Welcome to Mipsy

Mipsy is your scratchpad for talking to AI models that run on your own Mac. Mipsy was designed by Michael Mangelsdorf at Swirlsea.org in 2026.

“Mipsy is a tool that I like using and which I personally find helpful for a variety of tasks. I would love to hear from you if this app is useful to you over time! In the case that you would like to show your appreciation, I have a printshop where I sell decorative prints I make if you are interested.”

Mipsy is one half of Acopa; the other half is Roller, a stage for playing scenes with a model. Click Roller next to Mipsy at the top left to switch (and Mipsy there to come back; in the ZIM view the two are hidden, so leave it first). Roller has its own Help. Both share the models, the speech settings, the port and one password, and Roller shows your Vault stories below its World tree.

In the words of the author of SQLite, which incidentally is a central piece of this app:

The author disclaims copyright to this source code. In place of a legal notice, here is a blessing: May you do good and not evil. May you find forgiveness for yourself and forgive others. May you share freely, never taking more than you give.

• Getting started
• The text boxes
• Asking a question
• Your files
• Running scripts
• Stories
• ZIM files · Your own books
• Long texts
• Thinking (Fold)
• Listen & Speak
• Getting models
• Presets & Config
• Handy shortcuts

Getting started

At the top of the middle column you pick who answers you:

• LM Studio uses the models you have loaded in the LM Studio app. Its server needs to be switched on there (Developer tab → Start server).
• Local lets Mipsy run a model by itself, no LM Studio needed. The first question takes a few seconds while the model wakes up.

Then choose a model from the list, type something into Query and press Query. That's it.

The text boxes

The middle column has three boxes:

• System: tells the model how to behave, for example “You are a friendly editor. Answer briefly.”
• Context: the conversation so far, plus anything you want the model to know. Questions and answers collect here.
• Query: what you want to ask next.

The column on the right holds six parking boxes, handy spots to keep snippets you want to reuse.

Below the file tree on the left is the STDOUT box. It shows what a script prints (see Running scripts), but otherwise it's a box like any other: type in it, drag text in and out, keep things there.

All boxes work the same way:

• Double-click to select everything in a box.
• Drag selected text into another box to copy it there (it's never moved away).
• The little dotted handle at a box's bottom right drags a copy of the whole box.
• Everything you type is saved automatically and is still there next time.
• Drag the gaps between columns or boxes to make them bigger or smaller.

Asking a question

Press Query (or ⌘↵). Your question and the answer are added to Context, so it grows into a running conversation the model can see. While an answer is coming in, the button says Stop. Press it (or Esc) to stop early; whatever already arrived is kept.

The first question after picking a model takes longer: the model has to be loaded into memory first. A window shows how far it has come (for Local models step by step, with how much is already in memory) and closes as soon as the model gives its first answer; Cancel stops the loading. After that the model stays loaded, and the next questions start right away. While you wait, the status line says what is happening: Loading the model, Reading the prompt (a long Context takes a while to read, even with the model loaded), Thinking (models that reason first, with a word count), then Generating the answer.

The bar at the bottom shows how much of the model's memory your Context uses. If it turns red, the text is too long for the model. Remove some, or pick a model with more room.

Clear empties all the boxes, STDOUT included, and the ZIM reader with its back/forward history, closes everything in the ZIM tree, and makes the six parking boxes equally tall again. The Payload switch shows exactly what will be sent to the model, which is useful when an answer surprises you.

Your files

The left side shows your files and folders. Every file remembers its last 10 versions, so you can't lose a good draft by accident.

• Save text: select text in any box and drop it on a file. It becomes the newest version. This also works while its versions are shown; the version rows themselves can't be overwritten. Drop it on a folder to make a new file.
• Use text: drag a file into a box to insert its newest text. Drag a folder to insert all of its files, one after another.
• Older versions: click the little arrow next to a file to see its versions. Drag one into a box to get it back, or click its × to remove it.
• New, rename, delete: right-click. The icons at the top also make new files and folders.
• Organise: drag files and folders onto other folders to move them.
• Bring files in: drag files or whole folders from Finder onto the tree.
• File names include their type, like notes.md or script.sh.
• Crush (the small grey link at the top) tidies up everything: each file keeps only its newest version and all older versions are deleted. You'll be asked first, because this can't be undone.

A folder as one document

Dragging a folder into a box puts its files together into one text. So you can split a long text into parts, keep each part as its own file, and edit just the bit you care about, with its own version history.

