0.6.3 — Long chats stop hitting the wall

Long chats stop hitting the wall.

A chat used to grow until the model refused it outright, and the only way out was
wiping its memory entirely. Most of that growth wasn't your conversation — it was
the same document being re-sent with every single message. That's fixed, and when
a chat does fill up you now have something gentler than a reset.

What's new

Compact — click the ring next to the model picker and pick Compact. The chat
writes itself a summary, hands it to a fresh session, and carries on. You keep the
thread; the model travels light. Clear is still there when you want a clean slate.

Pick the integrations a chat can use — the plug icon in the composer scopes
integrations per chat. Every connected integration used to be loaded into every
conversation, which is a lot of baggage for a chat about your notes. Fewer
integrations means faster, sharper answers.

Your actual model list — the model picker now reads what your Claude account
really offers instead of a fixed three, so newer models show up as soon as you have
them. Models that support it also get a reasoning-effort setting, per chat.

Stop actually stops — hitting Stop ends the current answer without throwing
away the conversation. You can just keep talking.

Fixes

• The document you're viewing is sent once, not re-sent with every message
• A chat whose memory expired now says so, instead of quietly answering as if it
  had forgotten nothing
• A turn that fails explains itself instead of leaving an empty bubble
• #tag mentions list the files they matched rather than dumping all of them in
• Big images are resized before they're attached
• Fixed the integration tool set vanishing between turns

One thing to know

Integrations now come only from Pilea Brain. If you had MCP servers configured
outside the app — in a workspace .mcp.json or your personal Claude config — those
no longer load into chats here. Anything you set up under Integrations works
exactly as before. This makes the per-chat picker above tell the truth about what a
chat can reach.

---

Requires macOS 10.12 or later. Apple Silicon and Intel builds below — most Macs
from 2020 onward want the arm64 one.

0.6.2

Pilea AI answers again when you've got MCP servers connected.

Some MCP tools describe themselves in a way Gemini refused outright, which took
the whole chat down instead of just that one tool. Fixed — and a turn that fails
now tells us about it, so we find the next one before you have to report it.

Pilea Brain 0.6.1

Pilea AI works in chats you started with Claude

Switching to Pilea AI part-way through an existing conversation used to do
nothing at all — the thinking indicator would flash and disappear, and no answer
ever arrived. Any chat that already had tool calls in it was stuck. Fixed: those
conversations now carry over properly, so you can switch models mid-thread.

When something breaks, we can now see why

Pilea Brain keeps a diagnostics log on your machine and reports crashes back to
us. If the app misbehaves, there's finally a trail to follow instead of a
shrug — and a Copy diagnostics button on the error screen so you can send us
exactly what happened in one click.

Your notes stay yours. File paths and document titles are stripped before
anything is sent, along with email addresses and access tokens. Crash reports
never include screenshots or memory dumps, and the full log stays on your Mac
unless you choose to share it.

Fixes

• The app no longer double-mounts after certain errors, which could leave the
  window showing a red error screen instead of your workspace.
• Update failures are now recorded rather than passing silently.

0.6.0 — Pilea AI, built in

Pilea Brain now has an AI assistant of its own. Ask it questions about your notes, have it draft and edit documents in place, run commands, and pull things together across your workspace — no API key, no setup, it's just there.

What's new

Pilea AI — pick it in the agent panel and start typing. It reads the file you're looking at, follows your @-mentions and #tags, and can create and edit files directly in your workspace.

A calmer chat — the whole agent panel got a rebuild. Cleaner messages, better code blocks, attachments that actually show up, and suggestions you can click.

Faster, steadier sync — the backend moved to new infrastructure. Sync settles quicker after launch and holds up better on flaky connections.

One thing to know

You'll need to sign in once more after updating. The move to the new backend means a fresh sign-in — your files and workspaces are untouched, it's just the login. Sorry for the interruption; it's a one-time thing.

Fixes

• The chat no longer pastes your whole open document into the message you sent
• Fixed a loop where the assistant repeated the same answer instead of using what a tool returned
• Attached files and images now reach the assistant instead of being dropped
• Fixed a crash that could blank the chat panel mid-reply
• Deep links open the right window again

---

Requires macOS 10.12 or later. Apple Silicon and Intel builds below — most Macs from 2020 onward want the arm64 one.

0.5.0 — Transcribe anything, and know who said what

Turn a recording into a transcript, on your own machine — and know who said what.

Transcription

Drop in audio or video, get a transcript. Right-click any recording in the explorer, or open it and hit Transcribe. Everything runs locally with whisper.cpp — nothing is uploaded, and it works offline once the model is downloaded.

