ShibaCatReal-time VRM avatar
A custom shiba-cat VRM model rendered with Three.js and anime-style MToon shading, with spring-bone physics on the ears and tail.
A living 3D cat for your Windows desktop
ShibaCat is a small transparent window with a real-time VRM cat inside. It watches what you are working on, follows your cursor with its eyes, reacts with expressions and moods, and talks with you by voice.
- Windows 10 / 11
- Electron + Three.js + VRM
- Runs locally
- No telemetry
A cat with opinions about your workday
Everything below ships in the current codebase — no roadmap hand-waving.
Eyes that follow you
Gaze tracking chases your cursor, with natural blinking, random glances and micro-saccades layered on top — powered by an 18-blendshape expression engine.
Workspace awareness
A local Windows sensor reads the foreground app and infers context — coding, building, terminal work, video, music, chat, meetings, gaming — and the cat adjusts mood and behavior to match.
Living idle behavior
Breathing, tail sway, ear twitches, folded paws, stretches, scratches, grooming, yawning, walking and running are all procedural — no canned animation loops.
Voice conversations
Streaming speech-to-text, spoken replies with lip-sync, and barge-in when you talk over it. Conversations run in Bahasa Indonesia by default, powered by your own provider API keys.
Companion modes
Co-working, hangout and busy modes, a homeostatic mood engine and self-initiated remarks — the cat reacts to your session, not just to commands.
It knows the difference between a build and a meeting
ShibaCat polls the active window on your machine — locally — and maps it to a behavior profile. Switch apps and the cat switches personality.
- Coding & building — focused poses while you work, surprise when a build kicks off
- Terminal work — alert posture for CLI sessions
- Video & music — relaxed watching, head-bob energy for tunes
- Chat & meetings — calm presence while you talk to humans
- Gaming — excited, playful mood
- Idle & away — sleepy cat, and a welcome-back greeting when you return
An agent mode that asks before it acts
ShibaCat can do more than react — it can run tasks on your PC. It is opt-in, tightly scoped, and every risky step needs your approval.
Scoped by default
File tools are locked to a workspace folder you choose. Path traversal, UNC network paths, protected system areas and symlink escapes are rejected before anything executes. Secrets found in file output are redacted automatically.
Approval-gated actions
Terminal commands are risk-classified. Anything above low risk pops an approval card on the cat — allow once, allow for the session, or deny. Critical operations always ask, every time.
Tools & MCP
Filesystem, terminal, a sandboxed browser window, web fetch/search, long-term memory, sub-agents — plus your own tools via MCP servers (stdio or HTTP).
Audited & bounded
Sessions are capped at 24 reasoning steps and 15 minutes, repeated identical tool calls are stopped automatically, and every tool execution is written to an audit trail.
Heads-up: agent mode is experimental. The full pipeline (tools, approvals, skills, MCP) is implemented and unit-tested, but it is currently driven through the developer interface — the polished end-user toggle is still on the workbench.
Local-first, no telemetry
- Settings, memory and API keys live in a local SQLite database on your machine
- API keys never leave your machine except to the AI provider you configured them for
- Window titles are processed locally for workspace detection
- No analytics, crash reporting or usage tracking of any kind
- Browser URLs are discarded by the workspace sensor by design
- Everything works without any API key — voice and chat simply stay off
Questions, honestly answered
Is ShibaCat free?
Yes, it is a hobby-scale project you can build and run yourself. The repository is not publicly available yet while licensing is being sorted out.
Does it work on macOS or Linux?
No. ShibaCat is Windows-only today — the workspace-awareness sensor and packaging pipeline are built specifically for Windows 10/11 (64-bit).
Does it need an internet connection or API keys?
The avatar, animations, eye tracking and workspace awareness run fully offline. Voice and chat call third-party AI providers (OpenAI-compatible endpoints, Groq, Deepgram) and need your own API keys — without them those features simply stay idle.
What language does it speak?
Voice conversations are in Bahasa Indonesia by default — the speech pipeline is tuned for it. The settings interface mixes Indonesian and English.
Will it slow down my PC or spy on my screen?
It is a lightweight transparent window, and it can sit in the tray, stay click-through, or be turned off entirely. It only reads the title of the foreground window — locally — and never records your screen, microphone (unless voice mode is on) or keystrokes.