Lens 0.1.5

Smaller downloads

The installer was 143MB. It is 100MB now, and on Windows it is 1MB.

Most of the old size was not Chromium, it was payload that had no reason to
ship: react and react-dom were runtime dependencies even though they are
already bundled into the renderer, and PDF reading pulled in 24MB of prebuilt
native canvas binaries whose only job is rasterising pages, which this app
never does. Reading PDFs through pdfjs directly drops that. Bundled
dependencies also carried their own sources, sourcemaps, docs and test
fixtures. Together: app.asar 75MB to 23MB.

Windows now uses a web installer: a 1MB stub that downloads the app while the
installer is open, instead of a 116MB file to download first.

Updating

The app offers this release on its own. On Windows it can install itself, and
the update download is now 1MB rather than the whole app. On macOS and Linux it
downloads the installer and opens it for you.

Not code-signed yet, so your machine will ask once on first launch.

Lens 0.1.4

Settings

Each provider in the answers list now carries its own mark, in its own colour:
Claude in Anthropic's clay, Groq in their orange, Gemini as a gradient. The
OpenAI mark is monochrome by design, so it follows the text colour and flips
with the light theme, and the local option takes sage, which is what the app
uses everywhere to mean the work happened on your own hardware. Five rows of
text with a status dot each meant finding Groq took reading rather than
glancing.

Screen capture

Capture failures are diagnosed and reported instead of guessed at, so a missing
Screen Recording permission says so rather than producing an answer about
nothing.

Linux

The AppImage, deb and rpm are built on CI now rather than by hand, so they stop
falling a version behind. Audio is no longer offered where it cannot work: call
transcription needs ScreenCaptureKit, which is macOS only.

---

Updating: the app checks for this release on its own and will offer it. On
Windows it can install itself. On macOS and Linux it downloads the installer and
opens it for you, since replacing a running app in place is not something to do
behind your back.

Not code-signed yet, so your machine will ask once on first launch.

Lens 0.1.3

If you are on 0.1.2, Lens will offer this to you itself. Earlier versions need one manual install, after which updating is automatic.

New

• Slash commands. Type / in the message box for /screen, /summarise, /translate, /reply, /export, /new, /model, /docs, /web and more. Arrow keys to choose, Enter to run.
• Export a chat as Markdown, plain text or JSON.
• Rate an answer. Thumbs up or down under any answer. Not a statistic: a rejected answer is never reused as context again, and an approved one is preferred.
• Light theme, plus an option to follow your system.
• A thinking indicator that says what it is doing. The word types itself out, and while a reasoning model thinks you see elapsed time and the reasoning tokens spent so far. When the answer arrives it reports how long it thought.
• API keys in Settings for Claude, Gemini, OpenAI or Groq, at any time rather than only during first run. Keys are encrypted with your OS keychain.
• Model server address in Settings, so Lens can use another computer's GPU.
• Shortcuts reference in Settings, with the right keys for your platform.
• More models suggested, fourteen now, every one able to read a screen and matched to your hardware.

Fixed

• Answers were printed as raw Markdown, so **bold** showed its asterisks.
• Documents can be removed one at a time.
• A model that is not installed now names the server and lists what is available, instead of a bare 404.
• A missing installer no longer leaves you on the release page with no explanation.

Notes

Not signed with a paid certificate, so on first install: macOS right-click the app and choose Open; Windows choose More info then Run anyway.

Lens 0.1.2

The app can now update itself, and answers render properly.

Updating

Lens checks for new versions on launch and offers them in the panel. Press Update now and it downloads the installer itself, with progress. On Windows it installs and restarts. On macOS it opens the disk image for you to drag across, because replacing a running app needs a signing certificate this build does not have.

Nothing downloads or installs without you choosing it.

New

• Slash commands. Type / in the message box for /screen, /summarise, /translate, /reply, /export, /new, /model, /docs, /web and more. Arrow keys to choose, Enter to run.
• Export a chat as Markdown, plain text or JSON.
• Rate an answer. A thumbs up or down under any answer. This is not a statistic: a rejected answer is never reused as context again, and an approved one is preferred.
• Light theme, plus an option to follow your system.
• API keys in Settings. Add or change a key for Claude, Gemini, OpenAI or Groq at any time, not only during first run. Keys are encrypted with your OS keychain.
• Model server address in Settings, so you can point Lens at another computer's GPU.
• Shortcuts reference in Settings, showing the right keys for your platform.

Fixed

• Answers were printed as raw markdown, so **bold** appeared with the asterisks.
• Documents can be removed one at a time rather than only all at once.
• A model that is not installed now names the server and lists what is available, instead of a bare 404.

Notes

Not signed with a paid certificate, so on first run: macOS right-click the app and choose Open; Windows choose More info then Run anyway.

Lens 0.1.1

Fixes for the first release. If you installed 0.1.0, replace it with this.

Fixed

• Windows showed macOS shortcuts. The app told Windows users to press ⌘, a key their keyboard does not have. Shortcuts now render for the platform you are on.
• "model not found" with no way to fix it. A packaged app could only ever reach localhost, so a model hosted on another computer was unreachable. Settings → Model server now lets you point Lens at any machine, and a missing model names the server and lists what is actually installed on it.
• The close button was missing. Hiding and quitting are different actions; both are back in the title bar.

Added

• Settings → Shortcuts, a reference showing the keys for your platform. If another app already owns the capture shortcut, Lens registers the next one available and the list shows what was actually registered.
• Switching model now says it is loading, since only one model fits on most graphics cards and a silent wait looked like a hang.

Notes

Builds carry the author's name in their metadata and are ad-hoc signed. They are not signed with a paid certificate, so:

• macOS: right-click the app, choose Open, then Open again.
• Windows: on the SmartScreen prompt, choose More info then Run anyway.

Lens 0.1.0

The first release of Lens: a screen and audio aware AI assistant that runs on your own machine.

Downloads

Platform	File
macOS (Apple Silicon)	Lens-0.1.0-arm64.dmg
macOS (Intel)	Lens-0.1.0.dmg
Windows	Lens-Setup-0.1.0.exe

These builds are not code-signed yet, so the OS will warn you the first time.

• macOS: right-click the app and choose Open, then Open again.
• Windows: on the SmartScreen prompt, choose More info then Run anyway.

What it does

• Ask about your screen with ⌘⇧Space. It attaches the capture and asks what you want done with it, rather than guessing.
• Add your own documents (PDF, Word, text, markdown). Only the passages relevant to each question are used, so a large library still works with a small model.
• Turn on listening and it transcribes a call on-device, then answers the last thing the other person said.
• Chats are kept in a searchable sidebar, and it recalls relevant past exchanges.
• Optional web search, off by default.

Running it for free

Install Ollama and Lens will recommend a model that fits your hardware, then download it for you. Nothing leaves your machine.

If you would rather use a key you already own, Anthropic, Gemini, OpenAI and Groq all work. Keys are encrypted with your OS keychain.

Using another machine's GPU: set OLLAMA_HOST=0.0.0.0:11434 on the machine with the GPU, open port 11434, and point Lens at http://<that-ip>:11434 in Settings.

Known limits

• The audio helper is macOS only. It needs Screen Recording, Microphone, and Speech Recognition permission, and Dictation enabled in System Settings.
• A local model is slower than a frontier cloud model, and reasoning models such as qwen3-vl trade speed for depth. Both are switchable in the dropdown.