Frozen Fish
Local-first AI orchestration.
MCP-native. GPU-aware.
Frozen Fish orchestrates the LLM backends you already run on your own hardware. It routes every prompt with a live GPU-aware scoring router, exposes 12 sandboxed MCP servers to your agents, and serves context through a three-layer engine (CAG → CAR → RAG) with a YAML workflow engine, RBAC, a PII-redacted audit log, and an egress allowlist on every outbound call.
Your data never leaves the box.
12 MCP Servers
- filesystem
- sqlite
- web_scraper
- git
GPU-Aware Router
Three-Layer Context Engine
- CAGPrecomputed KV
- CARTool partitions
- RAGRetrieval fallback
LLM Backends
- Ollama
- vLLM
- LM Studio
- llama.cpp
- Transformers
Security First
- RBAC
- PII-redacted audit log
- Egress allowlist
- Sandboxed tools
$ ff orchestrator start [✓] Discovering backends... [✓] Loading 12 MCP servers... [✓] Initializing context engine... [✓] Starting dashboard on 127.0.0.1:8321 Frozen Fish is running.