Orchestrate your projectsand files
In the Pibox chat — or from Claude, Cursor and VS Code over MCP.
- Neon Skyline_v4.wavWAV 48k · 3 comments
- Neon Skyline_v3.wavWAV 48k · superseded
- Stems12 files
An example thread. Every move in it — search, review history, tasks, properties — is something the agent does through the same API the app uses.
The hours nobody bills for
Hunting for the right version. Rewriting comments into tasks. Checking a delivery twice. The work around the work is where projects leak time — and it is exactly the part an agent inside the project can take.
Find the take, not the folder
Ask for “the last bounce with the new vocal” and get the file. It knows versions, dates, who uploaded what, and which one the client actually saw.
Catch up in one answer
What moved since Friday, what is waiting on whom, what the client asked for at 1:12 — read back from the real thread, not from a summary somebody retyped.
Feedback becomes a checklist
A review round turns into tasks on the right folder, with the timecode attached, instead of a producer transcribing comments at midnight.
Deliveries that pass the first time
Loudness, true peak, clipping, phase and format, checked on every audio file against streaming, EBU R128 and ATSC targets — before the package leaves the building, not after the label bounces it.
Search sound and picture by what is in them
AI tags for genre, mood and instruments, tempo and key measured from the signal, artist and ISRC read from the file itself — and in video, what is said and what is on screen. “Warm analog drums, around 90 bpm” becomes a query, not an hour of auditioning.
An archive that stops growing sideways
Ask which takes an approved version has made redundant and move them to an archive folder in one go — so the storage bill is a decision you make rather than a surprise you get.
The questions engineers actually ask
Not prompts we invented. These are the questions working engineers and producers asked in their own sessions. Ask any of them in the assistant, on any track — pick one here to see what comes back.
On an open track
When the client has commented
Across every project
On a project
Live works in the assistant today. Soon is queued for the next releases.
Integrated −8.2 LUFS · true peak −0.1 dBTP · loudness range 5.4 LU.
- Streaming · −14 LUFS
- 5.8 LU too loud
- EBU R128 · −23 LUFS
- 14.8 LU too loud
- ATSC A/85 · −24 LKFS
- 15.8 LU too loud
- True peak · −1.0 dBTP ceiling
- 0.9 dB over
At the streaming target it gets turned down by 5.8 LU, so the extra loudness buys nothing — and the true peak can clip once a platform transcodes it.
Examples of what comes back, shortened to fit the page.
No new place to work
The agent has no library of its own. It works on the projects your team already keeps in Pibox, with the access you already gave.
It sees the project the way you do
The agent works as the person who invited it — the same projects, the same permissions, nothing more. No export, no second copy of your library living somewhere else.
You ask in plain words
In the Pibox chat next to the work, or from the assistant you already use. Through MCP, Claude, Cursor and VS Code talk to the same projects.
It does the work in the open
It creates the task, files the version, updates the properties, answers in the thread. Everything lands inside the project where your team can see that it happened.
More than a chat box
It listens, it watches video, it lets you pick the engine — and it shows you exactly where the credits went.
Talk instead of typing
Send the assistant a voice message instead of typing, and point it at the files you want it to read. Before it does anything, it shows you the plan.
Find the moment in a video
Ask where something happens and jump straight there. It transcribes what is said, with speakers, describes what is on screen, and searches across both.
Choose the model
Gemini runs by default. For longer, harder work, switch to a stronger model — Claude Opus, Sonnet 5 or Gemini Pro — and on Enterprise, the most capable model we offer.
Credits you can read
One pool for the whole organisation: a daily bonus first, then the monthly allowance from your plan. Uploads, comments and versions never spend credits, and nothing is charged twice.
The agent and the MCP server come with Pro and up. See what each plan includes.
Bring the assistant you already use
Pibox speaks the Model Context Protocol, the open standard for connecting assistants to real tools. Point your client at our endpoint with a personal token and it is working inside your projects — no plugin to install, no glue code to write.
https://mcp.pibox.com/mcpSetup guideClaude, Cursor and VS Code are configured today. Anything else that speaks MCP works the same way.
One direction is running in production today. The other is being built — what lands there is listed further down, with no dates we cannot keep.
