Iron Jarvis
Your own local-first AI operating system.
A desktop AI workspace that runs on your own computer. One chat box answers quick questions in seconds and quietly hands bigger jobs to a team of AI agents that plan, build and review the work. Every action is logged, risky steps ask first, and code changes wait for your approval before they merge.
- GitHub stars
- 59★
- Versions shipped
- 398
- Automated tests
- 10,800+
- Commits
- 584

Iron Jarvis chat workspace with model picker, project threads and composer
Two ways to read it
In plain English
A desktop AI workspace that runs on your own computer. One chat box answers quick questions in seconds and quietly hands bigger jobs to a team of AI agents that plan, build and review the work. Every action is logged, risky steps ask first, and code changes wait for your approval before they merge.
For engineers
A FastAPI daemon (Python 3.12, SQLite) plus a Next.js 15 dashboard inside an Electron app with tray, global hotkeys, auto-update and a Windows installer, and a Chrome/Edge browser add-on. Inside: a supervisor/subagent runtime in git worktrees, a fail-closed permission engine, a provider router with typed errors and a circuit breaker, budgeted context planning, an MCP client and server, visual workflows, a ConPTY terminal workspace and offline speech-to-text.
What it does.
- One chat surface that escalates itself into a full tool-using agent
- Supervisor + specialist subagents in sandboxed git worktrees, with human-approved merges
- Fail-closed permission engine: agents can only call the tools they were armed with
- Four-layer memory rendered as an orbitable 3D constellation
- Visual workflow canvas, schedules, webhooks and two-way Telegram/Slack chat
- Browser add-on, PII redaction, Office documents in and out, offline voice
The story
Why it exists
Cloud AI assistants are brilliant and opaque. Iron Jarvis is the opposite bet: an assistant that lives on your machine, explains every step, keeps its secrets encrypted on your disk and never silently ships your data to a cloud model when a local one is down. For an accounting practice handling client data, that rule is not a preference — it is the whole point.
How it works
A Python daemon owns the agents, memory, tools and permissions; a Next.js dashboard wrapped in Electron is the face. Quick questions are answered directly. Bigger asks are planned by a supervisor, split across specialist subagents that each work in their own git worktree, reviewed, and only merged when you approve the diff.
Shipped like a product
Version 1.300.0 means three hundred numbered releases, each one gated by a full test suite in CI and published as a Windows installer. The codebase holds roughly 176,000 lines of Python and 106,000 lines of TypeScript, guarded by more than 10,800 automated tests.
Built with
- Python 3.12
- FastAPI
- SQLite
- Next.js 15
- React
- TypeScript
- Electron
- MCP
- WebGL
- React Flow
- xterm.js
Status: Shipped · started June 2026 · open on GitHub
Keep exploring
- AI agents & orchestrationExperiment
RealDeal CPA Agent
An AI accounting employee, not a chatbot.
117Skills
- AI agents & orchestrationShipped
BetaBot
An AI beta tester that files real bug reports.
70Tests passing
- AI agents & orchestrationShipped
AI-Powered RPA
Plain English in, desktop automation out.
186Tests passing