I just completed building out a master per library instance VE Pro Server project using Claude with the new AI interface features in VEP 8.1
Let me just say, I understand some people have philosophical objections to AI. Please don’t use this thread to voice those concerns, its purpose is limited to the subject title.
Secondly, some would say bypass LLM and access the API yourself. Fine, those who have that experience, confidence, and ability will of course want to do that. Again, for me that’s outside the scope of the intent of this thread.
There was a LOT for me to learn when I launched the Claude MCP API project, but after working through hurdles we arrived at a good buildout technique. Using a spreadsheet I was able to have Claude verify expected outcomes at every step, and identify and correct any inconsistencies or errors on my part.
My hope for Dorico would be to use a system like this to make the the creation and organization of expression maps less tedious and quicker to build, also the creating, naming and organization of both playing and playback techniques. Down the road perhaps there’s also a use case for player and instrument creation and organization, etc. But I’d be happy to postpone that idea for now.
Again, I’m not wanting to argue the point with those philosophically opposed to it, I just mention it after seeing its value for VEP. Of course I can’t prevent anyone from saying what they want.
For me the benefit of an MCP API for Dorico would not be so much in terms of making expression maps, but instead doing repetitive operations to the music. Being able to put for instance a sketch score on a grand staff in 4 parts and telling it to split it up for string orchestra with divided cellos and the melody doubled up in the 1st and 2nd violins. It isn’t necessarily having to do creative things but automating busy-work. Sometimes it is easy to explain what you want in natural language but it is time consuming to do the work regardless. I could see it speeding up the workflow.
The main risk as always would be the potential for a hallucination to cause things to go awry, but at least the undo capability in Dorico should help in that case.
I’ve messed around with having Claude write Dorico Percussion Maps, with some mixed results. Percussion Maps seem particularly annoying and time intensive to make, so it would be great to eventually be able to offload that to Claude.
I’m glad that others are beginning to voice a desire for a Dorico MCP. I’ve used Claude successfully to make expression maps, a great time saver; even if I’ve had to tweak and customize them, it is a great starting point. I’ve also used Claude to repair broken Dorico files, albeit with mixed results.
I’d love to have Claude (or better yet, a local-running LLM) take care of a lot of the “donkey work” connected with notating music — not the composition, not the creative aspect of score layout, but tasks like connecting sample libraries, resolving display issues, page turns, etc etc etc.
And now that VEP has broken the ice and demonstrated use cases for pro-level music software…
Most of the people doing the kinds of Dorico workflows you’re describing are probably using a paid Claude plan rather than the free tier.
The free version of Claude can certainly help with things like:
Generating expression maps
Drafting percussion maps
Writing Lua/Python/XML snippets
Analyzing Dorico project files
Suggesting orchestration workflows
However, free-tier limits tend to become frustrating when you’re working with large XML files, extensive expression maps, or long back-and-forth debugging sessions. Dorico-related tasks often involve thousands of lines of configuration data, and that quickly consumes context windows and message allowances.
Many power users appear to favor:
Claude Pro — for larger context windows and higher usage limits.
Claude Max (where available) — for very heavy professional use.
Local models via tools such as Ollama, LM Studio, or Open WebUI when privacy, cost control, or offline operation are priorities.
Your examples actually highlight what an MCP integration would be good at. Expression maps and percussion maps are essentially text-generation problems, which current LLMs can already assist with. But splitting a sketch score into orchestral parts, fixing page turns, assigning divisi, reconnecting sample libraries, resolving notation inconsistencies, and batch-editing a project are action-oriented tasks. That’s where an MCP server connected directly to Dorico could become much more powerful.
For example, a command such as:
“Take this four-part piano sketch, distribute it to string orchestra, divide cellos where necessary, double the melody in Violin I and II, keep voice-leading intact, and show me a summary of changes before applying them.”
isn’t really an AI composition task. It’s closer to a sophisticated automation macro. The AI’s role is translating natural language into a series of precise Dorico operations.
The undo system is a significant safety net here. Many musicians would probably be comfortable with AI-assisted score editing if every operation were:
Previewable,
Reversible with a single Undo,
Logged so they could inspect what changed.
That feels much more realistic than expecting the AI to make artistic decisions.
The recent integration work from Vienna Ensemble Pro and similar professional music tools suggests that developers are beginning to see AI as a workflow layer rather than a creative replacement. For notation software, the biggest gains may indeed come from eliminating hours of repetitive score-management work rather than generating music itself.
Todd, out of curiosity: why is that whole expression / percussion map stuff so important to your work?
