MCP Server
Connect Claude Desktop and AI coding agents to Videoslash via the Model Context Protocol.
MCP Server
The Videoslash MCP server exposes your account, scripts, Creative Studio, AI Lab tools, and the Video Editor through the Model Context Protocol (MCP). Use it from Claude Desktop / claude.ai Connectors (OAuth) or from Claude Code, Cursor, Codex CLI, and OpenCode (API key).
Server URL
Use the www host — the apex domain can redirect with an HTTPS→HTTP downgrade that MCP clients reject.
https://www.videoslash.com/api/mcpAuthentication
Claude Desktop / claude.ai
Add a custom connector with the URL above, click Connect, and Allow on Videoslash. No API key is pasted into Claude. See Configuration.
Coding agents (API key)
Create a key in Account settings under Developers and pass it as a Bearer token. Each key's scopes control which tools the agent can see:
| Scope | Tools |
|---|---|
read | Account, credits, Creative Studio model catalog, AI Lab tool list, price estimates and run status |
scripts:generate | Title and script generation |
creative:generate | Creative Studio generations, AI Lab runs (upscale_video, lipsync_video, …) and media uploads |
editor:read | Video Editor inspect tools (list_projects, get_timeline, list_assets, …) |
editor:write | Video Editor mutations (add_clips, split_clip, generate_image, export_project, …). Implies read. |
A key only lists tools it is allowed to use. OAuth Connectors receive full MCP account access instead (including the Video Editor).
Each tool declares MCP annotations: a title; readOnlyHint: true and idempotentHint: true for tools that only read; destructiveHint: true only for tools that delete or overwrite data (delete_…, remove_…, replace_…), while tools that create or add data declare destructiveHint: false. Every tool declares openWorldHint: false. In Claude, read-only tools can run without a confirmation prompt; tools that change data prompt.
Quick start
- Claude Desktop: Add connector URL → Connect → Allow.
- Coding agents: Create an API key, then follow Configuration.
- Ask the agent to check credits, list models, or start a generation.
Example prompts:
- "Check my Videoslash credit balance."
- "List fast video models and generate a 5-second product clip."
- "Generate a script title about ancient civilizations."
- "Upscale this clip to 4K at 60 fps."
- "List my Video Editor projects and split the A-roll at 3 seconds."
Workflow examples
Chain a prior generation as input
- Call
create_creative_generationto produce an image. - Poll
get_creative_generationuntilstatusiscompleted. - Pass the returned
outputUrlinmediaUrlswhen callingcreate_creative_generationagain with a video model.
The agent can reuse any completed CDN outputUrl from list_creative_generations the same way.
Compare models on the same prompt
- Call
list_creative_modelswith filters (for examplecategory: Fast). - Call
create_creative_generationmultiple times with differentmodelIdvalues and the same prompt. - Poll each generation and compare the
outputUrlresults.
AI Lab tools
Every AI Lab tool from the dashboard has its own create tool: recast_video, reshoot_video, erase_from_video, extend_video, edit_video, recreate_video, create_thumbnail, upscale_video, lipsync_video, change_video_voice, translate_video, edit_image_background and add_video_sound. Helpers: list_ai_lab_tools (tools, inputs and choices), estimate_ai_lab_run (price quote) and get_ai_lab_run (status). In Claude the AI Lab widget collects the files; coding agents upload them first. See AI Lab tools.
Plans and credits
Generation tools (create_creative_generation and every AI Lab create tool) need an active plan, the same as on the dashboard; without one they return subscription_required with a billing link. Pay-per-use credits alone don't unlock generation. When credits run short, tools return insufficient_credits with advice (the pack that covers the shortfall, or when plan credits refill). get_credits returns the balances plus the plan, refill date and the credit packs the user can buy.
Async generations
Creative Studio tools mirror the REST API: create_creative_generation returns immediately with a generation id. Poll get_creative_generation until status is completed or failed. Completed outputs expose a CDN URL on cdn.videoslash.com.
AI Lab create tools return a run with its outputs processing. Poll get_ai_lab_run every 10–20 seconds until status is no longer processing.
Editor generate_image / generate_video / add_captions follow the same pattern: they return a jobId with status=processing. Poll poll_editor_job every 5–10 seconds. Cloud exports return a job id; poll get_export_job.
Claude Connectors inline previews (MCP Apps)
On Claude Desktop / claude.ai Connectors, create tools mount an MCP App widget next to the tool call. Only create tools mount a widget: the get, list and poll tools return plain results.
| Widget | Resource URI | Mounted by |
|---|---|---|
| Creative Studio | ui://videoslash/creative-generation@v15 | create_creative_generation |
| AI Lab | ui://videoslash/ai-lab@v1 | Every AI Lab create tool (upscale_video, lipsync_video, …) |
The version suffix changes whenever the widget changes, so Claude fetches the new one. Reconnect the connector if Claude still shows an old widget.
The Creative Studio widget shows:
- An aspect-ratio Generating card while the job runs
- Inline image or video when
outputUrlis ready - Recreate to rerun the same model, prompt, and parameters
This uses the official MCP Apps extension (sandboxed HTML iframe + _meta.ui.resourceUri). It is not automatic URL rendering from tool text.
Claude Code, Cursor, Codex, and OpenCode do not mount MCP App iframes — they still show JSON text with outputUrl. Behavior is unchanged for API-key clients.
The AI Lab widget collects the tool's files from the user, shows its options and price, then shows progress and results. See AI Lab tools.
See ADR 0008 for implementation details.
Local file uploads
In Claude Connectors, the widgets collect attached files from the user and upload them from the browser; Claude doesn't upload them itself.
Coding agents have two ways to upload a local file. Both return an uploadId:
- Presigned PUT (preferred, any size up to the upload limits): call
request_creative_media_uploadwith the file metadata, then upload the bytes withcurl -Tto the returneduploadUrlusingrequiredHeaders. The bytes go straight to storage, not through MCP. - Inline base64 (small files): call
upload_creative_mediawith the file as base64, up to 8MB decoded. The bytes travel in the tool call, so pass the base64 only as the tool argument and never print it into the conversation.
Pass the uploadId in mediaUploads on create_creative_generation, or in the file field of an AI Lab create tool (for example video on upscale_video).
See the Creative Studio API guide for the same presigned upload flow over REST.
Editor library files use request_editor_upload → PUT → complete_editor_upload (requires editor:write).
Tool reference
See the full list of available tools, parameters, and scopes under Tools.