Bridge AI agents into your video review loop.
Generative AI can render video, but it cannot read client minds. Dusken connects your agents directly to frame-accurate critiques, visual sketches and spoken director notes. Close the creative feedback loop.
Dusken MCP is an open Model Context Protocol server that gives AI agents (Claude, Cursor, OpenAI) programmatic read and write access to video timelines, frame-accurate annotations, speech transcripts, version stacks and NLE markers over standard HTTP JSON-RPC 2.0.
Generative video is fast. Creative review is still broken.
Generative pipelines can spin up dozens of cut variations in minutes. But the moment a client or director needs to request an edit, the workflow falls apart.
Today, using AI for video review means manual translation. An editor copies approximate timestamps out of an email, captures screenshots of comment pins, pastes them into an LLM chat prompt and hopes the model understands which corner of the frame to fix.
Dusken removes the manual translation layer. By exposing video timelines, frame coordinates and audio transcripts through the Model Context Protocol, your AI agents have a native pair of eyes and ears on the cut.
Two workflows. One connected feedback loop.
Whether your pipeline starts with an autonomous agent or a human editor in an edit bay, Dusken closes both directions.
The Agent-First Loop
AI Generates → Human Reviews → AI Iterates
1. Your agent renders Cut V1 and pushes it to Dusken with an automated upload call.
2. Directors and clients review on any browser with zero login: drawing vector arrows, marking ranges and recording voice notes.
3. Your agent queries Dusken MCP, extracts timecodes, speech transcripts and spatial quadrants, updates the cut to V2 and stacks it automatically.
4. Humans compare V1 and V2 with synced Wipe or Pixel Diff sliders and approve the final version.
The Human-First Loop
Editor Cuts → AI Reads Notes → NLE Markers
1. An editor uploads Cut V1 to Dusken for team and client review.
2. Collaborators pin frame feedback and voice comments across the timeline.
3. An AI agent reads the frame notes via MCP, handles automated fixes and checks off resolved feedback items.
4. The agent generates native DaVinci Resolve EDL or Premiere Pro XML timeline markers so the editor sees actionable notes directly on their edit track.
How studios and editors run Dusken MCP
Concrete pipelines operating on the Model Context Protocol today.
Generative Ad Pipelines
ComfyUI / Runway → Dusken → Claude
An AI studio renders 20 variations of a product commercial. The creative director reviews on mobile, circles packaging defects and leaves a voice note. Claude queries Dusken MCP, extracts the timecodes and spatial quadrants, updates the prompt generation script and triggers the next batch render.
Editorial Grunt Work Sync
Human Cut → AI Marker Sync
A documentary editor uploads a 30-minute cut. Collaborators leave 45 timecoded notes. An AI agent reads the feedback via MCP, isolates audio cleaning tasks, checks off completed items and generates a native DaVinci Resolve EDL or Premiere Pro XML marker file directly on the edit track.
Automated Client Sign-Off Audit
Cut V1 vs Cut V2 Audit
Before sending an updated cut to an enterprise brand, an agent calls compare_versions to verify that every requested client critique from Cut V1 was addressed, producing a signed verification log and eliminating missed revisions.
What your agent actually sees and understands
Dusken does not just pass raw text strings. It translates human video interaction into structured spatial and temporal data.
When a director circles an object on screen, Dusken calculates spatial quadrants (top-left, lower-third) and delta arrow vectors. Your agent receives exact coordinates alongside high-res composite frames.
Reviewers hold Ctrl+Alt to speak and sketch simultaneously. Dusken automatically transcribes the audio, detects spoken timecodes ("cut from 0:12 to 0:24") and serves searchable timestamps directly to your agent.
Agents can inspect comment resolution across Cut V1 and Cut V2, verifying that every client critique was addressed before notifying human collaborators.
With long-polling webhooks (wait_for_feedback), your agent sleeps until human notes land, wakes up instantly and kicks off automated re-rendering without burning API credits.