Take a C program. Make a folder calc.c and put one file in it for each piece:

calc.c/
  01 header comment.c
  02 includes.c
  03 struct stack.c
  04 push and pop.c
  05 main.c

Now you can drag just 04 push and pop.c into Query, ask the model to improve it, and drop the better version back onto that file. Drag the whole calc.c folder into a box to get the complete program again.

How the parts are put together:

• Order: by file name, just like in the tree. Numbers count as numbers (2 comes before 10). Starting names with 01, 02 … keeps the order easy to control.
• Newest versions only: each file adds its newest version. To go back to an older version of one part, drag that version into a box, select its text and drop it on the file's own row; it becomes the newest.
• Joined with one blank line between parts. No file names are added, so the result is exactly your text.
• Only the files directly in the folder: folders inside it are left out. Use this to park unused parts or alternatives in a subfolder.
• Always current: a very long folder dropped into Context becomes a placeholder (see Long texts). Mipsy reads the folder again for every question, so your latest edits to any part are always used.

Running scripts

Any file ending in .sh is a little program you can run. Right-click it and choose Run .sh. Right-click one of its older versions to run that version instead.

• Whatever it prints shows up in the STDOUT box below the file tree, live while it runs, errors included. It replaces what was there before.
• It can read your Query, Context, System and parking box texts. Each one is handed over as a little file: $MIPSY_QUERY, $MIPSY_CONTEXT, $MIPSY_PROMPT and $MIPSY_PARKING_1 to $MIPSY_PARKING_6 (top to bottom). Nothing else is passed in; the script picks what it needs.
• The script can also change those files. When it finishes, the boxes whose files it changed show the new text. That way a script can work like a little add-on for Mipsy.
• It runs in the folder where it sits in the tree, so it can read and write the files next to it.
• While it runs, a red Kill button appears at the bottom. Press it to stop the script and anything it started.
• Didn't mean to? Press ⌘Z in STDOUT right afterwards to get its old text back.

A tiny example, a script that runs whatever Python code you typed into Query (the result appears in STDOUT):

python3 "$MIPSY_QUERY"

One that counts the words in Context:

wc -w < "$MIPSY_CONTEXT"

And one that turns the first parking box into capital letters:

tr a-z A-Z < "$MIPSY_PARKING_1" > up.tmp
mv up.tmp "$MIPSY_PARKING_1"

Let an AI write a script for you

You don't have to write scripts yourself. Copy the text below into Query, add what the script should do at the end, and ask your model (or any other AI). The text explains to the AI how Mipsy works, in the AI's own technical language, so you can simply talk about “Parking Box 2” or “the Context box”.

Write a shell script for Mipsy, a local LLM scratchpad web app (Node.js server, macOS). Runtime facts:

UI LAYOUT (three columns, left to right)
- Left: the file tree (a view of the directory $MIPSY_FILETREE on disk), and below it a text pane "STDOUT" that receives the script's output.
- Center: three text panes stacked vertically, top to bottom:
  1. "System": the system prompt.
  2. "Context": the chat transcript. Turns start with a line "### User" or "### Assistant"; the text before the next such line is that turn's content.
  3. "Query": the next user message.
- Right: six "Parking Boxes" stacked vertically, numbered 1 (top) to 6 (bottom). Free-form text storage.

EXECUTION
- The user right-clicks the script in the tree and chooses "Run .sh". Interpreter: /bin/sh, unless line 1 is a shebang (e.g. #!/bin/bash or #!/bin/zsh; it is split on whitespace, extra arguments allowed).
- Working directory: the directory containing the script, inside $MIPSY_FILETREE.
- stdin: empty (closed immediately). All input comes from the files below.
- stdout and stderr: streamed interleaved into the STDOUT pane as they arrive, replacing its previous content. Cap 5 MB, then the script is killed. A non-zero exit appends "[exit N]". The user may kill the script at any time (SIGTERM to the process group, SIGKILL after 2 s).
- Environment variables holding absolute paths of temporary UTF-8 text files (hidden dotfiles in the working directory):
  $MIPSY_QUERY      = Query pane
  $MIPSY_CONTEXT    = Context pane
  $MIPSY_PROMPT     = System pane
  $MIPSY_PARKING_1 … $MIPSY_PARKING_6 = Parking Boxes 1 … 6
  Read them like normal files. To change a pane, overwrite its file (e.g. write a temp file, then mv it onto "$MIPSY_PARKING_2"). After the script exits (any exit code; not when killed), each file whose content changed is loaded into its pane; unchanged or deleted files leave the pane as it was. The files are deleted afterwards. STDOUT has no file; it only receives the output.
- $MIPSY_FILETREE: absolute path of the file tree's root directory.

