A study and research companion for Windows. Not a chatbot with a face bolted on — a character who sits beside you while you read, revise and write, keeps every conversation separate, and remembers what you are working on.
She ships with no model weights: you point her at a provider, or at Ollama on
localhost, and she comes alive. Everything else — your chats, your skills, your
settings — stays in %APPDATA%\Yura as plain files.
- Explain it, then test it. Ask for the shortest true explanation, then ask her to quiz you on it. The Socratic tutor skill makes her teach by question instead of lecture.
- Digest what you are reading. Paste an abstract or a page; the Paper digest skill returns the claim, the method, what the evidence supports, and the strongest objection.
- Recall drills. Recall drill quizzes only from material you have already covered in that conversation, one question at a time, with a running score.
- Keep the thread. Every conversation is its own file: pin the ones you are mid-way through, rename them, delete them. She never mixes two subjects.
A skill is a markdown file with instructions. Write one in Settings → Skills,
or drop any .md you already have into %APPDATA%\Yura\skills\:
---
name: Paper digest
description: Summarise a paper as claims, method, evidence, limits.
enabled: true
---
When the user shares a paper, answer in this order:
1. The claim, in one sentence.
2. Method, three bullets maximum.
3. What the evidence actually supports.
4. The strongest objection a reviewer would raise.Enabled skills are compiled into her system prompt on every request. Toggle one off and it stops applying immediately — nothing is deleted, and nothing needs a restart. Three starter skills are installed on first run.
Off by default. Settings → Abilities: turn on Python, pick one workspace folder, and she can write and run scripts inside it.
- 20 second limit with a hard kill, output capped,
-Iisolation so yourPYTHONPATHcannot leak into the run. - Every path is checked twice — lexically for
.., then again after canonicalisation, so a symlink cannot walk out of the folder. - Every script, file write and notification asks first, in the chat, showing the exact code or path, with Approve and Deny. Sixty seconds of silence is a no, and the denial is reported back so she can adapt instead of hanging.
The protocol is plain text — a fenced ```yura:run_python block — which is
why it works with every provider, including local models with no tool-calling API.
She can also set timers, keep a task list, and send a notification through the same gate. Timers survive the window being closed and fire exactly once.
Settings → MCP adds a Model Context Protocol server — name, command,
arguments. Connect it and its tools appear in her hands: she calls
list_mcp_tools to see what is on offer, then call_mcp, and the arguments are
shown to you before anything runs.
The client is stdio JSON-RPC only, on purpose: that covers the servers people
actually use (npx @modelcontextprotocol/..., uvx ...) without pretending to
implement transports nobody asked for.
Under the conversation, always visible while she works:
- Context — tokens used against the model's window, with a bar that warns at 60% and turns red at 85%. You can see a long PDF eating your memory before it does.
- Tokens per second — measured over a sliding window, so a local model crawling at 4 tok/s looks like what it is.
- Thinking — if the model exposes a reasoning stream (DeepSeek-R1, QwQ, Claude extended thinking), her scratchpad appears as it arrives. Never spoken, never saved.
Ships with Sendagaya Shino — the VRM Consortium's own sample model, CC0, bundled and attributed. She breathes, blinks with a natural jitter, tracks your cursor with her eyes and a fraction of her head, changes expression mid-sentence from the emotion tags in her replies, and stands on a floor with a contact shadow rather than floating in a void.
Any VRM 0.x or 1.0 file works: drop it in ui/public/assets/models/ and pick it
in Settings → Avatar. With no model at all she appears as a designed spirit form.
Download from Releases:
| Build | File | What it does |
|---|---|---|
| Installer | Yura_0.2.0_x64-setup.exe |
Per-user install, Start-menu entry, uninstaller. 2.1 MB. |
| Portable | Yura_0.2.0_portable_x64.zip |
Unzip anywhere and run yura.exe. No install, no registry. |
Both need WebView2, which ships with Windows 11 and current Windows 10. The builds are unsigned, so SmartScreen warns once — More info → Run anyway.
Your data, all in %APPDATA%\Yura\:
settings.json provider, key, persona, voice, avatar, theme
chats/*.json one file per conversation
skills/*.md your skills
Delete that folder and Yura forgets everything.
OpenAI · Anthropic · DeepSeek · OpenRouter · Groq · Mistral · Together · xAI · Ollama · LM Studio · any OpenAI-compatible base URL. Provider, endpoint, key, model, temperature, reply length and memory depth all live in Settings, and the key never enters the webview.
A companion workload wants short turns and a fast first token, not a reasoning
model — keep max_tokens around 300–500. Details in
docs/providers.md.
| A conversation | Settings |
|---|---|
 |
 |
Needs Node 20+ and Rust. On Windows, the MSVC toolchain. On Linux, cargo-xwin
cross-compiles the whole thing, installer included.
npm install # task runner + Tauri CLI
npm --prefix ui install # renderer
npm run tauri:dev # hot-reloading dev window
npm test # 104 unit tests across the crates
npm run build:windows # NSIS installer, cross-compiledCross-compiling on Linux also needs lld llvm clang nsis, the
x86_64-pc-windows-msvc target, and cargo install --locked cargo-xwin. The
gotchas — the pre-build hook's working directory, where artifacts actually land,
why MSI needs a Windows host — are written down in
docs/architecture.md.
crates/yura-core persona · emotions · messages · skills · reply parser
crates/yura-llm LlmProvider trait · OpenAI-compatible · Anthropic · deltas
crates/yura-tools sandboxed workspace · python runner · timers · tasks
crates/yura-docs PDF text · magic-byte sniffing · prompt blocks
crates/yura-mcp stdio JSON-RPC client · tools only
crates/yura-avatar emotion → VRM expression weights · viseme tables
crates/yura-store chats · settings · skills · atomic writes
src-tauri the window: commands, event streaming, cancellation
ui/src/scene renderer, floor, contact shadow, bloom/grain/vignette
ui/src/avatar VRM rig · emotion mixer · idle · gaze · lip-sync
ui/src/chat conversation rail · transcript · composer · landing
ui/src/settings the settings page (categories → rows)
ui/src/telemetry context meter · tokens/s · thinking strip
Three rules hold it together: nothing in crates/ knows Tauri exists,
nothing in ui/src/scene knows what a message is, and the API key never
enters the webview. Adding a provider is one row in PRESETS; adding an
emotion is one enum variant, one expression row and one hue — and the tests fail
if those tables drift apart.
Shipped in 0.4: MCP servers, abilities with the approve/deny gate, the library folder, downloadable voice resources and tray/start-on-login. Next is wiring the microphone to the Whisper resource.
The design for each is written down in docs/plan-0.3.md.
MIT for the code — see LICENSE. The bundled avatar is CC0 and attributed in ATTRIBUTION.md; no voice models are bundled.