Janua is a new fork of Apache Guacamole with comprehensive support for modern Wayland-based systems.
It is built for GNOME Remote Desktop and KDE KRdp compatibility:
- FreeRDP 3 support
- RDPSND version 8
- H.264/AVC codecs for RDPGFX (required by KDE KRdp)
A visual node-based programming environment for AI workflows, featuring a decentralized message bus architecture with a PostgreSQL data plane.
- Visual drag-and-drop module interface
- Decentralized message bus architecture
- Multi-provider LLM support (Anthropic, OpenAI, Mistral, Ollama)
- Agent team management with hierarchical roles
A management dashboard for LLM inference infrastructure, providing control over models, instance pools, GPU allocation, and distributed storage.
- Multi-model management (Mistral, DeepSeek, Qwen, Kimi, LLaMA)
- Speculative decoding with draft model support
- Instance pool orchestration with per-GPU assignment
- Speedloader for hot/cold model storage tiers
- RDMA-enabled distributed block storage with RAID aggregation
EagleBranch
llama.cpp fork with TurboQuant 3-bit KV cache compression and speculative decoding (EAGLE v1/v2/v3, NextN/MTP) for current, legacy, and CPU inference. Includes multiple bug fixes that enable legacy NVIDIA hardware and a fused rotated-domain matvec kernel ("Lane B") that closes most of the throughput gap to IQ4_NL on MLA models.
- TurboQuant 3-bit KV cache compression — 5× smaller than f16, +2.27% PPL on Mistral Small 24B
- EAGLE v1/v2/v3 speculative decoding for Mistral Large 3 (675B)
- NextN/MTP native speculative decoding for DeepSeek R1/V3 (671B)
- Pascal SM61 through Hopper SM90; CPU inference path on x86
- Full MLA (Multi-head Latent Attention) and MoE support
- MLA quantized-K bug fixes available under MIT on the public branch
Life management tools. A sovereign-cloud alternative to NextCloud — self-hosted or cloud-hosted, with zero footprint on your devices.
- Self-hosted with HTTPS reverse proxy and Tailscale — your data never leaves your infrastructure
- Web client works on any device including mobile — delete the bookmark and nothing exists on your phone
- 18 modules: tasks, email, calendar, recipes, finance, health, weather, knowledge base, and more
- Native Qt6/C++ cross-platform desktop client — Swift for Apple coming soon
Distributed multi-room audio routing in Rust. A software AV receiver that connects speaker nodes across a network with AI control via MCP.
- Route TTS, music, or system audio to any room, group, or broadcast everywhere
- Per-room volume and mute with live web dashboard
- STT microphone routing — select which room's mic feeds speech recognition
- AI control via MCP — Claude can manage routing, volume, and STT sources
- Rust server + Python nodes with system tray integration
Fast hybrid RAM/storage management for LLM models. Written in Rust. Keeps frequently-used models in RAM for instant access while storing others on cold storage.
- Multi-tier RAM storage — local DDR5, remote NVMe-oF over RDMA
- One-click promote/demote with automatic tier selection
- Fastest-fit placement with failover across pools
- Remote pool lifecycle — connect/disconnect with progress feedback
- Model pinning and async background transfers
- Support for Ollama, llama.cpp, vLLM, and HuggingFace