FILE TREE CONVENTIONS (directory $MIPSY_FILETREE)
- Folders are plain directories. Names starting with "." are hidden from the tree.
- Each file in the tree is a version stack of up to 10 versions, stored as separate files with the version number before the last extension: notes--0.md (newest) … notes--9.md (oldest). Without an extension: notes--0 … notes--9. The lowest number present is the newest version. Logical file names may not end in "--<digit>".
- A plain file without a suffix (notes.md) is a single-version file and counts as version 0.
- To read the newest version of "notes.md": use notes.md if it exists, else the lowest-numbered notes--N.md.
- To add a new newest version: write the plain file notes.md next to the existing notes--N.md files. When the tree refreshes (automatically after the script ends), Mipsy renames it to notes--0.md, shifts the older versions down by one and drops anything beyond 10 versions. Do not renumber the files yourself.
- To create a new file: write the plain name (e.g. summary.txt).

REQUIREMENTS
- Return one complete script, POSIX sh unless a shebang says otherwise, quoting all paths ("$MIPSY_CONTEXT" etc.; paths may contain spaces).
- Print only what should appear in the STDOUT pane; put results meant for other panes into their files.

TASK

Scripts can do anything you could do in Terminal, including deleting files. Only run scripts you understand.

Stories

The book icon in the bottom-right corner of the tree view switches to your stories: a place for longer writing, organised as folders → stories → chapters → scenes → moments.

• Public stories are always visible. The Vault is locked with a password: type it and click set after each start. Please remember it; it can't be recovered. It is the one password of Acopa: it also unlocks Roller (and Roller's password field unlocks the Vault here).
• In Roller: the Vault (not Public) also appears in Roller mode, below Roller's World tree. Drag a row from it into Roller's boxes to copy its text, or drag text selected in Roller onto a scene, moment, synopsis, story or folder, just like here.
• Right-click to add folders, stories, chapters and scenes. Drop text on a scene to add a moment.
• Drag a chapter or scene into a box to use its text. A scene can have a short synopsis; tick its box to use the synopsis instead of all the moments.
• Folders and stories organise, chapters tell. Folders and stories sort your work and can hold their own notes, such as a premise, a character list or ideas. Dragging one into a box gives just those notes, not everything inside it. The narrative itself lives in the chapters: dragging a chapter brings in all of its scenes, in order. So you decide exactly which chapters go to the model.
• Right-click a scene → Scene Editor to edit all its moments side by side.

Synopses: keep big chapters small

A model can only take in so much text at once. Once a story grows, dragging whole chapters into Context quickly fills the bar at the bottom. Synopses help here: give a scene a short summary, or write an LLM prompt to generate one for you, and you choose with a checkbox per scene whether the model sees the full text or just the summary.

• Write a synopsis: drop text on the scene's synopsis row. A model can write it for you: drag the scene into Query, ask for a summary, and drop the answer on that row.
• Switch it on: once a scene has a synopsis, a checkbox appears next to it. Ticked, the synopsis stands in for the scene's moments; unticked, the full moments are used again. Nothing is deleted either way.
• Where it counts: when you drag a chapter into a box (a story gives only its own notes, see above), ticked scenes come in as their synopsis, marked “(synopsis)” in the scene heading, and all others in full. Dragging a single scene always gives all its moments; dragging the synopsis row gives just the synopsis.
• See the saving: the token numbers in the tree count a ticked scene by its synopsis, so you can watch the chapter shrink as you tick.

A good mix: summaries for the early chapters the model only needs to remember, full text for the scene you're working on and the ones right before it.

ZIM files (offline Wikipedia & co.)

A .zim file is a whole website packed into one file for reading offline: Wikipedia (all of it, or a topic), Wiktionary, Stack Exchange sites, Project Gutenberg and many more. Get them from library.kiwix.org. Many come in three sizes: maxi (full articles with pictures), nopic (full articles, no pictures) and mini (only the opening part of each article).