• Six model sizes, from a 75 MB tiny model to a 3 GB large one, downloaded on first use
• Norwegian recordings get specialised models (NB-Whisper) trained by the National Library of Norway
• Pick the language, or let it detect one — worth setting, since auto-detect on Norwegian audio sometimes decides it's English and produces confident nonsense
• Live preview as the words land, and the transcript is saved as markdown next to your file

Who said what

Turn on Identify speakers and an interview reads as a conversation instead of a wall of timestamps. Speakers are labelled in the order they first talk, so you can rename them once and be done.

This holds up on long recordings. Speaker detection normally falls apart past about fifteen minutes — two voices quietly collapse into one — so this splits a recording into overlapping chunks, takes a reference clip of each voice, and uses those clips to keep everyone straight from the first minute to the last. Measured on hour-long test conversations: 100% correct on English, 99.9% across a German/French/Japanese three-way. Two genuinely similar voices in similar languages are the hard case and land around 85%.

Tell it how many people are talking — it's a much easier question to answer than it looks, and it makes a real difference.

Editor and explorer

• Cmd+F finds text in the document
• Menu shortcuts are rebindable, and there's a shortcuts overview under Help
• New File has a proper menu entry
• Single newlines are treated as soft line breaks, so pasted text keeps its shape
• Tables you haven't resized now hug their contents instead of stretching
• The selection bubble menu is white, not cream
• Leaving a focused folder in the explorer steps up one level instead of jumping home
• Binary files open properly instead of failing silently

---

Requires macOS 11 or later. Apple Silicon and Intel builds below.

0.4.0 — Integrations, and renames that don't lose your work

Your integrations, a much safer editor, and a canvas that does what you'd expect.

Integrations

28 apps, one click. Notion, Linear, GitHub, HubSpot, Intercom, Webflow and more connect through a browsable store — sign in through your browser, and the agent picks up what that app can do. No config files, no API keys to paste.

Fixes that were losing your work

Renaming a document no longer leaves the old one behind. Renaming looked like it created a copy: the original came back alongside the new name. Closing a document saves it first, and it was saving to the name the file had before the rename — so it reappeared moments after being moved. The same bug hid in drag-and-drop, where renaming a folder restored the whole original folder.

Deleting works. Files could come back moments after you deleted them, for the same underlying reason. Renaming and deleting now also keep your unsaved edits, scroll position and undo history intact.

@-references follow renames

Rename a file or folder and every @-reference to it updates. Links used to record only the words you saw, so once a document was saved and reopened there was nothing left saying what it pointed at. Existing links keep working exactly as before, and nothing needs converting.

Clicking a reference to a folder also used to quietly create a stray document named after that folder. Fixed.

Canvas

• Paste a screenshot, drag an image in from Finder, or use Add image
• Hold Option and drag a card to duplicate it, like Figma
• Select two or more cards to align them or arrange them as a grid
• Set card text size from xs to xxl — headings and code scale with it
• Website cards preview the real desktop page instead of the mobile layout
• Image cards sit bare on the canvas, no filename bar

Files & sidebar

• Favourites — star any file or folder, then find them with is:favorite. They show on the home screen too
• Double-click a folder to focus the explorer on it, with a one-click way back
• The sidebar tabs have proper labels, search fills the panel header, and each panel has a single clear + New action
• The right-click menu is grouped instead of one long list

A snappier explorer. Deleting, creating and renaming now show up almost instantly — the explorer was waiting a flat half-second on every change, including your own.

---

Requires macOS 11 or later. Apple Silicon and Intel builds below.

Pilea Brain 0.3.4

Fixes a crash when opening files

This is a focused hotfix on top of 0.3.3.

• Opening a document no longer crashes the editor. On some machines, opening a file could throw a render error and leave the editor blank. The editor now safely handles the brief moment during a file switch when the previous document is being torn down. Everything from 0.3.3 is included.

If you were seeing a red "Render error" screen when opening files, this update resolves it.

Pilea Brain 0.3.3

The headline: PDF export got a whole new engine. Plus a big round of editor and file-explorer polish.

