Mellow v1.5.0

Mellow can now draw on your screen while it explains, and look things up on the web for you.

What's new

Screen drawings (experimental)

Mellow could already point at things with the bone. It can now draw too: when you ask about something on your screen, the bone highlights, circles or outlines what it is talking about, adds a short caption, and explains it out loud.

• Turn on Screen drawings in Settings → Abilities → Screen guidance. It is off by default and needs an engine that can see images.
• Try asking “Explain this chart”, “Circle the save button” or “Where do I click to open a new tab?”
• When something keeps changing on its own, such as a terminal that is still printing, Mellow highlights the whole area once and explains it instead of chasing each new line.
• A drawing disappears if you scroll or the screen changes.
• If you only want the bone, say “only point”, or turn off Screen drawings.

Each drawing sends a screenshot to the engine you chose and makes one planning request, plus at most one closer look when needed. Ordinary answers make no drawing requests.

Web search

Mellow can now answer questions that need the internet. There is no special phrase to learn: Mellow decides when to look something up.

• Quick facts like weather, prices, scores and the latest news are answered out loud.
• Ask Mellow to research or look into something, and a coloured bone parks in the top-right corner while you keep talking. Click it when it is ready to see what it found, with its sources. Follow-up questions use what it found.
• It is on by default. Turn it off with Search the web in Settings → Abilities → Web research.

If your engine is Gemini, OpenAI, Anthropic or OpenRouter, Mellow uses that provider's own web search, which your provider may charge for. Otherwise your question goes to free public search engines such as DuckDuckGo, and your engine writes the answer from the results. Weather comes from Open-Meteo. Nothing on your screen is sent for a search.

What's changed

• One place for Mellow's abilities. Screen guidance, Web research and Screen-aware writing now share a single Abilities page in Settings. Each has its switch up front and the details under How it works. Switches that can't be used right now are greyed out.
• The bone arrives before Mellow talks about it. Mellow now waits for the bone to reach a control before describing it. If the bone can't get there, Mellow says so instead of describing the wrong spot.
• Faster guidance. With drawings on, Mellow starts reading the screen while you are still speaking, so explanations begin sooner.

Fixed

• Your saved colours now load correctly when Mellow starts.
• Showing a drawing no longer takes focus away from the app you are using.

Known issues

• Screen drawings are experimental. Where a mark lands depends on the model, and drawings have mostly been tested with a cloud model that can see images. Small local models may place marks less accurately.
• Mellow can't draw on the desktop itself yet. With nothing open, it points with the bone instead.
• The free search can be busy now and then. When it is, Mellow says so and the bone offers Try again.

Mellow v1.4.0

Mellow can now remember what matters to you between conversations.

What's new

Personal memory

Previously, a new conversation meant telling Mellow about yourself again. Mellow can now keep a short, editable summary and use it when answering.

• Turn on Memory in Settings → Personalization. It is off by default. When enabled, Mellow starts learning from your saved chats in the background.
• Say “remember that…” to save something for the current answer and future conversations, or “forget that…” to remove it.
• Open Memory summary → Manage to review, edit, or delete what Mellow remembers.
• Add a nickname, occupation, or other details under About you. These details work even when Memory is off.

The memory summary is stored on your computer. Learning sends redacted chat text to the AI engine you selected; using a cloud provider may consume additional tokens. Memory also adds context to answers. Deleting memory does not delete your saved conversations in Sessions.

Point to several things in order

For questions with several visible steps, Mellow can now move the bone between up to three controls, explaining each one as it goes. The controls must already be visible on the current screen. Mellow does not click them or wait for the screen to change.

Licensing

Mellow's source is now licensed under AGPL-3.0, with additional attribution and name/art terms in NOTICE.

Known issues

• Learning has been tested with a cloud model, but its results with small local models and Claude Code/Codex have not yet been verified in live use. You can correct the summary in Settings.
• Multi-step pointing can still miss a control or stop early if the screen changes or a target cannot be confirmed.

Mellow v1.3.0

Agent mode works properly now, and meetings can identify individual speakers.

What's new

Identify individual speakers in meetings (experimental)

Meetings used to label everything coming out of your laptop as one Other
participant. Mellow can now separate those voices into Speaker 1,
Speaker 2, and so on.

