How The Bridge Got Remade in a Day and a Half: Astra, Codex, and Runway's MCP
The making-of for The Bridge (2026): GPT-6 Astra running Codex, Runway's MCP driving Seedance 2.5, a plain-HTML production dashboard, directing by voice, and a DaVinci Resolve grade, title cards, and audio mix I never touched a mouse for. Plus the two places the agent fell down, and what the whole thing cost in usage.
A few days ago I put out The Bridge (2026), a remake of one of my first AI films. I’m not usually a fan of going back and tinkering with old work, but given what the tools have done lately, if I was ever going to revisit something, now was the time. So I went full Lucas on it — and the whole thing took about a day. A day and a half if you count cleanup. This is the breakdown of how, including a couple of moments that had my jaw on the floor. Quick disclosure up front: Runway sponsored the remake. More on them further down.
The original vs. the remake
The 2025 original was built on an early-access version of Veo 2, which at the time was about as good as it got. No audio, so any dialogue or lip sync meant a trip through a third-party tool. No multi-referencing either — this was right around the start of the hack where you take the first frame and put “cut immediately” in the prompt. That version took about a week of pounding nails into the wall with my forehead.
| The Bridge (2025) | The Bridge (2026) | |
|---|---|---|
| Video model | Veo 2, early access | Seedance 2.5 on Runway |
| Audio | None native; lip sync via a third-party tool | Native |
| References | No multi-referencing | Character sheets plus 360 turnarounds |
| Who ran the tools | Me, by hand | GPT-6 Astra in Codex, over Runway’s MCP |
| Time | About a week | About a day and a half |
Most of that gap is down to GPT-6 Astra and Runway’s MCP. Astra is the same model that replaced my daily driver a couple of weeks ago; this is what it looks like on an actual production.
The setup: Codex, Astra, and Runway’s MCP
If you’re not familiar, MCP stands for Model Context Protocol, and all you really need to know is that it’s the thing that lets your LLM drive the Runway platform — or whatever else you point it at. Combine that with Astra and things get strange fast.
To follow along at home: download the ChatGPT desktop app, set it to Codex, and have a Runway account. Then hop into Codex and say “hook up the Runway MCP.” That’s it. Genuinely the easiest part of the whole project.
The studio film skill
The other big piece is my Theoretically studio film skill, which I’m still constantly revising. It’s a folder structure: a CLAUDE.md and a SKILL.md at the top, then subfolders for references, brief, assets, and tracking, organized by production phase — story and style, cast and locations, shots and stills, and the videos themselves. The point is that I can point Claude at it, or Codex, or Kimi K3, and all of them know exactly how my films are structured and how to execute them. It’s the descendant of the production office from the Paperclip Heart breakdown, and the folder plus a full tutorial goes out in this week’s newsletter.
Script, characters, and 360 turnarounds
First job for Astra: literally watch the original, then generate a script and a shot list from it. Meanwhile I went to Midjourney for the initial character references and location plates. Working up the ghoul, I flashed back to why the original used a Death figure at the gate: I didn’t have to lip sync him. Not a constraint this time.
From the character sheet, the next step is a 360 rotation generated with whatever video model I’m using — here, Seedance 2.5 on Runway. Why video rather than an image editor? Pure speculation on my part, but I have a sneaking feeling that if the video model generates the back of the character, then somewhere in its brain that’s what the back looks like, and you’ve got a better shot at holding it through the rest of the process.
The production dashboard, and Astra’s diary
The studio folder gets dense enough that even with a big-brain model keeping it organized, I couldn’t tell what I’d actually made. The fix this time was a production dashboard: a plain HTML file sitting on my desktop, pointed at the project folder. Tabs for every shot with its prompt and notes, the storyboards, the references (including a tribute tab to the source film that I never actually opened), and a breakdown of every location, scene, and sequence. For me, this was the single biggest improvement of the whole round.
One of those tabs is a diary. At the end of each session, Astra wrote up what we’d done, so that when I forgot something, it could remind me.
