clip. Give the URL to the user for
  feedback or approvals; anyone with it can watch and download the clip.
- `update_clip` renames a clip or rewrites its description. `generate_social_copy` has AI rewrite
  the description from the clip's transcript (1 credit) — offer it when the user wants punchier copy.
- `recut_clip` re-renders a clip with new start and end seconds (5 to 120, inside the video).
  Rendering is free; dubbed copies cannot be recut. Existing media stays live until the new render succeeds.
  The web clip page offers the same operation with a visual timeline, keyboard handles, and
  selection preview when the source is retained; agents pass exact timestamps via `recut_clip`.
- `reselect_moments` re-runs moment selection on a finished project with new settings (goal, count,
  minimum score). Old clips remain until replacements render successfully; only selection (1 credit) is charged again.
- `rate_clip` records whether a clip worked (up or down, optional note). Ask the user after they
  review clips; ratings tune future selection.
- `delete_clip` permanently removes one clip (confirm with the user first).

## Channels

`list_channels` remains available for existing connections. `create_channel`, `update_channel`,
and `sync_channel` are advertised only when the operator explicitly enables YouTube ingestion.
When offered, channel creation requires ownership/permission confirmation and
`rights_confirmed: true`; preferences apply to future videos. Do not offer disabled automation.

## More tools

- `add_format` renders finished moments in another format without re-transcribing. It,
  `reselect_moments`, `recut_clip`, and `retry_project` need original video, removed after 14 days.
- `reattach_source` restores a purged source without new project or AI charges: call `create_upload`,
  upload the exact original combined file, then pass `project_id`, `upload_id`, and explicit
  `rights_confirmed: true`. It does not start processing; retry the desired edit afterwards.
  Exact saved fingerprint, duration, video and audio must match. Alternate encodings, separate
  streams, and legacy projects without saved fingerprints are rejected; create a new project instead.
- `correct_transcript` takes `project_id` and `segments` (1–10000 objects containing only numeric
  `start`, numeric `end`, and nonblank `text` up to 10000 characters). Segments must be ordered,
  non-overlapping, and inside source duration, with end greater than start. Requires finished
  processing and retained original source. It rerenders original-audio clips free; dubbed clips
  are not retranscribed or redubbed. Poll `get_project` after correction.
- `correct_transcript`, `recut_clip`, and `reselect_moments` reject projects with pending or
  uncertain publications. Wait for publishing or resolve its uncertain outcome first. Edits keep
  existing ready clips until replacements finish; failed replacements do not discard originals.
  `add_format` also waits until any staged project edit finishes.
- `get_clip` returns score breakdown, rationale, captions, cover links, and publication states.
- `get_transcript` returns timestamped transcript segments for posts or correction.
- `generate_hook_cover` makes a new preview cover from a frame of the clip's video with your hook text
  (`mode: "frame"`, free). To change the headline, pass different `custom_hook` text.
- `list_projects` finds earlier projects with optional `status`, `limit` (integer 1–100, default 20),
  and signed opaque `cursor`. Pass returned `next_cursor` unchanged with the same filter to get
  the next stable newest-first page; null means no more. Invalid cursors return errors.
- `delete_project` is permanent: confirm with the user first.
- `export_account_data` returns the user's own data as JSON (projects, clips, credit purchases,
  AI usage). It is read-only. Account deletion is not available to agents; point the user to
  Settings in the web app.

## Rules

- Stop and report the exact error if any step fails. Do not guess or silently change approach.
- Never claim clips exist until `get_project` shows `completed` and returns clips.
- Do not put the agent token in chat, files, or commits.

---
name: aiclipper
description: Turn the user's own long video (a local file or their YouTube video) into short captioned social clips with aiclipper.video. Use when the user asks to clip, cut highlights from, or repurpose a video for TikTok, YouTube Shorts, Instagram Reels, LinkedIn, or X.
---

# aiclipper.video

Creates short, captioned clips from long videos through the `aiclipper` MCP server
(https://aiclipper.video/agents). The AI finds the strongest moments, reframes them for each
feed, burns in captions, and writes social copy. Processing takes a few minutes and uses the
user's AI credits.

