OpenMonoAgent.ai

2026-07-05 · source · source ↗ · raw capture · wiki page

OpenMonoAgent is a terminal-native coding agent powered by local LLMs. It runs entirely on the user's own hardware, requires no API keys, and offers unlimited tokens at no cost. The project is in beta, built by StartupHakk LLC.

Founding quote: "AI shouldn't be a subscription you rent. It should be infrastructure you own"

Architecture

Built on C#/.NET — "AI tooling should be infrastructure, not a subscription"
Embedded llama.cpp — two hardware paths: NVIDIA GPU runs Qwen3.6 27B; CPU fallback runs Qwen3.6 35B A3B
Docker-native sandboxing — agent mounts project into container, "can't leave the box"
Fully offline — no cloud dependency, no data leakage
Outbound-only dual-box mode for separating laptop client from home GPU rig

Features

Embedded inference, zero setup — hardware auto-detection, no config or API keys
TUI for long sessions — full-screen ANSI with streaming tokens, tok/s meter, context-window tracking, auto-compaction
Docker-sandboxed — permission gates on destructive operations
20 tools + MCP — file I/O, shell, search, web fetch, LSP, patches, sub-agents, plan mode
Built for .NET — Roslyn integration gives "real compiler intelligence: type hierarchies, call graphs, cross-assembly references"
LSP for C# and TypeScript — hover, go-to-definition, references
Playbooks — "typed, composable, stateful workflow automation" with step sequencing and gates
Dual-box mode — works behind NAT/CGNAT via relay
Persistent sessions — JSONL transcripts, cross-session memory, snapshot-based file undo

Performance Benchmarks (tok/s)

Hardware	Model	tok/s
Ryzen 9 7940HS (CPU)	Qwen3.6 35B A3B	~17–20
Apple M5 Pro (64GB)	Qwen3.6 35B A3B	~45–48
RTX 3090 (~$700 used)	Qwen3.6 27B	~42–45
RTX 4090	Qwen3.6 27B	~47–50
RTX 5090	Qwen3.6 27B	~75–80
RTX 3060 (12GB)	Qwen3.5 9B	~38–40

The site calls a used RTX 3090 "indistinguishable from a cloud API" and notes a NUC with 32GB RAM delivers "a solid 20 tok/s."

Installation

bash <(curl -fsSL https://raw.githubusercontent.com/StartupHakk/OpenMonoAgent.ai/refs/heads/main/get-openmono.sh)

Then run: openmono agent

Pricing

"Unlimited tokens. Forever." — Free, no per-token billing, no rate limits, no subscriptions.

Comparison

Positioned as the "third option" vs. Claude Code and OpenCode: fully offline privacy, Docker-native sandboxing (vs. host install for competitors), C#/.NET (vs. npm), free unlimited tokens (vs. per-token pricing).