Protocol Architecture and Specifications
Built strictly on Anthropic's Model Context Protocol specification for deterministic agent execution.
| Specification | Technical Implementation |
|---|---|
| Transport Protocol | Streamable HTTP (POST) over TLS 1.3 with chunked transfer encoding |
| Message Standard | JSON-RPC 2.0 complying with the open Model Context Protocol specification |
| Authentication | Bearer Token authorization with granular per-workspace read/write scopes |
| Canvas Coordinate System | Normalized 0.0 to 1.0 vector coordinates with spatial quadrant mapping (top-left, lower-third) |
| Timecode Precision | SMPTE (HH:MM:SS:FF) frame matching and millisecond floating-point timestamps |
| Timeline Marker Formats | DaVinci Resolve CMX 3600 EDL, Apple Final Cut Pro XML, Adobe Premiere Pro XML/CSV |
| Media Encoding and Transcripts | Adaptive bitrate HLS proxy streaming with automated Whisper speech-to-text transcription |
Built for the tools your team already uses
Dusken connects with desktop apps, coding IDEs and autonomous agent runtimes.
Add Dusken to your Claude Desktop connectors or register it in your terminal with Claude Code for zero-friction review analysis.
Connect ChatGPT, OpenAI Agents SDK and LangChain pipelines via our standard Streamable HTTP endpoint.
Connect your coding IDE to review data. Build video render scripts and check client revision notes right in your editor.
Ready to connect your client?
Run this command with Claude Code or copy the endpoint for your environment:
What changes once you connect Dusken MCP
Copying approximate timecodes from Slack and email, taking screenshots into ChatGPT, guessing what the director meant and losing revision history between renders.
AI agents connect directly to frame-accurate review data, vector coordinates, audio transcripts and automated DaVinci / Premiere timeline markers.
| Workflow Capability | Disconnected AI Generation | With Dusken MCP |
|---|---|---|
| Review and Approval Flow | Manual screenshot copy-pasting and timecode typing into chat prompts | Native MCP JSON-RPC protocol queried directly by agents |
| Spatial Understanding | Vague written descriptions of visual adjustments | Quadrant-localized arrow vectors and composite image frames |
| Timeline and NLE Sync | Manual re-editing or manual retyping of sequence markers | Automated DaVinci Resolve EDL and Premiere Pro XML exports |
Built for production video pipelines
Close the creative loop
Transform disconnected AI video generation into a continuous, human-directed feedback and iteration loop.
Included on all plans
Native MCP server access is included across Free, Solo and Team tiers without enterprise surcharges.
Unlimited free reviewers
Add directors, clients and team members with zero seat fees. Clients review free without creating accounts.
Frequently Asked Questions
No. As of 2026, Frame.io does not offer a native Model Context Protocol server for AI assistants. Developers must write custom webhook scripts against private APIs. Dusken provides a native, built-in MCP server on all plans, allowing Claude, Cursor and OpenAI agents to query frame comments, inspect visual canvas sketches and export timeline markers directly.
Generative video workflows use Dusken MCP to close the feedback loop. An automated pipeline uploads generated video cuts to Dusken via upload_video. Directors review the cut in any browser without logging in, drawing arrows on visual defects and recording spoken voice critiques. Connected AI agents query list_feedback to retrieve exact timecodes, speech transcripts and spatial quadrants, update prompt scripts and trigger the next batch render.
Yes. Connected agents can use post_feedback to reply to client threads with timecoded suggestions, ask clarifying questions on specific frames, or update reviewer action item statuses to resolved once an edit is applied.
No. Clients and external directors review through zero-login browser links with standard drawing and voice tools. They never interact with terminal commands, API keys or model settings. The MCP bridge operates transparently in the background.
Yes. Agents invoke export_nle_markers to convert all human review notes, speech transcripts and drawing references into native DaVinci Resolve (EDL) or Adobe Premiere Pro (XML) timeline marker files. Editors import these markers directly onto their sequence track with zero retyping.
Yes. Model Context Protocol server access is included across Free, Solo and Team tiers without enterprise surcharges. External reviewers are 100% free and unlimited on every plan.
Got something else? Contact us
We actually read every message. Bugs, ideas, rants, all welcome.
Ready to connect your AI agents to video review?
Give Claude, Cursor and your autonomous pipeline direct access to frame-accurate review data today. Free forever on Dusken.