Beautiful PDF export

• A new export engine — documents render through a bundled typesetting engine (Typst) with print-shop quality: smart page breaks, a real table of contents with page numbers, tables that continue across pages with their header repeated, footnotes at the page bottom, and native math typesetting.
• Everything exports — charts render as real charts (bar, line, area, pie — in Pilea colors), images (both ![[wikilink]] and markdown styles), callouts, task lists with drawn checkboxes, code with syntax highlighting, [toc], emoji.
• Export options where you need them — the Download-as-PDF button opens a small dialog: engine (Beautiful / Quick), cover page on/off, contents on the cover, page numbers, page size. No more digging through Settings for a one-pager.
• Fixed a save bug that could quietly rewrite tables as HTML and drop chart blocks from documents. Files now round-trip cleanly.

File explorer

• Sort your files — by name or date, folders-first, from a button next to the (renamed) "Search files" field, with an indicator when a non-default sort is active.
• Native OS file icons — optional, in Settings; crisp everywhere including tabs and Recent activity.
• Renames commit on Enter, dropping a folder into another folder no longer springs it open, and the oversized image drag-preview is now a tidy thumbnail.

Editor

• Insert images from a picker — the image button shows your workspace's images (newest first, searchable) with Browse and URL options.
• "Add to chat" — select text and quote it straight into the AI chat from the selection menu.
• Table insert icons now point the way they act, chart blocks from AI chat are editable, the fold chevron is easier to hit, and tab-bar scrollbars no longer nudge content.

Around the app

• Cleaner header — the center shows your folder path with a clear switcher cue; Feedback and Settings moved into the user menu; the formatting toolbar centers when there's room.
• Sturdier under the hood — the app recovers from renderer crashes instead of blanking, a chat session race that caused "Too many active sessions" is fixed, moved images still resolve, external links open safely, and checkbox-only file updates render immediately.

Everything from 0.3.1 and 0.3.2 included.

Pilea Brain 0.3.2

A quick follow-up to 0.3.1 focused on the update experience itself.

Updating the app

• Install updates from Settings → About — when an update is available you now get a proper "Download" button, a progress bar, and a "Restart & Install" button right there. Previously About could only tell you an update existed.
• A clearer update card — the update notification is now high-visibility (hello, tangerine) and shows what's new in the version you're about to get, not the one you're on.

Small fixes

• Code blocks no longer grow a phantom trailing blank line.
• Removed the leftover "Daily notes folder" setting from Settings → Editor.

Everything from 0.3.1 is of course included — row resizing, the table hover fix, staying signed in, and the rest.

Pilea Brain 0.3.1

A big batch of editor polish, workspace fixes, and two long-awaited table wins.

Tables, finally behaving

• Resize rows — drag any horizontal table border, just like columns. Drag up to shrink back.
• No more table jump — hovering a border to resize used to shift the whole table. Found the real cause and fixed it for good.

Editor

• Nested checklists stick — indented sub-items no longer flatten when you reopen a page. (Items flattened by the old bug stay flat — re-indent once and they'll hold.)
• Properties live in the document — a quiet, collapsible "Properties" section above your first line, like Notion. No more full-width bar.
• Write faster — empty lines hint the "/" command menu and markdown shortcuts; turn any selected text into a linked document from the right-click menu.
• True WYSIWYG width — the editor column matches paper, so print and PDF output look like what you see.
• Obsidian-style tabs — the active tab connects to the page below it; your scroll position is remembered per tab.
• Auto-save fixed — a glitch could overwrite text you'd just typed. Gone.
• Dark mode — @-links and wiki-links are readable everywhere.

Skills

• Skills are real folders now — open them, read the skill text, browse scripts, and hit run.
• Schedule any skill — a calendar-style picker (daily, weekly, custom) with a visual indicator on scheduled skills. No cron syntax needed.

Canvas

• Native canvas view with Figma/Miro-style navigation: smooth panning and zooming, grab/select tools (V / H), and per-file viewport memory.

Workspaces & sign-in

• You stay signed in — a flaky network at launch could silently log you out (and made the app ask for Google sign-in every time). Sessions now survive restarts and reconnect on their own.
• Recent local folders — the workspace switcher lists folders you've opened recently, one click to jump back.
• Cleaner sign-in — Google leads (with the proper logo); email sign-in tucked behind a link.
• Faster switching — already-synced workspaces just open (no pointless re-sync), the old workspace's tabs close, and the workspace list scrolls properly.
• Tidier home — deleted files no longer linger in Recent activity, and untouched "Untitled" files clean themselves up.

Files & chat

• Lock files and folders — locked paths are protected from edits, including by the AI.
• Import folders, drag multiple selected files at once, and copy any file's path from the explorer menu.
• Audio & video get a friendly screen with a one-click "open in the right app" instead of garbled text.
• Chat — attachments show up after sending, markdown tables render, and @-file mentions keep their formatting.