██ ██ ███████ ██████ ██████ ██████ ██ ██ ███████ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ███████ ██ ██████ ██ ██ ██████ ███████ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ███████ ██ ██ ██████ ██████ ██ ██ ███████
Parallel agents, Bare-metal performance, No cloud dependencies
An ultra-lightweight, self-hosted, multi-agent AI operating system built in Rust. Run parallel coding loops locally, sandboxed, and 100% offline.
"Hiroshi's offline Ollama execution and local vector search give us full code privacy with under 5ms execution latency."
"Running parallel, background-active developer loops with whitelisted commands is an absolute lifesaver. We run tests overnight and review diffs."
"The Telegram integration and inquire setup wizard make it the most polished developer tool I've seen this year."
Setup.
Start the binary. The inquire onboarding wizard automatically detects Ollama ports and configures Telegram tokens.
Configure.
Declare agent prompts and whitelisted tools inside AGENTS.md. Tokio loops route agent actions dynamically.
Extend.
Drop custom scripts into ~/.hiroshi/skills/ or connect community JSON-RPC stdio MCP servers.
Does Hiroshi require an internet connection?
No. Hiroshi is 100% offline-capable, utilizing local Ollama server instances and native Rust vector computations.
Is execution secure?
Yes. Scripts execute strictly inside a path-sanitized workspace with whitelisted commands and a 10s timeout circuit breaker.
What is the resource footprint?
Hiroshi runs as a lightweight background daemon consuming less than 40 MB RAM with under 5ms execution loop latency.
Accessible from WhatsApp, Telegram, or any chat app you already use.
Download Hiroshi