Dan Osterman— AI Product Engineer & Creative Technologist

Case 03 · An AI video platform for filmmakers

Bring your own agent

The problemI already work through my own agent. I want it to drive your platform, not learn another chat box.

03.1Claude calls Update shot prompts. Shot 37 changes from 65 mm to 50 mm beside it
Play full video
Tool permission groups and examples from the expanded write/delete tool list
03.2Write tools and their approval controls — an excerpt of the tool list
View full size
The organization-level agent-access setting with Allow agents to connect enabled
03.3Agent access enabled for the organization
View full size

Notes · Case 03

In short

An MCP server for the platform so any agent client, Claude, Codex or a custom one, can create stories, add and edit shots, upload references and generate, under the user's own account and limits.

Role
Designed and built the MCP server and its authorization
Client
An AI video platform for filmmakers
Year
2026
Stack
Model Context Protocol · OAuth · platform API

What came out

  • A user connects once and their agent can read and change their stories through named tools
  • Uploads, including images and video references, go through the same path as the web app
  • Everything the agent does is scoped to the user's account, organization and spend limits

Where it started

Some of the most productive users never wanted a chat box inside the product. They had their own agents, their own scripts, their own way of working, and they wanted the platform to be a thing their agent could operate. Others wanted to run a batch overnight from a terminal. The web app could not do either.

What I built

An MCP server that exposes the platform as a set of tools: list and read stories, create and update shots, upload images and video as references, start and check generations, and the rest of what a user can do in the browser. A user connects their agent once through the platform’s own sign-in and the agent works under their account.

The point was not to add an agent to the product. It was to let any agent become a user.

What was hard

The tool surface has to be small enough that an agent picks the right one and complete enough to finish real work. Uploads are the hard case: an agent sending a reference video needs the same validation, storage and lineage as a person dragging it into the browser. And nothing the agent does can exceed what the person who connected it is allowed to do or spend.