
MCP: because even your AI needs an assistant sometimes
Top Stories - Overview
A lot of AI tools moved this week, and one of them did it without telling you. FLUX 3 turned up doing twenty seconds with sound in a single pass. Houdini 22 shipped with model training that queues on the farm like any other job. And Midjourney swapped its default model on Friday, so if you built a look last month, go rerun it because it might be better now.
Underneath all that, the plumbing got rebuilt. MCP, the thing that lets an assistant reach into your applications and actually do something, had its biggest revision since launch, and the useful part is that big jobs no longer need someone babysitting them (well, mostly...).
Then there's the other half of the week. A team of twelve people made a fifteen-minute animated film in two weeks and published their budget, which is the encouraging story. The discouraging story is a closed beta at Disney that turns a brand's logo and old spots into finished commercials, pointed straight at the kind of work that's been drifting in-house for years now.
Featured Stories
1. Black Forest Labs Ships Twenty Seconds With Sound
VentureBeat | MarkTechPost
What Happened: Black Forest Labs released FLUX 3, its first public video model, jointly trained across image, video, audio and robot action rather than assembled from separate systems. It generates up to 20 seconds with native sound in a single pass. Early access is gated and video is limited to "partners".
Why Is This Important? Twenty seconds matters more than it sounds. Most models top out at five to fifteen, so anything longer gets stitched, and stitching is where continuity breaks and a compositor loses an afternoon hiding the seams. Native audio in the same pass kills a second round trip. Don't quote the benchmark numbers, though: that chart is labeled a preliminary evaluation of an early candidate, not the model you'd (likely) get.
Tags: #AITools #VideoGeneration #GenerativeAI #AIAudio #BlackForestLabs
2. Midjourney Changed Your Default, and Is Finally Leaving Discord
Midjourney
What Happened: V8.2 replaced V8.1 as Midjourney's default on July 24. The update targets aesthetics, image quality and personalization, with a larger pool of images for building personalization profiles. Every new prompt runs through it automatically, no flag required.
Why Is This Important? If you built a look on V8.1, it's not your default anymore. Rerun your reference prompts before you trust an old sref, because personalization and style handling both moved and there's no system card or benchmark to tell you how far. We flagged V8.2 back in June as unshipped and ran V8.1 instead. This is that story finishing.
The bigger news landed the same day and got buried under an acquisition story. Midjourney is finally building a standalone image app. Four years of living inside Discord has meant every facility with a locked-down network treats it as a consumer toy, and every art director who won't install Discord has been going through somebody else to get frames. An app doesn't answer the licensing questions, but it removes the reason Midjourney kept getting left off the approved list. No name, no date, no window. Just that it's happening.
Tags: #AITools #GenerativeAI #PostWorkflow #Midjourney
3. Houdini 22 Puts Model Training on the Farm
CG Channel | Digital Production
What Happened: SideFX shipped Houdini 22 in mid-July, three weeks after its June keynote. The release adds a native 3D Gaussian Splat toolset, a KineFX and APEX rebuild, and machine learning training that runs through PDG and TOPs rather than as a separate specialist task.
Why Is This Important? The ML detail matters most if you're running a farm. Training a model now partitions and parallelizes the way you'd wedge a sim or a render, so it gets scheduled, queued and budgeted like any other job instead of tying up a workstation for two days. Autodesk argued something close to this last week with Bring Your Own Data. Two vendors in two weeks. Hmm...
Tags: #AITools #VFX #PostWorkflow #Houdini
EDITOR'S NOTE
SideFX
Speaking of Houdini, last week we listed Houdini 22 among the applications shipping MCP support, alongside Blender, Unreal, Silhouette and Adobe. That came from NVIDIA's SIGGRAPH roundup and it was too generous. SideFX's own documentation calls the AI-assisted APEX Script tooling a "preview for a tool that will be released later on". What does ship is an APEX Script package with a VS Code extension and a working MCP server for syntax and snippets, which is useful but isn't the creative-facing access the other integrations offer.
What Happened: Higgsfield premiered Kök Börü at SIGGRAPH, a 15-minute animated short set in 15th-century Central Asia. Twelve people made it in two weeks, pairing hand-drawn animation with generative video and native Kazakh voice actors. The company puts a traditional version at one to two years and one to two million dollars.
Why Is This Important? The number to focus on isn't two weeks, it's twelve people. That's a crew you could pull from a good post house, and the team published its budget and timeline instead of vaguely waving at them, which is rarer than the film itself. A word of caution before trying this particular workflow: it runs on Seedance. To paraphrase Led Zeppelin, it’s whispering tales of IP vulnerability. Maybe don’t upload those Marvel plates just yet.