• Turn it on in Settings → Meetings under Speaker recognition. It is off
  by default, and the switch can only be changed between meetings.
• The first time, Mellow downloads two speaker models (about 35 MB) and shows the
  progress there. Until they are ready, meetings record with the usual labels.
• After a meeting, click a speaker label in the transcript to name it, mark one
  as the same person as another to combine them, or separate them again.
  Combining is reversible, and giving two speakers the same name does not merge
  them. Regenerate existing notes afterwards so they use the new labels.
• Your microphone stays You. Speech that overlaps or stays uncertain remains
  Other participant.
• It all runs on your computer, and the voice measurements are discarded when the
  meeting ends. Nothing about anyone's voice is stored with the transcript.

This is experimental: its accuracy has not been measured against annotated real
meetings yet, so check the labels before relying on them.

Fixed

Agent mode

Using Claude Code or Codex as Mellow's brain was slow enough to look broken. Asking
Mellow to point at something took around 15 seconds and usually came back with
nothing at all.

• Pointing now takes a couple of seconds, and the bone appears as soon as
  Mellow has picked rather than waiting for it to finish speaking.
• Requests that used to return nothing now answer reliably.
• Mellow can reach menu items that stay hidden until something is expanded, like
  Spam under Gmail's "More labels".
• Agent mode now uses Sonnet by default, which was much faster and steadier than
  Haiku.
• Switching back to Cloud after using an agent no longer breaks Settings.

Optimizations

• Meeting transcripts no longer lose or duplicate words where recording windows
  meet, and phrases you genuinely repeat, such as "yes, yes", stay in the
  transcript. During uninterrupted speech the live text can lag a few seconds
  behind while Mellow waits for a natural pause; nothing is dropped, and this
  happens whether or not speaker labels are on.
• Holding the wake hotkey opens the microphone and starts listening on that same
  press, and letting go early cancels cleanly. Sleeping fully releases the
  microphone.
• Screen reading is cached and bounded, and a fresh check of the screen prevents
  Mellow pointing at something that has already moved.
• Switching to Just the pet now says what it actually does and takes effect
  immediately: AI features, the microphone and the bone pointer stop without a
  restart.
• The agent and model list shows its progress while checking, and offers Retry
  if the check fails, instead of appearing empty.
• Fixed installer signing and upload in the release scripts.

Known issues

• Pointing can still be inaccurate. The bone sometimes lands on the wrong
  control, most often with icons, crowded screens or ambiguous wording.
• Agent mode uses up your Claude Code or Codex limits faster than ordinary
  chatting in either app, because every request carries the screen with it. We
  have not found a way to reduce that yet.
• Agent mode still takes a few seconds to answer. If you want the fastest
  replies, using your own API key in Cloud mode is still quicker.
• Speaker labels are experimental and opt-in. Overlapping, very brief or similar
  voices can stay unresolved, and long meetings, noisy audio and broader accuracy
  testing are still outstanding.
• Meetings saved before this version are not relabelled.
• Screen-aware writing and speaker labels are both off by default.

Mellow v1.2.0

Mellow 1.2.0 brings screen aware dictation, Mellow customization, and auto update alerts, plus a much wider catalog of things he can open for you.

What’s new

Screen aware dictation

• Put your cursor in any text box, hold Ctrl + Shift + Space, and say what you want written.
• Dictate word for word, or ask for a draft: “reply saying I’ll be twenty minutes late”.
• Works in documents, email, chat apps, and terminal prompts.
• Ask again to revise the draft, as long as you have not touched the text yourself.
• It never presses Enter, never moves your cursor, and hands you a copyable draft if it cannot confirm the text landed.
• Off by default. Turn it on in Settings → Screen-aware writing.

Mellow customization

• Click a part of Mellow in the preview, then pick a colour for it.
• You choose three things: the fur, the patch, and the blush.
• The shading follows your patch colour, and the eyes and nose stay dark so his face always reads clearly.
• Seven presets to start from: Original, Golden, Chocolate, Husky, Rose, Sky and Lilac. Or pick any colour you like.
• Nothing changes on your desktop until you save, and Reset to original brings back the coat he shipped with.
• Find it in Settings → Customize Mellow.

