Praxis: Human + Agent NLE

Inspiration

A long time ago, in another life, I was the editor for the Pan Sonic DVD Kuvaputki and the director would bark what he wanted achieved. And I figured out how to do it. With WebMCP, now it's my turn. I wasn't a tireless agent. This is more humane.

What it does

The human directs. The agent assists. Both work on the same project. Praxis is a human-and-agent video production studio with a script, scenes, assets, preview, timeline, and action ledger. Through WebMCP, an agent can read the current project and selection, create checkpoints, make bounded edits, inspect change history, and run structural quality checks. The collaboration is visible: an agent changes a script beat, the revision advances, and the operation appears in the ledger. The director can restore a checkpoint without leaving the editor. The submitted demo uses existing media. Live media generation, rendering, and background delegation are disabled.

How we built it

We built the studio with React, TypeScript, and Vite. A Cloudflare Worker connects it to a SQLite-backed Durable Object that maintains the project’s authoritative state. The human interface and WebMCP tools use the same revision-aware command system. Changes are validated, checked against locks and revision conflicts, and recorded in history. The hosted demo exposes six WebMCP tools: project context, current selection, change history, project operations, checkpoints, and structural quality checks. Codex was our development collaborator throughout.

Challenges we ran into

The hard part was not letting an agent change something. It was making that change understandable, bounded, and reversible. We worked through browser-specific authentication behavior, WebMCP availability, asynchronous project loading, and keeping browser state synchronized with durable state. We also had to distinguish successful tool registration from proof that an agent could actually discover and invoke those tools. Deadline pressure forced a useful decision: present a narrow collaboration loop we could demonstrate honestly.

Accomplishments that we're proud of

An agent and a human can work through the same production model, with visible attribution and a human-controlled way back. The demonstration shows an actual semantic tool operation—not simulated clicks—followed by a director restoring the baseline. We also reached 250 passing tests across the application and control plane, plus passing TypeScript checks.

What we learned

Trust is something you build into the workflow. Shared context, clear boundaries, visible history, and reversibility matter more than a reassuring chatbot response. We also learned that a smaller working demonstration can communicate an ambitious idea better than a larger collection of unfinished promises.

What's next for Praxis: Human + Agent NLE

Connect the existing architecture to provider-backed media generation and rendering, deepen timeline editing, and improve project isolation and collaboration. The guiding principle stays the same: give the agent useful production responsibilities while keeping creative authority with the human.

Built With

• cloudflare-durable-objects
• cloudflare-workers
• git
• github
• html/css
• openai-codex
• react
• sqlite
• typescript
• vite
• vitest
• webmcp
• zod