Tags: #AITools #GenerativeAI #VideoGeneration #FilmProduction #Higgsfield
5. The Protocol Behind AI Integration Just Got Rebuilt
VentureBeat | SiliconANGLE
What Happened: The Agentic AI Foundation published the largest revision to the Model Context Protocol since Anthropic open-sourced it in 2024. The spec drops persistent sessions, hardens authorization, and promotes two capabilities to official extensions: MCP Apps for server-rendered interfaces, and MCP Tasks for jobs that don't finish in one round trip.
Why Is This Important? Tasks is the part that touches our work. A render, a batch transcode or a long conform doesn't come back in one exchange, and the old MCP spec meant something had to hold the connection open and hope nothing dropped. Now the server hands back a durable ticket and the artist can walk away. That's the difference between agent integrations that survive a real job and ones that only work in a demo.
Tags: #AITools #PostWorkflow #GenerativeAI #Orchestration
What Happened: Disney began a closed beta of Ad Creative Studio, a tool inside Campaign Manager that turns a brand's logos, guidelines and prior creative into CTV spots, with multiple variations by audience, geography and campaign goal. Small and mid-sized businesses are the stated target, agencies and larger brands are also in scope.
Why Is This Important? For those doing commercial production or post work, the honest take isn't that AI is coming for our jobs. Brand and agency in-housing has been siphoning off this type of work for years now. Disney just removed the last practical reason a spot like this had to leave the client’s building. What seems likely to still make it out-of-house, however, is work where being wrong is expensive.
Tags: #AITools #VideoGeneration #MotionGraphics #PostProduction #Disney

This week’s Term of Week is MCP which stands for Model Context Protocol. It's the standard that lets an AI assistant reach into an application to see exactly what you’re working on like read a timeline, build a node tree, adjust a shape, or kick off a render.
Before it existed, every pairing of an AI assistant and an application needed its own custom connection, which meant almost none of them got built. MCP turned that into one connection anyone can use. That's why the acronym started showing up in release notes everywhere, all at once this year.
But, it's also worth knowing what it isn't. MCP doesn't make you or your assistant smarter and it doesn't decide what anyone is allowed to touch. It's a doorway. Who gets a key is still decided by whoever set it up.
General AI News
Adobe is dropping Intel Macs. Version 27.0 of Premiere, After Effects, Media Encoder and Audition will require Apple silicon. Intel machines stay on 26.x with security updates only, and Adobe points to Apple dropping Intel in macOS 27 as part of the reasoning. If there's an Intel Mac in your bay, that's now a hardware decision with a date attached. Adobe
Anthropic shipped Claude Opus 5, its fourth model in under two months, at half the price of its top tier with a per-request effort dial. The cadence is the story: standardize on one model and it's superseded in weeks. TechCrunch | Axios
An open-weights letter doubled in a weekend. Twenty-five companies published "Open Weights and American AI Leadership" on July 24, urging Washington against restrictions on downloadable models. Within a day it carried fifty names, including OpenAI and Google (both reported absent at launch). Anthropic and Amazon are still out. Read the list by business model rather than by principle: almost everyone who signed sells something that free models drive demand for. CNBC | Forbes
Four days later, the same industry asked for a brake. More than a thousand employees at OpenAI, Anthropic, Google and Meta signed a letter asking the US government to build the tools to deliberately pace frontier AI development. It isn't a contradiction of the first letter. One is about distributing what already exists, the other about a mechanism for what doesn't exist yet. CNN | The Register
Topaz Video is on a clock. The FTC granted early termination on Adobe's acquisition, clearing it well ahead of schedule. The part that matters if Topaz Video is on your drive: the app was already frozen at v7.1.5 with no further model updates and no support for future operating systems. The clock was running before the deal. It's running faster now. Seeking Alpha | ProVideo Coalition
Utopai Studios is developing a theatrical animated feature from Mike Bender's children's book The Most Serious Fart, with Bender writing and directing. The interesting part isn't the title. Bender and his producer spent years trying to make it conventionally and the economics never worked. Variety | AWN
IN CASE YOU MISSED IT
EU AI Act Article 50 takes effect Monday, August 2. If your work reaches EU audiences, synthetic content needs marking and deepfakes need disclosure, whether you built the tool or just used it on a job. We covered the final guidelines on July 17.