← Back to all projects
🐚

My product Open source

OmniShell

Plain English → OS workflows, with human approval

A service that turns plain-English requests into operating-system workflows on Linux, macOS and Windows — with a human approval step before anything risky runs.

19 capability types6 LLM providers16 risk rules

Built for: Developers and IT teams who automate tasks across different operating systems.

↓ Scroll to explore the problem, architecture and stack

😩

The problem

Automating everyday system tasks means remembering commands for each OS, and letting an AI run commands directly is dangerous.

💡

Why it matters

OmniShell plans the steps with an LLM, scores them against risk rules, and waits for a person to approve anything risky — automation without the fear.

What you can do with it

01Describe a task in plain English and get a step-by-step plan.
02Run it on Linux, macOS or Windows (Bash or PowerShell).
03Approve or reject risky steps before they run.
04Track jobs reliably, even across restarts.

How it’s built

A simplified view of the architecture — follow the dots as a request moves down the layers.

You
Plain-English request
Approvalone-time token
Planner
FastAPI router19 capability types
LiteLLM plan6 providers · validated JSON
Scheduler
PostgreSQL queueSKIP LOCKED · leases
Executor
Risk rules16 checks
Bash / PowerShellverified with psutil

A request, step by step

  1. You describe a task; the router picks one of 19 capability types.
  2. One LiteLLM call returns a Pydantic-validated plan, falling back across providers if needed.
  3. The plan is scored against 16 risk rules; risky steps wait for an expiring approval token.
  4. Workers claim jobs with FOR UPDATE SKIP LOCKED and heartbeat leases, then run and verify each step.

Tech stack

Backend
PythonFastAPINestJS
AI
🔀LiteLLMGemini
Data
PostgreSQLRedis
Platform
Docker

What I built

  • A FastAPI service that routes requests across 19 capability types and makes one LiteLLM planning call returning Pydantic-validated JSON.
  • LLM fallback across 6 providers (Mistral, OpenRouter, Groq, Gemini, NVIDIA NIM, Hugging Face) with error classification and per-model cool-downs.
  • A PostgreSQL job scheduler using FOR UPDATE SKIP LOCKED claims, heartbeat leases and expiring SHA-256-hashed approval tokens.
  • A cross-platform executor (Bash and PowerShell) with 16 risk rules, syntax pre-checks, approval gates and process verification with psutil.