What it does OpenMono is a terminal-native coding agent built in C#/.NET 10. One curl-bash installer sets up a Dockerized llama.cpp inference server, auto-detects your GPU/CPU/Apple Silicon, and drops you into a TUI or classic CLI. The agent reads your project files, plans, writes code, verifies with Roslyn or LSP, and loops up to 25 iterations per turn. No API keys, no cloud, no subscription.

The interesting bit The “doom-loop detection” is a nice touch: if the agent repeats the same tool sequence three times, it aborts. More structurally, the project treats inference as infrastructure rather than a service — it ships quantized Qwen models, runs them locally, and even supports splitting agent and inference across two machines via an outbound tunnel (no port forwarding). The 12-step tool pipeline with pre/post hooks and parallel read-only execution suggests someone actually thought about what happens when LLMs get tool-happy.

Key highlights

• Bundled llama.cpp with zero-config model selection (24 GB GPU → 27B dense, CPU → 35B MoE, Apple Silicon → Metal)
• 5 specialist sub-agents with locked tool sets and turn budgets (Explore, Plan, Coder, Verify, general-purpose)
• Docker sandbox: project mounts as /workspace, nothing else visible
• Deep code intelligence via Roslyn (C#) + lazy LSP for TS/Python/Go/Rust + optional semantic graph tools
• YAML playbooks with typed parameters, composable and resumable
• Vision support via @image.png with auto-resize for VRAM budget
• AGPL-3.0 license

Caveats

• Public beta, shipping fast — expect rough edges
• macOS less than 64 GB unified memory is “not encouraged” (falls back to smaller models, tight context windows)
• Linux requires Ubuntu 26.04 LTS or 25.10; Intel Macs limited to agent-only mode
• OpenAI/Anthropic/Ollama providers are marked WIP
• No Windows support listed

Verdict Worth a look if you have the hardware and want a fully local, auditable coding agent with actual sandboxing. Skip it if you’re on Windows, low-end Macs, or need the polish of mature cloud tools.