## Before you start

- The `aiclipper` MCP server must be connected (endpoint `https://aiclipper.video/mcp`, header
  `Authorization: Bearer <agent token>`). If its tools are missing, tell the user to create a
  token in Settings → Agent access and follow https://aiclipper.video/agents.
- Only process videos the user owns or has permission to use. `create_clips` requires
  `rights_confirmed: true`: ask the user to confirm ownership or permission first, and never set it
  on their behalf without asking.
- Call `get_account` first if you are unsure the user has credits. New projects need credits.
  If they run out, `list_credit_packs` shows the one-time packs and `buy_credits` returns a
  checkout link the user opens and pays in their browser (you cannot pay). Credits arrive after
  payment.
- Before creating a project, call `estimate_credits` with the video's length in minutes (and the
  formats, `target_languages`, or `cover_mode` you plan to use). `required_now` is what the project
  needs to start (transcription plus clip selection) and `enough` says whether the balance covers it.
  Tell the user the `total` before you start. If `enough` is false, `create_clips` would fail at once,
  so buy credits first.

## Clip a local video file (preferred)

Upload original files. YouTube URL ingestion, channel automation, and the browser extension are disabled by default; use those paths only when tools/list advertises their inputs or tools.

1. Work out the file facts:
   - size in bytes: `wc -c < FILE`
   - MIME type from the extension: `.mp4` → `video/mp4`, `.mov` → `video/quicktime`,
     `.webm` → `video/webm`, `.mkv` → `video/x-matroska`
   - Base64 MD5: `openssl md5 -binary FILE | base64`
   - Limits: 2 GB, 15 seconds to 2 hours, must contain audio.
   - Source resolution is retained; YouTube downloads and extension SABR imports select up to 2160p (4K). Exports keep existing format dimensions (up to 1080p), with H.264 CRF 18. A 4K source improves crop detail; it does not enable 4K exports.
2. Call `create_upload` with `filename`, `byte_size`, `content_type`, `md5_base64`.
3. PUT the file to the returned `url` with exactly the returned `headers` (Content-Type,
   Content-MD5, Content-Disposition) and the file bytes as the body. The `example` field is a
   ready-made curl command. The upload must return HTTP 200. The URL expires in one hour.
4. **The upload is not a project yet.** Call `create_clips` with `upload_id` (do not also pass
   `source_url`) and `rights_confirmed: true` to start processing. An upload that is not turned into a project with create_clips is deleted automatically about a day after it was created.

For separate streams already uploaded by the browser extension, pass the completed video stream's
ID as `upload_id` and the completed audio stream's ID as `audio_upload_id`. Both uploads must belong
to the signed-in user, be unused, and total no more than 2 GB. `audio_upload_id` requires `upload_id`;
do not pair audio with `source_url`. Ordinary local video files need no separate audio upload.

Server YouTube URL/channel imports are disabled in the current deployment. Use local uploads or completed extension stream uploads; `extension.enabled` is independent of the server `youtube.ingestion_enabled` gate.

## Clip a YouTube video the user owns

Only when `create_clips` advertises `source_url`, pass a single owned video URL and
`rights_confirmed: true`. Default deployments accept uploads only. Never download someone else's
video without permission. If URL ingestion fails, report the error and request the original file;
do not repeatedly retry a blocked link or promise the extension can bypass restrictions.

