The AI desktop for people who build with AI.
A local-first autonomous agent for macOS. Watches your screen, listens to your meetings, remembers your work, spawns parallel sub-agents, and plugs into the MCP ecosystem (GitHub, Composio, Context7, HuggingFace).
Built on the patterns of Fazm — rebuilt local-only, with real sub-agents and production-grade RAG on top.
Alpha / heavy WIP. Phase 0 (foundation) is committed. Phase 1 (Node bridge) is active. See PLAN.md for the full build spec — every module designed before a line ships.
| Capability | Detail |
|---|---|
| Parallel sub-agents | Atomic SQL checkout, persona injection, depth & budget guards |
| Hybrid RAG | sqlite-vec + FTS5 in one SQLite file — auto-ingests code, papers, meetings, screen history |
| Knowledge graph | Live-updating on every tool call. Query with /kg <topic> |
| MCP ecosystem | Declarative ~/.shiro/mcp.yaml — Playwright, GitHub, Composio, Context7, HuggingFace, filesystem |
| Meeting mode | ScreenCaptureKit audio + Deepgram Nova-3 streaming STT — action items extracted live, approval-gated |
| Skills system | Claude Code-compatible .skill.md format — 60+ skills planned across 6 packs |
| Bring-your-own-model | Hot-swap LM Studio, Ollama, or Claude — local-first by default |
Swift + SwiftUI macOS native app (floating bar, task board, meeting overlay, KG viewer)
GRDB + sqlite-vec local memory, hybrid RAG, knowledge graph
Node 20 + TypeScript agent bridge, tool dispatch, ACP protocol
LM Studio / Ollama local inference (primary) — Claude as cloud fallback
Deepgram Nova-3 streaming speech-to-text for meeting mode
MCP servers external tool integrations
| Role | Model | VRAM |
|---|---|---|
| Brain | google/gemma-4-26b-a4b |
18 GB |
| Fast | qwen/qwen3-8b |
4.6 GB |
| Vision | qwen/qwen2.5-vl-7b |
6 GB |
| Embed | text-embedding-embeddinggemma-300m-qat |
0.2 GB |
All run simultaneously in LM Studio (~29 GB resident).
git clone https://github.com/abhisheksharma001/shiro.git
cd shiro
./setup.sh # installs Node bridge deps + MCP servers
open Shiro.xcworkspace # build & run in Xcode with ⌘RRequires macOS 14+ (Sonoma), Xcode 15+, and LM Studio or Ollama running locally.
- Phase 0 — project scaffold, agent coordinator, bridge architecture
- Phase 1 — Node/ACP bridge fork, LM Studio proxy
- Phase 2 — hybrid RAG + knowledge graph
- Phase 3 — MCP registry + bundled servers
- Phase 4 — meeting mode (audio + STT + action items)
- Phase 5 — skills system + slash commands
- Fazm — Swift↔Node ACP bridge pattern and SkillInstaller inspiration
- Paperclip (Mediar-AI) — task state machine + atomic checkout pattern
- Claude Code — skill format, slash commands, hooks engine
MIT — see LICENSE.
Built in the open by @abhisheksharma001