I was under the impression that you are embedded in a very vibrant culture of live musicians, and tight deadlines. So I would have thought that even semi-careful mock-ups can never hold up with the need for speed in your community (and, frankly, with the expertise musicianship that I would expect in that field).
Basically everything I write or work on will be played by live musicians, but …
I do a lot of work in various Afro-Cuban, Brazilian, and other Latin styles, so sometimes having a basic groove playing back can be helpful when writing. I’m still afraid of the “Clave police” so hearing the rhythms in context helps to confirm everything works and nothing is crossed.
For big band commissions, I often have to provide a mockup. Generally no one really cares about the drum part in those, but some basic playback can be useful.
Drum playback can be useful in an educational context, like my jazz comp and arranging or analysis classes.
I’m mostly out of the copying game now, and generally try to throw that work to former students when it comes in, but I still get calls for that. There was a time when I did a lot of orchestral copying. (I even got a call for a last minute orchestral recording date last week, but it had to be in Sib. My Sib chops were never that strong to begin with, and I certainly didn’t have time to get them back up, so I threw the job to a friend.) Often I’m asked for a mockup for that stuff though that will need percussion playback. I copied a new 50-minute ballet in 2024 that was choreographed to the composer’s piano sketch, but then eventually the dancers rehearsed to my mockups, so obviously the percussion playback was important.
I don’t mind spending a bit of time on Expression Maps or Percussion Maps if they will produce “better” playback since these can be applied to any project going forward. I never spend any time editing MIDI on a project though as that seems like a waste of time when it’s never the final product. Whatever I get with the Expression or Percussion Maps is it.
I’ve seen some old-school percussionists absolutely cuss out an arranger for writing something cruzado. I never want to be on the wrong side of them (or the clave).
I’m using Claude Pro at the $20/month level It’s excellent at getting my business receipts and statements together for tax purposes and can be very useful with expression maps.
For serious research, Claude (and Gemini) can be very misleading. Originally, the confidence with which results are expressed encouraged me not to double-check its assertions. But I quickly learned that everything Claude asserts has to be analyzed and examined — and very carefully. The upshot is that if I’m trying to understand a complicated subject about which I know nothing — for example, Sanskrit grammar — I’d get into a lot of trouble if I depended upon the “answers” that Claude provides. The assertions are often beside the point or flat-out wrong — but very articulate!
OTOH, Claude can be very useful for tedious tasks like receipt tracking and renaming files (I have an IMSLP-file-renamer skill) and that is easily worth $20/month. And a good Dorico MCP could be very useful.
A bit too complicated and off-topic to really go into here, but in the Cuban tradition there is a two-bar rhythmic pattern that is the rhythmic foundation of any music considered to be “in Clave.” This can be 6/8 or 4/4, “3-2” or “2-3,” Rumba or Son. Certain rhythms work better on the 3 side or the 2 side, so if you continually write rhythms that sound “crossed” on the wrong side, the groove will be off and the percussionists will hate you. Once established, the clave does not change, so it is up to the arranger to change the phrase lengths to match the clave if the rhythms of the melody require it.
For example, below is Tito Puente’s Complicación where there are many phrases with odd numbers of bars to make the melody fit the clave. It starts in 2-3 Son, then has an 11 bar phrase so the vocal melody around the 0:14 mark is in 3-2. There’s then another 11 bar phrase so when the Coro comes in at 0:28 it’s on the 2 side. The Montuno around 1:36 is in 3-2, the Mambo at 1:57 is in 2-3, etc. The clave never changes though, the melodic phrase structure changes to fit the clave.
If you write something cruzado then you risk the wrath of the Clave police!
If there’s one thing essential to working with AI LLM’s it is that you have to know enough about your subject to recognize when responses are off track or just plumb wrong, and then push back. From my experience Claude and ChatGPT are pretty good about admitting their errors. Another big thing is to train it that YOU the human are the boss and IT the software is the assistant. I even insisted they change their tone with me, and they did pretty well with. AI is a useful assistant, not a replacement.
Just gave it another try, and Claude has definitely gotten better with this. Here’s what Claude generated when I asked it to create a map for VSL Studio Jazz Drums, which I don’t think I’ve seen posted anywhere yet. (The VSL Dorico Wizard doesn’t generate Percussion Maps.) I asked for PDF documentation too so I can check it over.
I need to go through, define all the techniques in the kit, and give it a whirl, but it looks pretty accurate so far. This was using the paid ($200/yr) version and the Opus 4.8 model.
This is a really cool thread. Thanks to @DaddyO and the others who’ve contributed.
As someone who’s yet to become “AI-curious” enough even to try using a tool, it’s interesting to think about how one might benefit a user’s Dorico workflow.
