MEL

MEL is a filmmaking app for generative AI. You bring an idea (a logline, a script, a vibe) and MEL develops it into a shootable spine, breaks it into sequences and shots, and writes each cinematic prompt with craft and continuity. You generate the video, images and audio right there in the same window, react to what comes back, and MEL gets measurably better at the next prompt because it learned from the last one.

It runs on your own Claude subscription rather than a metered API bill — the agent work is headless Claude Code under the hood — so directing and iterating all night never runs a meter.

The learning loop

Most AI tools forget everything the moment you close the tab. MEL compounds. Every render is watched, judged and remembered, and that memory flows back into the next prompt automatically:

• It watches your renders — when a generation finishes, an agent looks at the actual frames, compares them to the prompt, and drafts a verdict: what worked, what failed, and why.
• It captures your intent — the change you asked for is logged as the hypothesis, so your reasoning is never lost to a blank field later.
• It distills lessons — patterns across many renders get promoted into a growing rulebook of hard-won craft rules, with one click of your approval.
• It briefs the agent — before writing a prompt, the agent is handed exactly the lessons that matter for this change, plus the shot's own history of what already failed.
• It reviews before you spend — a pre-flight check reads your prompt against the whole rulebook and predicts what will go wrong, before the render costs you money and minutes.

Render → verdict → lesson → back into the agent's head. You stay editor-in-chief at every gate; the tool does the remembering. The longer you use it, the fewer wasted renders you burn — your studio's hard-won knowledge finally accumulates instead of evaporating.

In the app

• MEL, your director — a conversational film director that plans coverage, blocking, camera and edit, then writes every shot's prompt and a continuity doc to hold the sequence together.
• Generate — in-app video, image and audio generation (Seedance, Flux, GPT Image and more), driven straight from the prompts the agent wrote.
• Library — every render you've ever made, durable and browsable, organised by project. Nothing thrown away; favourites one click away.
• Shots & Stills — iterate any prompt with an agent that respects the discipline: one variable at a time, diffed against your last good version, linted before you spend.

Thin context, fat repo. Your prompts, lessons and renders all live in a plain, portable repository you own — the source of truth; the cockpit is the window onto it. Single-user and local-first: no cloud, no account, nothing leaves your machine except the render calls you choose to make.