## Choose formats

Call `list_formats` for details. Pass `formats` as an array of ids:

| id | Ratio | Best for |
|---|---|---|
| `punch` (default) | 9:16 full-bleed crop | TikTok, Shorts, Reels |
| `studio` | 9:16 centered fit | podcasts and interviews with several people |
| `square` | 1:1 | LinkedIn, X, Instagram feed |
| `portrait` | 4:5 | Instagram and Facebook feeds |
| `cinema` | 16:9 | YouTube, X, Facebook Watch |

Other useful `create_clips` options: `clip_count` (an integer 1-5, or `"auto"`; default 3), `clip_length`
(15-120 seconds, the maximum length, default 30), `goal` (`virality`, `education`, `storytelling`), `caption_style`
(`pop` karaoke highlight, `clean` sentence case, or `impact` one big word at a time — ask the user which look they
want; colors and size follow via `caption_color`, `caption_highlight_color`, and `caption_size`),
`cover_mode` (`frame` is free and
makes the cover from a frame of the video with your headline; `none` skips it), `target_languages` for dubbed versions.

Branding: pass `watermark_text` (up to 60 characters), hex `watermark_color`, and
`watermark_position` (use the advertised enum). For logos, call `create_brand_logo_upload` with
filename, byte_size, content_type, and md5_base64 (PNG/JPEG/WebP, at most 5 MB), PUT bytes using
returned headers, then pass its `upload_id` as `brand_logo_upload_id` to `create_clips`. The logo
must belong to this account, be fully uploaded, and not already attached elsewhere.

`clip_count: "auto"` lets the AI decide how many clips the video yields: it picks every strong, self-contained
moment and sizes each clip to its natural start and end, up to `clip_length`. Auto stops at 12 clips per project
(fewer when many formats are rendered) and can return fewer when the video has few strong moments. Use it when
the user asks for "all the good moments" or does not name a number. Clips are scored by AI editorial estimates,
not predicted views, so say that rather than promising a clip will go viral. `estimate_credits` prices Auto at its
upper bound and says "up to".

## Wait for the result

1. `create_clips` returns a `project_id`. Poll `get_project` every 20-30 seconds until `status`
   is `completed` or `failed`. Statuses go queued → downloading → transcribing → analyzing →
   rendering → completed.
2. If `failed`, report `error_message` exactly. For a failed project, `retry_project` re-queues it
   and reuses finished steps.
3. When completed, give the user each clip's `title`, `score`, `format`, `social_description`,
   and `video_url`. Download links expire after one hour: call `get_project` or `get_clip` again
   for fresh ones.

## Publish, share, and refine

- `publish_clip` uploads a ready clip to the user's connected channel: YouTube Shorts (square or
  vertical, under 3 minutes), TikTok (vertical, under 3 minutes), or Instagram Reels (9:16, 5-90
  seconds) — pass the matching `provider`. The user must connect the channel in Settings →
  Publishing first; if it is not connected, tell them to. TikTok and Instagram fetch the video
  through the clip's public share link, which is turned on automatically. Read `publications` in
  `get_clip` or `get_project` to follow pending, published, failed, or uncertain status. Only
  published confirms success; uncertain requires checking the provider before retrying, to avoid
  duplicate posts. Read-only fields include provider, status, external_id, url, published_at,
  error, and operation_stage. Resumable credentials and operation IDs are never exposed.
- `share_clip` turns a public link (`/s/<token>`) on or off for a clip. Give the URL to the user for
  feedback or approvals; anyone with it can watch and download the clip.
- `update_clip` renames a clip or rewrites its description. `generate_social_copy` has AI rewrite
  the description from the clip's transcript (1 credit) — offer it when the user wants punchier copy.
- `recut_clip` re-renders a clip with new start and end seconds (5 to 120, inside the video).
  Rendering is free; dubbed copies cannot be recut. Existing media stays live until the new render succeeds.
  The web clip page offers the same operation with a visual timeline, keyboard handles, and
  selection preview when the source is retained; agents pass exact timestamps via `recut_clip`.
- `reselect_moments` re-runs moment selection on a finished project with new settings (goal, count,
  minimum score). Old clips remain until replacements render successfully; only selection (1 credit) is charged again.
- `rate_clip` records whether a clip worked (up or down, optional note). Ask the user after they
  review clips; ratings tune future selection.
- `delete_clip` permanently removes one clip (confirm with the user first).

