The Model Context Protocol (MCP) lets an AI agent connect to a running application, read its state, and take actions inside it — the same way a person would. For most software, this is an interesting experiment. For a video editing workspace, it changes what's possible.
The three-actor model
In a SnipChamp project, three kinds of participants operate on the same footage with the same permissions:
- The human editor — works in the browser UI, drags clips, applies transitions, exports.
- The in-app assistant — answers questions, suggests cuts, generates subtitles.
- An external MCP agent — reads open notes, calls edit functions, marks notes resolved.
All three see the same project state. All three can resolve notes. All three actions are logged in the same audit trail. There is no special privileged path for the AI.
What the agent actually does
When an MCP agent (say, Claude) connects to a SnipChamp project, it can call functions like trim_clip, split_remove, and render_preview. It reads the list of open notes on a clip, understands the timestamps and instructions, and executes.
A typical agent run looks like this: connect to the project, read two open notes (“[00:11:25] cut to the boardwriting, hold 8s” and “[00:37:02] drop this tangent entirely”), call trim_clip(src, 00:11:25, +8s)and split_remove(src, 00:37:02 - 00:39:48), render a preview, post the preview link to the project, and mark both notes resolved. A human reviews the preview and approves it in one click.
Why the notes model matters for agents
The alternative — asking an agent to watch raw footage and decide what to cut — is expensive, slow, and unreliable. The agent is making interpretive judgements about content it may not fully understand. The results require careful human review.
With timestamped notes, the agent is executing human decisions, not making them. The human watched the footage, formed a judgement, pinned a note. The agent reads the instruction and acts. This is a much tighter loop: the human's time goes to watching and deciding; the agent's time goes to executing at machine speed.
Audit trail by default
Because every action — whether by a human, the in-app assistant, or an MCP agent — is logged against the project, you always have a full history of what was done, when, and by whom. If the agent resolved a note, the log shows it. If a human undid that change, the log shows that too. Nothing happens off-record.
This makes AI-assisted editing accountable in the same way human editing is accountable — through a trail that exists in the project, not just in someone's memory.