Everyone on the project, not just the one who set it up
Sound or picture, staff or freelance, in the studio or three time zones away — the agent answers from the same project everybody is already working in.
Mixing & mastering engineers
Every note on every revision in one place, with the timecode it belongs to — and the round closing itself into tasks instead of a list you keep in your head.
Producers & artists
Ask what is left before the release instead of asking five people. Versions, approvals and who is still sitting on something, in one answer.
Content & project managers
Deadlines, deliverables and approvals across every project you run, without chasing them through chat, email and three spreadsheets.
Video editors & post teams
Frame-marked notes, a clean version trail, and cutdowns that keep the history of what they came from — for picture the same way it works for sound.
Studios & label operations
Delivery specs verified before anything ships, metadata filled in where it is missing, and the repeat questions answered without a person in the loop.
Agencies & their clients
Clients comment where they watch and listen. The agent turns that into work on your side, so nothing is lost between the review link and the edit.
Unreleased work deserves clear rules
Most of what passes through Pibox is not public yet — and some of it never will be. So the boundaries are simple enough to repeat to a client.
The same rules are written into our Terms (section 10) and Privacy Policy, including which providers run the models and where.
It acts as you, never above you
The agent inherits your access and nothing else — projects you cannot open, it cannot open. It never deletes anything, or shares it outside the team, without you confirming.
You see the plan before it acts
It shows what it is about to do, and you can call it off. What it remembers about you is yours to read and clear.
Your content is not training data
Pibox does not use your files, messages or transcripts to train models, and the model providers it calls do not either. A request sends only the part it needs, and the provider keeps it no longer than the request takes.
Everything leaves a trace
Tasks it creates, files it moves, versions it shares — all of it lands in the project with a name on it. Nothing happens in a private window.
What lands next
Listed here rather than mixed into the page above, so it is obvious which half you can use this afternoon and which half is still being built.
- The questions above as one-tap buttons on every track, so nobody starts from a blank chat
- Your team’s own MCP servers, connected from settings, with an admin approving each one
- Bots you invite to a project the way you invite a person, answering in the team thread
- A project brief the assistant keeps in mind: genre, references, deadline, what ships per song
Questions
What can the agent do today?
Find files, projects and moments by name, metadata, tags and transcript text; read the loudness and delivery checks on any audio file; work with folders, versions, statuses, playlists and share links; create and assign tasks, reply to comments and remind people of deadlines; transcribe audio and video; tag tracks by genre, mood and instruments; build xlsx reports; and answer questions about Pibox itself.
Does it work for video as well as audio?
Yes. Versions, frame-marked comments, review threads and tasks work the same for a mix or a cut. On Team and Enterprise it can also transcribe a video, describe what is on screen and search across both, so you can ask where something happens and jump straight there. AI tags for genre, mood and instruments are for audio.
Do we have to change how we work?
No. The agent reads the projects you already have — the same folders, versions and comment threads. If your team keeps working exactly as before, it simply has more to answer from.
Which models does it use?
Gemini runs by default. Stronger models — Claude Opus, Sonnet 5 and Gemini Pro — are on Team and Enterprise, and the most capable one on Enterprise. Anthropic and Amazon models run through Amazon Bedrock in the EU, Google models through Vertex AI, on European endpoints wherever the provider offers them.
Can I use Claude or Cursor instead of the Pibox chat?
Yes. Pibox runs an MCP server at https://mcp.pibox.com/mcp. Point any MCP client at it with a personal API token and it works inside your projects from wherever you already are.
Do you train AI models on our files?
No. Pibox does not use your content to train machine-learning models, and the model providers it calls do not either. When a request needs a file, only the relevant part is sent, processed on Pibox’s instructions, and kept no longer than the request requires. Section 10 of the Terms and the Privacy Policy have the details.
What does it cost?
The agent and the MCP server come with Pro and up. AI work spends credits shared across your organisation — a daily bonus first, then the monthly pool from your plan. Ordinary work in Pibox never spends credits, and a track that is already tagged or a transcript that is already done is never charged twice.
Planning a rollout across a studio or a label? Book a walkthrough and we will set it up with you.
Put it on one project and see what comes back
Upload a round of revisions, ask it what changed, and let it turn the comments into tasks. That is a five-minute test with your own work.