## Channels

`list_channels` remains available for existing connections. `create_channel`, `update_channel`,
and `sync_channel` are advertised only when the operator explicitly enables YouTube ingestion.
When offered, channel creation requires ownership/permission confirmation and
`rights_confirmed: true`; preferences apply to future videos. Do not offer disabled automation.

## More tools

- `add_format` renders finished moments in another format without re-transcribing. It,
  `reselect_moments`, `recut_clip`, and `retry_project` need original video, removed after 14 days.
- `reattach_source` restores a purged source without new project or AI charges: call `create_upload`,
  upload the exact original combined file, then pass `project_id`, `upload_id`, and explicit
  `rights_confirmed: true`. It does not start processing; retry the desired edit afterwards.
  Exact saved fingerprint, duration, video and audio must match. Alternate encodings, separate
  streams, and legacy projects without saved fingerprints are rejected; create a new project instead.
- `correct_transcript` takes `project_id` and `segments` (1–10000 objects containing only numeric
  `start`, numeric `end`, and nonblank `text` up to 10000 characters). Segments must be ordered,
  non-overlapping, and inside source duration, with end greater than start. Requires finished
  processing and retained original source. It rerenders original-audio clips free; dubbed clips
  are not retranscribed or redubbed. Poll `get_project` after correction.
- `correct_transcript`, `recut_clip`, and `reselect_moments` reject projects with pending or
  uncertain publications. Wait for publishing or resolve its uncertain outcome first. Edits keep
  existing ready clips until replacements finish; failed replacements do not discard originals.
  `add_format` also waits until any staged project edit finishes.
- `get_clip` returns score breakdown, rationale, captions, cover links, and publication states.
- `get_transcript` returns timestamped transcript segments for posts or correction.
- `generate_hook_cover` makes a new preview cover from a frame of the clip's video with your hook text
  (`mode: "frame"`, free). To change the headline, pass different `custom_hook` text.
- `list_projects` finds earlier projects with optional `status`, `limit` (integer 1–100, default 20),
  and signed opaque `cursor`. Pass returned `next_cursor` unchanged with the same filter to get
  the next stable newest-first page; null means no more. Invalid cursors return errors.
- `delete_project` is permanent: confirm with the user first.
- `export_account_data` returns the user's own data as JSON (projects, clips, credit purchases,
  AI usage). It is read-only. Account deletion is not available to agents; point the user to
  Settings in the web app.

## Rules

- Stop and report the exact error if any step fails. Do not guess or silently change approach.
- Never claim clips exist until `get_project` shows `completed` and returns clips.
- Do not put the agent token in chat, files, or commits.

Example

You: Clip ~/Recordings/episode-48.mp4 into 3 vertical clips for Shorts and a square one for LinkedIn.

Agent: create_upload(filename, byte_size, content_type, md5_base64)
       curl -X PUT --upload-file episode-48.mp4 … (signed URL, straight to storage)
       create_clips(upload_id, formats: ["punch", "square"], clip_count: 3)
       get_project(project_id) … status: completed
       → 6 clips with titles, viral scores, captions, social copy, and download links

Security and limits

Tokens are stored hashed, can be revoked any time in Settings, and show when they were last used.
Download links are signed and expire after one hour. Call get_project or get_clip again for fresh ones.
Original source videos are deleted 14 days after processing finishes. Download them before then; generated clips, captions, covers, and transcripts remain until you delete the project.
Agents share your account’s credits and limits: 20 projects per day, 3 processing at once, 120 MCP requests per minute per token.
Only process videos you own or have permission to use. See the Terms of Service.

Questions or a client that will not connect? Email support@aiclipper.video.

Clip from any agent.

Create an agent token

Add the server

Ask for clips

Claude Code

Cursor

VS Code

Claude Desktop and other stdio-only clients

Tools

Skills

Claude Code

Cursor, Codex, and other agents

Example

Security and limits