How BuzzRolls Studio works

AI Agents interact with software that is built for humans via tool calls. For example, if you have the Gmail connector for Claude, it can draft an email using the create_draft tool call followed by the recipient, subject, and email message. The Claude App sees that tool call, and uses the Gmail connector to create that draft inside your Gmail account. The Claude LLM model (Opus, Fable, etc.) is the brain, the Claude Desktop app is the hands, the tool call describes the action you want to take, and the connector gives it the access to carry out that action.

BuzzRolls Studio is a connector that gives your AI Agent access to around 60 tool calls that turn it into a Premiere Pro video editor.

There are three categories of tool calls from BuzzRolls Studio:

1. Tool calls that directly control Premiere Pro through the BuzzRolls Studio companion Premiere Pro plugin.

In order to edit videos for you, your AI Agent needs somewhere to do it, and ideally, the place it edits videos is also a place where the human can edit the same video. AIs and humans are better at different things, and by working together they can produce a better result than either alone. But in order to do that, we have to have the AI work in a place that is also easy for humans to work in. That’s why we chose Adobe Premiere Pro. Typically, our customers will let their AI do 90% of the work, producing the first draft of the video. Then the human comes and does the remaining 10% of the work, making final edits and polish to the video. Since BuzzRolls Studio works in the program you already know how to use, it is easy for the human to do the parts of editing that humans are best at.

But to make this work, your AI Agent needs some way to actually use Premiere Pro. That’s where BuzzRolls comes in. BuzzRolls Studio makes editing in Premiere Pro intuitive for your AI by turning it into a format that is easy for it to process and understand. That format is tool calls. For example, when your AI wants to understand the timeline, it uses the get_timeline_state tool call. The BuzzRolls Premiere Pro plugin sees this tool call, reads the timeline, then turns the visual timeline into text with what clips are where, on what track, and what their start and end times are. This information then gets handed back to your AI so it can understand the timeline.

After that, let’s say that your AI wants to split a clip. It calls the split_clip tool call, instructing which clip to split and where. This sounds simple but actually requires many complex pieces behind the curtain to make this work well. Feel free to read the technical explanation section for more info. Thankfully, as the user, you don’t need to worry about any of this complexity or what each tool call is or what it does. Your AI will understand exactly all the features of BuzzRolls Studio and how it works. You can just talk to it like how you would talk to your assistant editor.

But wait, can’t AIs just control your computer mouse and keyboard directly?

Yes, kinda. Computer use is evolving and getting better. Your AI can click around on your screen and take actions if you allow it to. Guess how it works. Yep. Tool calls. Your AI uses the screenshot tool call to see what’s on screen, then the click tool call with screen coordinates to click somewhere. The problem with this is that it is extremely slow, extremely token expensive (meaning it uses your usage really fast), and highly error prone. Dedicated tool calls to control a program will always be cheaper, faster, and more consistent than computer use, even if computer use gets really good. Premiere Pro is basically the hardest case for computer use to do too because taking any one action requires so many clicks, drags, re-zooms to do. In the same way that it is hard for a human to look at 100 lines of code and understand what is going on, it is hard for an AI to look at a picture of Premiere Pro’s timeline and click around to understand what is going on.

It’s also worth pointing out that not all AI Premiere Pro connectors are made the same. Open source Premiere Pro MCPs exist, but they don’t replicate what BuzzRolls Studio does. These open source projects are a good starting place if you’d like to rebuild the functionality we made for BuzzRolls, but if you just want something that works and you want to focus on churning out high-quality videos instead of focusing on fixing obscure Premiere Pro tool call bugs, then BuzzRolls Studio is the right answer.

2. Tool calls that teach your AI how to be a video editor.

Just because your AI can control Premiere Pro, doesn’t mean it will be good at editing videos. It’s exactly the same thing for humans. Video editing is hard. Premiere Pro is hard. I don’t have to tell you this. You once had to learn how to use Premiere Pro, remember how long it took to learn? It’s the same thing for AIs.

BuzzRolls Studio solves this through something called “progressive disclosure”. The basic concept is that your AI is given just the right knowledge of how to do something at just the right time when it needs to do that thing.