Directing by voice
From there it’s the standard loop: generate images, review them, send them to Runway. The strange part is that I did most of it through Codex’s voice mode — chatty, conversational directions in, properly structured prompts out, because the skill folder already tells the system how I like to prompt. It played like Astra as my production manager handing instructions to Runway as the creative house, with the footage circling back to me to approve or nix. It felt very much like stepping into the future.
What the agent missed
Keep one eye on the platform yourself. As good as the agentic stuff is, it’s not great at thinking ahead, or at knowing what options even exist. Case in point: Seedance 2.5 on Runway can output ProRes 422 HQ, which apparently sidesteps a lot of the banding we’ve been seeing in 2.5, because you skip the MP4 compression. That would have been very handy to know in the early stages. Astra never reached for it.
The wait, and cutting it myself
Less friction, but it feels slower — which is a weird thing to say about a film that took a day and a half. Most of that is the round trip: Astra assembles the prompt, sends it to Runway, brings the result back, files it in the right folder. There’s dead time.
I spent some of it testing Astra as an editor, placing clips on my Premiere timeline. Adobe doesn’t have an MCP yet, so this ran on computer use: it took over a monitor, moved slowly, and got very upset with me if I touched anything. Runway did announce Premiere Pro and After Effects plugins while I was mid-project, so you can now generate straight onto the timeline — that’s one to circle back to. For now, I’m just faster. I made the first cut myself and left a couple of spots open.
A word on Runway
They sponsored the remake, so here’s the pitch, no hard sell. Runway hosts just about every model on the market — Seedance, Kling, MiniMax, Nano Banana, GPT Image, and their own Gen-4.5 — and I’ve always thought of them as the bridge between indie projects and full professional ones. If you’re starting out, the apps are broken out by films, shorts, marketing, social, and educational content. If you’re at the other end, there’s ProRes output, HDR via Ruby, and the Premiere and After Effects integrations. The research side is the part I keep an eye on: they just showed GWM Worlds 2, real-time interactive generated worlds with audio, alongside a paper on real-time video that lays out where they think it’s going. Plans start at $12 a month, and first-time subscribers get 25% off any plan with code THEO25 at Runway.
The Resolve pass: grade, Ruby, titles, and mix
With a first cut done, I wanted to see what Astra could do with color, which is not my strength. I can fumble through a Resolve edit, but the color page and its nodes lose me completely. So I exported the XML from Premiere, imported it into DaVinci Resolve, and said, more or less, “go ahead and color correct this.” It got to work.
To be clear, this is a grade pass on Rec. 709 footage, which is what Seedance outputs. It’s not log or raw, so we’re adjusting the colors that are there rather than recovering anything. We added Dehancer for film grain.
If you want more headroom, that’s what Ruby is for: Runway’s model that takes SDR clips and redistributes brightness and color into an HDR container. Before anyone types it — no, that is not the same as log. You’re not recovering detail. But it does give you room to actually push a grade, and I had it go ham on a few shots just to see how far it would go.
While we were in there, a few small edits: pull a shot back a few frames, add a cross dissolve, that kind of thing, all by giving it a timecode. Then the opening and closing title treatments, which it built in Fusion. I still don’t know what any of those nodes connect to. I can move them around, though. Finally it did the audio mix — dialogue, music, and ambience leveled and smoothed across three tracks. Not a lot to do, but I didn’t touch the mouse once in Resolve.
What it cost in usage
About 70% of my weekly usage, on the top plan. That includes some genuinely dumb spend: I left it on Ultra by accident several times when medium or low would have done, and I had it download and install Resolve for me over computer use — a total waste of tokens, and so much fun to watch.
Where this goes
Set the two films side by side and it’s a little mind-boggling how far things moved in so short a span. The next phase, I think, is world models and real time, which is why that Runway research page is worth a visit.
I know for some of you this is anxiety-inducing. I just color graded and mixed audio without touching my keyboard. I get it. But this is a tool that unlocks other tools in service of your creativity. If you’re apprehensive, engage with it: learn what it can do, what it can’t, and what it can do for you. There are no rules, and there never were.
Clock starts now on The Bridge 3.0, two years out.