• Put .zim files into the ZIM folder in the Acopa folder. Subfolders are fine: they show up as folders, which is handy for sorting. Another folder can be picked under Config → ZIM library. The ZIM folder in the Acopa folder stays in the tree anyway, as Local ZIM folder below your library.
• A large library is split into three groups, about a third of your ZIM files (and folders) each, in alphabetical order: the group rows say which letters they cover, for example A–J. They are closed when Acopa starts; open the one you need. Selecting a group makes searches cover just that group.
• Books: drag an .epub book from Finder onto the ZIM tree, and it becomes a ZIM file you can read and search like any other. See Your own books (EPUB) below.
• The small gray ZIM link in the bottom-right corner of the tree switches the left pane to your ZIM library, and the Context box to a reader. Click ZIM again to get your file (or story) tree and Context back. Context isn't touched meanwhile.
• Inside each ZIM: Main page, A–Z (all titles, in groups that split further, A → Ab, Ac …, until they're short enough to list), and your latest search results. Click an article to read it. Full articles can be expanded to their sections; click a section to jump there.
• Search with the field at the top: typing finds titles, Enter searches the full text. It searches the ZIM (or folder, or group) selected in the tree, or every ZIM if nothing is selected. The × on a results row removes it. See Searching well below.
• The reader shows text only: headings, paragraphs, lists and quotes. Tables, footnotes, reference lists, formulas and the like are left out on purpose. Underlined words are links to other articles: click to follow, ← and → (or ⌘[ / ⌘]) to go back and forth. Dotted underlines are links to the web: double-click to open them in a new window. A 🖼 line stands for a picture: hover over it to see it (only in ZIM files that include pictures).
• Asking about an article: while the reader is shown, it is the context. Query sends your System prompt, the article and your Query to the model, and the answer appears below the article (your regular Context isn't used or changed). Each question is asked fresh, without earlier answers, which suits summaries and pulling out facts. The bar at the bottom counts the article. The answers stay with the article while Mipsy is open; hover one and click × to remove it. To ask about your regular Context instead, click ZIM to leave the ZIM view.
• ZIM lookup: select a few words (fewer than 20) in any box or in the reader, right-click and choose ZIM lookup: it searches your ZIM files for them, just like typing them into the search field and pressing Enter (and switches to the ZIM view if needed). Every word of the selection has to be found, so two or three telling words work better than a whole sentence.
• Citing: right-click a parking box and choose Place reference to put in a line saying where the text came from, for example [Sea level rise § Impacts — Wikipedia, revision of 2026-10-08 · https://en.wikipedia.org/w/index.php?…&oldid=… · ZIM wikipedia_en_climate-change_mini_2026-10]. The section is the one where you last selected or clicked in the reader (or the one at the top of the reader). Wikipedia links point to the exact version in your ZIM file; for other ZIMs the link is added when one is known, otherwise the reference names the page inside the ZIM (› path). The menu appears while an article is open in the reader; hold ⌥ for Safari's usual menu.
• Taking text out: select text in the reader and drag it into a parking box (or any box). It arrives as plain text, without the 🖼 lines. Or drag an article or a section from the tree into a box to get all of its text.

Searching well

Where it searches: in the ZIM file, folder or group that is selected in the tree, and only in the ZIM files that are ticked. Each ZIM file has a checkbox: untick the ones you don't want in your searches (Stack Exchange sites, say, when you're after encyclopedia articles); they are shown faded. A ZIM file you select itself is always searched. Clicking an article selects its ZIM file, so the next search stays in that file; the search field says where it searches (“Search Wikipedia …”, “Search 3 ticked ZIM files …”). To search all ticked files again, click the top ZIM row of the tree. The ⌃⌄ button at the top closes everything in the tree again (Collapse all).

The two kinds of search work differently:

• While you type (from 2 letters on), titles are searched. The last word may be unfinished: clim already finds Climate change.
• Enter searches the whole text of every article. All your words have to appear in an article, in any order and anywhere in it, not only the first one: ocean coral acid finds only articles that mention all three. The best matches come first, up to 50 per ZIM file, each with a short excerpt; the gray number is an estimate of how many articles match in all.
• Word forms count as the same word in ZIM files of a known language (glaciers also finds glacier), and very common words such as the or of are skipped.
• For more precise searches: "sea level" in quotes finds that exact phrase; OR and NOT in capitals combine words (coral OR reef, reef NOT barrier); -word leaves out articles with that word.
• Some ZIM files have no full-text index. There Enter searches titles instead, and a note says so.

Nothing found? Use fewer words: each extra word must also be in the article. Two or three telling words usually work best. And check where it searched (see above).