I’ll give a real example. Let’s say you ask your AI to edit down your podcast and switch camera angles depending on who is talking. The first thing your AI will do is load up BuzzRolls Studio so it can start editing. When it does that, it sees that BuzzRolls provides instructions to read when editing podcasts. It reads those instructions and knows which tool calls it should use to get the job done. Great. The first step is for it to understand the timeline and use the get_timeline_state tool to do that. When it reaches for the get_timeline_state tool call, that tool call has its own instructions that tell it how to use that tool, and what gotchas it can encounter when using it. Your AI agent then uses the tool call to see the timeline, and OOPS, it did something wrong when it used the tool and was given an error. No worries though, along with the error was another instruction of how to do it right to fix the error. Great. Your AI agent sees what it did wrong and tries the tool call again. This time it succeeds and can see the timeline.

This whole process sounds complicated (and it is), but it happens at the speed of AI. Most people that use BuzzRolls Studio have no idea how much is going on to make this work. That whole loop happens for hundreds of tool calls that your agent uses to accomplish the task you gave it. From the human perspective, they asked their AI to do something, they got up, came back 10 min later, and the job is done.

3. Tool calls that enhance your AI to ‘see’ and ‘hear’ the video.

Our mission at BuzzRolls is to turn your AI into the best Premiere Pro video editor we can. Even if your AI can use Premiere Pro, and your AI understands how to edit video, it’s still limited by the fact that it primarily sees and processes the world as text. It can’t hear or see your video in the same way humans do. Imagine how hard that would be to edit a video without audio and with a blurry pair of glasses. So, our solution is that we give your AI access to BuzzRolls cloud tools that let it ‘hear’ and ‘see’ the video more effectively.

For example, when we were first building BuzzRolls, we noticed that the AI would often make obvious errors and cut the video at the wrong times. Upon investigation, we realized that it actually wasn’t making any mistakes according to what Premiere Pro’s transcript was telling it. As you’ve probably noticed if you use the text-based editing feature in Premiere, Premiere’s transcription is… bad. The timings are inaccurate and the words themselves are very often transcribed to the wrong words. It works fine for a human though because we can still listen to the video in addition to what we can read on the transcript. But for an AI agent, who can ONLY go off of the transcript, this is a major problem.

So, we developed our own agent-grade transcription system that is SIGNIFICANTLY more accurate than Premiere’s. Yes, we ended up building our own transcription platform just so your AI can do its editing more precisely. That’s not to say that we are smarter than Adobe, but our transcription systems are built for different use cases. Premiere’s is built to be fast and cheap. Ours is optimized for being accurate and in a format that is ideal for an AI agent to look at. After we rolled out the new transcription system, our edit accuracies improved significantly.

Another example is ‘seeing’ the video. Your AI can export screenshots and look at them when it needs to verify something visually. You can imagine why this is not ideal, for many reasons. So, we created the visualize_clips tool call. Your AI agent can send BuzzRolls Cloud the video, ask it a question such as “At what time does the snowboarder do the 180 backflip?” and it gets back an answer that it uses to place the video. Often it will then export a screenshot of exactly that time to verify that’s when the backflip happens.

There are a couple of things worth noting about these cloud tools. They cost BuzzRolls Credits. These services are good because they use expensive transcription and visual analysis AI models. Not hugely expensive, on the order of ~$1 for most videos. Expensive enough that we had to create a credits system so that we don’t lose money on our power users that churn out videos using BuzzRolls Studio. If you are on the BuzzRolls Studio Pro plan you’ll pretty much never need to think about credits. Your subscription covers it. If you are on the BuzzRolls Studio Lite plan, you may need to refill your credits here or there. We make the majority of our profits on our subscription fees, not the credits system. We try to price everything near cost for what it costs us.

Also, in order to use BuzzRolls cloud tools, your AI agent has to send your video or audio up into our cloud to do the analysis, and I know some people may be concerned that we are stealing your IP or using it to train AI models. We’re not. We want your money, not your data. We delete your content off our servers immediately after the processing is complete. We have no reason to hold onto your info. That’s a liability headache for us and serves us no purpose. We just want BuzzRolls Studio to work well enough that we keep you as a customer month after month.

You probably still have questions. If you do, you can put them in this handy questions box. Thanks for reading.

