Openclip

Video editing from your agent via the OpenClip MCP server. Use when the user wants to repurpose/clip a long video or URL into shorts/viral moments, transcrib...

OpenClip: video editing from your agent

OpenClip is a video editing MCP server. Its headline pipeline turns a long video into short, captioned, ranked clips, and around that it exposes 28 tools in five groups, identical in every client:

• Clipping pipeline: submit_video, get_video_status, list_clips, render_clip, get_render_status, list_caption_presets, get_transcript, list_supported_providers, list_videos.
• Uploads: create_upload, complete_upload (see the gotcha below: complete_upload starts the PAID pipeline; free tools skip it).
• Free media utilities (account only, NO subscription): transcribe, edit_video, convert_media, extract_thumbnails, edit_image, remove_background, get_tool_job_status.
• UGC generation: generate_ugc, get_ugc_job_status.
• Account & presets: get_account, get_usage, list_agents, create_agent, update_agent, describe_agent_settings, create_agent_logo_upload, set_agent_logo.

Setup (once)

The remote server supports OAuth one-click as the primary path: paste the server URL (https://openclip.app/mcp) into your client and sign in with your OpenClip account, no token to copy.

1. Add the remote server to your agent (server URL: https://openclip.app/mcp):
   • Claude Desktop: Settings → Connectors → Add custom connector → paste https://openclip.app/mcp → Connect → sign in.
   • Claude Code: claude mcp add --transport http openclip https://openclip.app/mcp, then run /mcp and authorize in the browser.
   • Cursor (.cursor/mcp.json): { "mcpServers": { "openclip": { "url": "https://openclip.app/mcp" } } }, Cursor runs OAuth.
2. Manual Bearer token (alternative, for header-only / programmatic clients): generate an MCP token at openclip.app/settings/connect ("Generate MCP token", shown once) and connect to the /mcp/key endpoint with it as a Bearer header:
   • Claude Code: claude mcp add --transport http openclip https://openclip.app/mcp/key --header "Authorization: Bearer <token>"
   • Cursor (.cursor/mcp.json): { "mcpServers": { "openclip": { "url": "https://openclip.app/mcp/key", "headers": { "Authorization": "Bearer <token>" } } } }
   • Any MCP client: POST to https://openclip.app/mcp/key with header Authorization: Bearer <token>.
3. Sanity check: call get_account. Confirm credits_remaining > 0.

(Self-hosting/dev only: run the local stdio server with php artisan mcp:start openclip.)

The loop (processing is ASYNC, always poll)

1. Submit: submit_video(url) → returns a job_id with status:"queued". This ALWAYS succeeds (queues the job) when authenticated, it does NOT pre-check subscription or credits.
2. Poll status: get_video_status(job_id) until status is completed (or failed). Do NOT call list_clips before completion. Read status for the real state:
   • failed (often almost immediately after submit) usually means no active subscription - advise the user to subscribe at openclip.app.
   • pending_credits means the team is out of credits, advise the user to top up.
   • download_failed means the source URL couldn't be fetched, ask the user to check the link.
3. List clips: list_clips(job_id) → viral moments with virality_score (0-10), title/hook, start/end (ms), and clip / rendered_clip URLs.
4. (Optional) Render captioned: render_clip(moment_id, caption_preset). Async too, poll list_clips and watch the moment's rendered_clip for the finished URL. Preset keys: list_caption_presets.

Free media utilities (no subscription needed)

transcribe, edit_video, convert_media, extract_thumbnails, edit_image, and remove_background are FREE: they only need an account, not a subscription or credits. All follow the same async pattern:

1. Get a file in: for a local file call create_upload(filename, content_type), PUT the bytes to the returned upload_url, and use the returned id as file/video. Do NOT call complete_upload for free tools - that starts the paid clipping pipeline; free tools read the uploaded file directly. (Videos already in list_videos work as inputs too.)
2. Call the tool. It returns a tool_job hashid with status: "queued".
3. Poll get_tool_job_status(tool_job) every 5-10s: queued/processing are in-flight, completed/failed terminal. On completed it returns permanent CDN outputs URLs (e.g. json/srt/vtt for transcribe, a zip of frames for extract_thumbnails). On failed it returns error.