Your own books (EPUB)

E-books in the .epub format can be turned into ZIM files, so you can read them in the reader, search their full text, drag passages into your boxes and ask the model about them, just like Wikipedia articles.

• How: switch to the ZIM view and drag one .epub file from Finder anywhere onto the ZIM tree (it lights up). A line Converting “…” into a ZIM file… appears at the top; after a few seconds (longer for big books with many pictures) the new book shows up, selected, under Local ZIM folder.
• Where it goes: always into the ZIM folder in the Acopa folder, whatever library folder you picked in Config. The file is named after the author and title, for example Herman Melville - Moby Dick.zim. If a file of that name is already there, Mipsy asks: Save replaces it, or type another name first to keep both.
• What's inside: a Main page with the title, author, publisher, cover, the book's description and its table of contents (click a chapter to open it); every chapter as an article, named as in the book's table of contents, so A–Z lists the chapters and each chapter expands into its sections; and the book's pictures (hover a 🖼 line). The book's cover becomes the icon in the tree.
• Searching: the book gets a full-text index, so Enter finds words anywhere in it, with word forms in the book's language (whales also finds whale). Select the book in the tree to search only in it.
• Limits: one book at a time, up to 70 MB. Books with copy protection (DRM), as sold by many shops, can't be converted; DRM-free books (Project Gutenberg, Standard Ebooks, many publishers' own shops) work. Layout, fonts and tables are left out, as for all ZIM files in the reader.
• Removing a book: delete its .zim file in Finder (Acopa folder → ZIM); the tree follows when you click Refresh or reopen the ZIM view.
• The ZIM file is stored as it is, not encrypted, like an exported file. Keep that in mind for private documents.

The ZIM view needs libzim, which acopa.command installs once by itself. If the tree says libzim is missing: in Terminal, in the Acopa folder, run export PATH=/opt/homebrew/bin:$PATH and then npm install (it needs Node 22 or newer, and Apple's command-line tools: xcode-select --install).

Long texts

When you drag something very long (over 1,000 lines) into Context, Mipsy puts in a short placeholder like ⟦#3 📄 novel · 4,402 lines⟧ instead, so everything stays fast. The full, current text is still used when you ask a question. Double-click the placeholder if you want the whole text in the box.

Thinking (Fold)

Some models “think out loud” before they answer. With Fold pressed in, that thinking is tucked away into a one-line placeholder so Context stays readable. Press Fold again to see all of it, or double-click a single placeholder to open just that one. The thinking is never sent back to the model.

Listen & Speak

• Listen lets you dictate into Query: click, talk, click Stop. Mixed languages are fine.
• Speak appears when you select a few sentences. It reads them aloud in a natural voice.

Both run entirely on your Mac and need a one-time setup.

Getting models

A model is one big file that holds an AI's “knowledge”. Mipsy's Local backend runs models in the GGUF format (files ending in .gguf). It finds them in two places:

• Acopa's own models folder (inside the Acopa app folder; Roller uses it too). Give each model its own subfolder, named after the model, and put the .gguf file in it:

  models/
    Mistral Nemo 12B Instruct/
      Mistral-Nemo-Instruct-2407-Q4_K_M.gguf
    Gemma 3 12B/
      gemma-3-12b-it-Q4_K_M.gguf

  The folder name is what you'll see in the model list. If a folder holds several .gguf files (say a Q4 and a Q8 version), each shows up as folder/file. Models in this folder are only for Local; LM Studio doesn't see them.
• LM Studio's model folder. Everything you download in LM Studio shows up in Local too, under the same name.

After adding a model, press the little refresh arrow next to the model list (with Local selected).

Where models come from: Hugging Face

Hugging Face is the big public library of AI models, a bit like an app store where anyone can publish. The companies that make models (Mistral, Google, Meta, Alibaba's Qwen and others) put them there, and volunteers convert them into GGUF files for programs like Mipsy and LM Studio. Downloading is free, and you don't need an account for most models.

• Find a GGUF: search for the model's name plus “GGUF”, e.g. Mistral Nemo Instruct GGUF. Converted versions usually come from well-known uploaders such as bartowski, unsloth, lmstudio-community or mradermacher.
• Pick a size: one model is offered in many “quantizations”, such as Q4_K_M, Q5_K_M or Q8_0. Smaller numbers mean smaller files and slightly lower quality. Q4_K_M is a good balance. The file has to fit comfortably in your Mac's memory; see How big a model can my Mac run? below.
• Pick the “instruct” or “chat” version. “Base” models aren't trained to answer questions; they just keep writing.
• Download it from the Files tab (the small download arrow next to the file) and move it into a new subfolder of models.
• Big models split into parts (…-00001-of-00003.gguf and so on): put all parts in the same folder; Mipsy lists them as one model. Files with mmproj in their name are picture add-ons, which Mipsy doesn't use.

