A little robot that sits on your desktop and keeps an eye on your work.
Pixy answers Claude Code's permission prompts, chats with a language model running on your own machine, remembers what you talked about last week, and tells you when the build breaks or mail arrives. Then it goes quiet until it has something to say.
Everything runs on your machine. No account, no telemetry, no server of ours — there isn't one. Your chats, notes and search index never leave the disk they were written to.
Status: used daily by its author on one machine. Windows only for now. Interfaces are not stable yet.
Claude Code asks permission before it runs a command or edits a file. Normally that prompt waits in whichever terminal you started it in. Pixy puts it on top of everything instead, so you can approve from wherever you are — and the answer goes straight back to Claude.
The real command, file path and diff are on the card. Several sessions can be waiting at once; they stack. When Claude asks a multiple-choice question, it becomes tappable chips.
 An edit, with the diff it is about to make |
 A question, as chips |
Point it at whatever you run locally. Streaming replies, syntax highlighting,
several named models you can switch between mid-thought. Attach documents
(pdf docx pptx xlsx csv md txt json html) or images. Edit a
message and send it again, or retry a single reply.
Ask "what did we decide about caching?" and it finds the conversation that said cache — searched across everything you have ever talked about, and quietly added to your new question. Add an embedding server and it will also find the one that made the same point in different words.
Unread mail with a one-line summary written by your local model. A morning brief. A nudge before a meeting. Merged pull requests, failing CI, and issues that just landed on you. All as notices that appear and go away — not another tab to check.
 Summarised locally — not the first line of the mail |
 Before a meeting starts |
 The morning brief |
 Something landed on you |
 The daily digest |
|
An on-device wake word, then transcription and speech through local servers. Nothing is recorded or sent anywhere before the wake word fires.
Text-to-image through a local server, started when you ask and shut down when you stop. A markdown notes pane with wiki-links, pinning and a focus mode — real files in a real folder, editable by anything. A graph of what Claude Code chose to remember about your projects. Music controls, per-app volume, and enough awareness of whether you're on a call or away to stay out of the way.
 Notes |
 Memory |
Pixy is a small pill that floats above your other windows. It changes shape and colour with what it is doing:
| Gesture | What happens |
|---|---|
| Drag | Move it anywhere on screen |
| Single click | Opens the music panel |
| Double click | Opens a terminal |
| Right click | Quick menu: Workspace, Settings, Hide |
| Esc | Closes the quick menu |
Clicks pass straight through to whatever is underneath unless the cursor is actually on Pixy, so it can sit on top of your editor without getting in the way.
 Right click |
 Single click |
When a card appears: press Approve or Deny and the answer goes back to Claude Code. × dismisses it, which counts as a deny. Chips plus Submit for a question. Anything else — mail, meetings, GitHub — you can click through to open.
The tray icon has everything the mascot can't reach: show and hide, open the workspace or settings, open a terminal, run the GitHub digest or morning brief on the spot, and quit.
1. Install it. Download the installer from Releases and run it. Windows will warn you that it's unsigned — choose More info → Run anyway. It needs Windows 10 or 11.
2. Open Settings from the tray icon and fill in whatever you want to use. Nothing needs to be set up in a particular order, and Pixy starts fine with none of it — every feature simply stays quiet until it's configured.
3. For the Claude Code cards, paste a few lines of hook configuration into
.claude/settings.json. This is the one step that can't be done from the
Settings window.
→ SETUP.md walks through all of it, one feature at a time: the local model, the hooks, voice, pictures, Gmail and Calendar, GitHub, and what to check when something doesn't work.
Prefer to build it yourself? See CONTRIBUTING.md.
Pixy reads web pages a language model chose, runs a local HTTP server, and holds OAuth tokens. That deserves a straight answer rather than a reassurance, so it has its own document: SECURITY.md.
The short version:
- Chat history, notes, memory and the search index never leave the machine. Remembered history isn't sent to a hosted model unless you turn that on.
- You supply your own OAuth client and tokens. There is no shared application and no secret compiled into the binary — a desktop app can't keep one.
- Secrets are stored in plaintext on your machine. A deliberate trade-off for a local-first tool, explained rather than hidden.
- Web content is treated as hostile input.
| SETUP.md | How to use it — every feature, step by step |
| SECURITY.md | What it trusts, what it stores, how to report a vulnerability |
| CONTRIBUTING.md | Building it, where the code lives, filing an issue |
| docs/DESIGN-NOTES.md | Why it is built the way it is |
Built with Tauri 2. Windows only today — porting is about four files, and the notes are in CONTRIBUTING.
MIT — see LICENSE. Third-party components redistributed here are listed in THIRD-PARTY-NOTICES.md.