Operation cheat sheet (full params in reference.md): edit_video does one operation of crop, trim, rotate, resize, compress, or mute; convert_media targets mp4/webm/mov/mkv, gif, or mp3/aac/wav/flac; edit_image does compress, resize, or crop; remove_background returns a transparent PNG. Free usage is rate-limited per day with per-file size caps; the error text explains how to lift a limit.

Note the transcript split: transcribe is the free tool for uploaded files; get_transcript reads the time-coded transcript of a video that went through the paid clipping pipeline (supports start_ms/end_ms windows at sentence or word level).

UGC generation

generate_ugc renders a short (~7-12s) vertical UGC-style clip from a creative brief (NOT a prose prompt: structured fields like character, or a server-side brief_name preset). Free accounts get ONE generation per day; paid accounts are metered in credits by the seconds rendered. It returns a ugc_job hashid; poll get_ugc_job_status every 10-15s (queued/rendering in-flight, completed returns video_url + result_meta, failed returns error). Renders are heavy: expect a couple of minutes.

Reusable presets (agents)

To make repeat workflows identical (same tracker model, caption preset, composition, logo watermark), save a processing agent once with create_agent and pass it to submit_video. describe_agent_settings documents every nested field and allowed value; update_agent patches only the fields you send; create_agent_logo_upload + set_agent_logo attach a watermark (max 2 MB). get_usage reports the credit balance (credits are MINUTES of processing) and recent activity.

Rules & gotchas

• Always poll; never assume a submit is instantly done. Poll get_video_status in a loop.
• Pick clips by virality_score (higher = better) unless the user says otherwise.
• Times are milliseconds. clip.url is a permanent CDN URL (not signed/expiring).
• submit_video queues regardless of subscription/credit state; the gate surfaces LATER in get_video_status (failed = no subscription, pending_credits = out of credits). Read the status and advise the user accordingly, do NOT expect a synchronous "subscribe" error.
• Projects/folders are not supported by this API, don't promise them.

Error handling

These are MCP tools, not a REST/HTTP API. A failed tool call returns a normal tool result with isError: true and a short plain-text message (no HTTP 401/402/403/422 status, no JSON error envelope). Read the message text and act on it:

• "Not authenticated. Reconnect your OpenClip token.", the connection isn't carrying a valid identity → reconnect. For the OAuth path, re-run the connector sign-in (e.g. /mcp in Claude Code); for the manual /mcp/key path, re-generate the MCP token on the connect page and re-add the server with the new Bearer token. (A token missing the mcp:use scope/ability is rejected at the transport before any tool runs, same fix: reconnect / re-mint.)
• "Job not found." (get_video_status, list_clips), the job_id is wrong, from a different account, or expired → re-check the id you passed; if it was from an earlier session, re-submit the video.
• "Viral moment not found." (render_clip), the moment_id is wrong or belongs to another account → re-fetch moments with list_clips and use an id from that response.
• "Video not yet available. Check get_video_status for progress." (list_clips), the job hasn't produced a video record yet (still queued/downloading) → keep polling get_video_status and only call list_clips once status is completed.
• Validation messages (e.g. a bad/missing url, an unknown caption_preset), the validation error text is returned as the tool message → fix the offending argument and retry.
• Subscription / credits are NOT errors here, they appear as get_video_status status values (failed = no subscription → subscribe; pending_credits = out of credits → top up), not as a tool error. See "The loop" above.

Example prompts → actions

• "Clip this: " → submit_video(url), poll get_video_status, then list_clips.
• "Give me the top 3 moments from job_…" → list_clips, sort by virality_score, take 3.
• "Render moment m_… with the Beast captions" → render_clip(m_…, caption_preset="beast"), poll.
• "Transcribe this recording" → create_upload, PUT bytes, transcribe(video), poll get_tool_job_status (NO complete_upload).
• "Convert this to a gif" / "compress this video" → create_upload, PUT bytes, convert_media(file, to="gif") / edit_video(file, operation="compress"), poll.
• "Remove the background from this image" → create_upload, PUT bytes, remove_background(file), poll, return the transparent PNG.
• "Make a UGC ad for my app" → generate_ugc(brief={...}), poll get_ugc_job_status.

See reference.md for status values, preset keys, operation params, and field shapes.