Auto update alerts

• See the version you are running in Settings → Updates.
• Check GitHub for a newer release whenever you want.
• Download and install it without leaving Mellow.
• Mellow reopens himself once the update is finished.
• When a new version appears, he mentions it once and then stays quiet about it.

A wider action catalog

• Mellow now finds your real Desktop, Documents, Pictures and Videos, including when OneDrive has moved them.
• Indexed items went from around 300 to over 1000.
• Drives use the names Explorer shows, so “open Local Disk C” works.
• Search a site directly: “search YouTube for lofi beats”, “look up python decorators on Wikipedia”.
• Shortened and abbreviated app names match far more reliably, so “open calc” just works.

What’s better

• Voice interaction: more reliable microphone handling, and smoother feedback while you hold the shortcut to speak.
• Screen aware dictation reaches every kind of text field now, including editors and terminals that do not report their own contents.
• Screen aware dictation has its own settings page explaining what it does, instead of sitting at the bottom of Advanced.

Updating from 1.1.0

• This is the first release that can update itself, so Mellow 1.1.0 cannot update to it automatically.
• Download the installer below and run it once. From this version onwards, new releases are offered to you inside Mellow.
• Installing over 1.1.0 keeps your settings, conversations, meeting notes and downloaded voice models.

Mellow v1.1.0

What’s new

Meeting transcription and notes

Mellow can now take notes while you focus on the conversation.

• Right-click Mellow to start transcribing your microphone and meeting audio.
• Pause, resume, or stop and save using the controls above him.
• Find saved transcripts in Settings → Meetings, with separate labels for You and Other participant.
• Generate structured notes covering key topics, decisions, action items, and open questions using your configured AI engine.
• Export transcripts or notes as Markdown, plain text, or JSON, or copy them to your clipboard.
• Search, rename, and delete saved meetings.
• Enjoy Mellow’s new glasses-and-notebook pose while he transcribes.

What’s better

• Cleaner transcripts: improved echo cancellation helps reduce duplicated speaker audio.
• More natural conversations: speech is grouped into readable turns, with improved ordering instead of fragmented, timestamp-heavy chunks.
• Clearer meeting controls: recording status, pause/resume, and stop-and-save actions stay within reach.
• Polished meeting pages: easier switching between transcripts and notes, with clearer generation progress and export options.
• Refined artwork: a detailed note-taking animation with colours that fit Mellow’s other poses.

Meeting labels distinguish your microphone from meeting audio—not individual remote speakers.

Mellow v1.0.0

Meet Mellow, a small pixel-art dog who lives on your Windows desktop.

Mellow is both a desktop companion and a hands-free assistant. Talk to him, ask questions, let him help around your computer, stay focused with timers and reminders, or keep him around simply as a pet.

What Mellow can do

• Listen through push-to-talk with Ctrl + Shift + Space
• Answer questions and speak responses aloud
• Look at your screen when you ask for visual help
• Point toward controls and objects using his bone pointer
• Open apps, folders, websites, and media
• Adjust the volume of individual applications
• Set reminders and notify you when they finish
• Run customizable Pomodoro focus sessions
• React when petted, dragged, woken up, or left idle
• Sleep, stretch, peek from screen edges, listen, think, and talk through pixel-art animations

Choose how Mellow works

Mellow supports several setups:

• Run language models locally through Ollama
• Use local Parakeet speech recognition
• Use local Kokoro text-to-speech with multiple voices
• Connect OpenAI, Anthropic, OpenRouter, Groq, NVIDIA NIM, ElevenLabs, or another compatible provider
• Use an existing Claude Code or Codex subscription
• Select pet-only mode with no AI or model downloads

Local speech models are downloaded once during onboarding and reused afterward.

Getting started

1. Download Mellow-Setup-1.0.0-x64.exe below.
2. Run the installer.
3. Complete the first-run onboarding.
4. Choose how Mellow should think, listen, and speak.
5. Hold Ctrl + Shift + Space and start talking.

System requirements

• Windows 10 or Windows 11
• 64-bit system
• Microphone access for voice input
• Internet access for cloud providers and initial local-model downloads
• Ollama installed separately if you want to run the answer model locally

Note

For the best visual experience, enable Animation effects in:

Windows Settings → Accessibility → Visual effects