If a model answers with nonsense or wanders off-topic while others work fine, its file may be missing its chat template, the instructions for how a conversation is laid out. Try the same model from another uploader.

How big a model can my Mac run?

The model has to fit into your Mac's memory (RAM) while it runs, with room to spare for macOS, your other apps and the conversation itself. Memory, not processor speed, decides which models you can use.

• “12B” means 12 billion parameters. Parameters are the numbers the model learned during training; more of them usually means more knowledge and better reasoning, but also a bigger file and slower answers. Common sizes are 1B, 3–4B, 7–8B, 12–14B, 24–32B and 70B.
• The Q number decides how many bytes each parameter takes. Squeezing the numbers into fewer bits is called quantization. Rough rule: a Q4 file needs a little over half a gigabyte per billion parameters, Q8 about one gigabyte, the unsqueezed F16 two. So a 12B model is about 7 GB as Q4_K_M, 13 GB as Q8_0 and 24 GB as F16. The file size on Hugging Face tells you directly.
• The conversation needs memory too. On top of the file, the model keeps the text it is reading in memory. How much depends on the Context size in Config: the default 8,192 tokens add roughly 0.5–1.5 GB; very long contexts can add several gigabytes.
• macOS lets the graphics chip use only part of the RAM, about two thirds to three quarters of it, and you'll want your browser and other apps to keep running.

A practical guide for Q4_K_M files on Apple Silicon Macs:

• Too big? The model fails to load, or answers crawl out a word every few seconds while the Mac becomes sluggish. Pick a smaller model, a smaller Q (Q4 instead of Q8), or lower the Context size.
• Spare memory? Rather than a bigger Q of the same model, a larger model at Q4 is usually the better use of it.
• Names like “30B-A3B” are “mixture of experts” models: all 30B parameters must fit in memory, but only 3B are used for each word, so they answer about as fast as a 3B model while knowing much more.
• To see your Mac's memory: Apple menu → About This Mac.
• Mipsy checks for you: after a Local model has loaded, Mipsy compares its Context size with what the model and your Mac's memory allow (from llama.cpp's own figures). If the setting is far too large (the Mac would start swapping), larger than the model was trained for, or much smaller than it could be, a short note explains it and offers a better size; one click sets it in Config, and the model is loaded with it at the next query. “Keep” keeps your setting, and Mipsy doesn't ask again for that model and size.

Presets & Config

Config holds the settings that shape the model's answers, such as Temperature (higher means more creative, lower means more focused). A set of settings is a preset; pick one from the list next to Config.

• Change values and press Save to update the preset. Type a new name to save a new one (up to 5).
• Reset fills in the original values again. Press Save to keep them.
• Below the presets, under For all of Acopa, Config has the settings that apply to Mipsy and Roller alike: the memory size (context) for Local models, the ports, the folder of your ZIM library and the voice settings for Speak. Roller's own Config only has its presets and nudge rules, and links here for the rest. Roller's presets are separate from Mipsy's.

Ports

A port is the “door number” a program on your Mac listens on. You only need to change it if another program already uses one of Acopa's, for example when Acopa won't start, or Speak, Listen or a Local model reports that its port is in use. There's one setting, under Config → Ports, saved right away:

• Acopa (normally 3230) is the address of this page, http://localhost:3230 (Mipsy and Roller both).
• The helper programs take the next three numbers: Speak +1, Local models +2 and Listen +3 (normally 3231, 3232, 3233). Config shows them next to the field.
• A new number is used the next time Acopa starts: close its Terminal window, double-click acopa.command again, and Safari opens at the new address. Update any bookmarks.
• Use a number between 1024 and 65532.

If a Speak, Listen or Local helper from an earlier Mipsy is still running (for example after a crash), Mipsy stops it by itself. If another program has the port, Mipsy tells you which one.

Handy shortcuts

• ⌘↵ asks your question. Esc stops an answer.
• Double-click a box to select all of its text.
• In the file tree: Enter renames, Delete deletes.
• Right-click almost anything in the trees to see what you can do with it.