BuzzRolls Studio is an agentic harness that optimizes a shell-capable AI Agent for video editing inside of Premiere Pro. All of BuzzRolls Studio’s functionality is mediated through tool calls hosted on BuzzRolls Studio’s MCP server.

There are three tool call types that BuzzRolls Studio serves:

1. Verbs

These tool calls control Premiere Pro through the companion BuzzRolls Plugin.

These verb tool calls abstract the raw UXP JS that controls Premiere Pro into a format that is optimized for AI agent usage. Premiere Pro is FULL of landmines and edge cases. We’ve developed a self-improving software system that catches edge cases as they happen, and patches new classes of bugs as they appear (shoutout to our early access users that helped catch many of the early bugs!). As an example, the UXP framework does NOT have a razor split function, but BuzzRolls Studio still provides one. We chain together thousands of lines of code and a slew of UXP methods to recreate a razor split and handle edge cases that would otherwise destroy your timeline. Another example: it turns out that the set_effect_param tool worked correctly on all versions of Premiere EXCEPT the German-language version. We would have never caught this obscure bug unless we had it error for our German users.

Verb tool calls also provide progressive disclosure. They give your AI the information it needs to complete the task at just the right time. For example, when your AI wants to get the transcript, it fetches the get_transcript tool call. The tool call then reveals the full description of how to use it. If the agent uses the tool call wrong or some other event causes the tool call to fail, it will be given a message that steers the model towards using the tool correctly or resolving the issue that caused it to error in the first place.

Verb tool calls are optimized for reducing token consumption. This is done through a variety of means that all come down to either decreasing the input tokens that are returned by the tool calls or reducing the output tokens that the model must expend to perform a given task.

Examples of input token reduction: scoped information retrieval, such as in the get_transcript tool call (your agent requests ONLY the exact transcript information it needs for a given part of the timeline and with the exact amount of timing detail that it needs, and it is compelled to do this by the progressive disclosure system).

Example of output token reduction: abstracting as many actions into as few tool calls as possible. The remove_bad_takes tool call creates a transcript document that the agent text-edits down to trim the video down. Without this tool call, the agent would need to chain together direct timeline edit tool calls, which is more output token heavy and prone to error.

2. Playbooks

These tool calls provide guidance on how to accomplish common video editing workflows.

You can think of playbooks as being like documentation. They provide jumping off points for the agent that steer them towards approaching common video editing tasks in ways that are proven effective. For example, the edit_podcast_playbook provides step by step guidance for how to edit different flavors of podcasts based on footage context. Not all editing work mediated by BuzzRolls Studio happens through Premiere. Some editing tasks happen best using bash commands or writing python scripts that manipulate the footage. Playbooks sometimes steer the agent towards sets of ffmpeg commands that are proven more efficient than doing the same task in Premiere.

3. Recipes

These tool calls are the entry point for your agent interacting through BuzzRolls Cloud hosted APIs.

There are two recipes in particular that significantly improve an AI Agent’s video editing abilities: the transcribe_audio and visualize_clips tool calls. Premiere Pro has an internal transcription engine that is… bad. Early on, we noticed that almost all the video editing ‘mistakes’ the agent made were from it receiving improper word timing data or incorrect transcriptions from the Premiere Pro transcription. For a human, this isn’t a problem. We can listen to the audio in addition to reading it via the transcript. AI Agents can ONLY see the text from the transcript. Therefore, we had to build our own Agent-Grade transcription service. It is far more compute expensive to run than what Premiere offers, but because it is so critical to the success of an AI agent, it is worth running that agent-grade transcription pretty much every time.

visualize_clips is a similar story. LLMs are optimized to see the world as text. Using visualize_clips, they can ask a model optimized for video understanding a question of the video, for example, “When does the basketball player make the shot?” and it gets returned the timestamp it can use to align the drop in the music.

Recipe tool calls generally cost us money to run 3rd party APIs and our own compute. This is why we charge BuzzRolls Credits for these services. Most people use a small amount of credits, a fraction of the cost they are spending on their BuzzRolls subscription. However, some power users use VASTLY more than everyone else, so to keep our service fair, we’ve separated the costs for these services out from our standard subscription.