Stop Being the Messenger Between Your AI Agents

When multiple agents are working, the most annoying part isn't that the AI isn't smart enough—it's that you have to relay messages between them. Copying and pasting, switching terminals, watching tasks run one after another.

Local Operator is a locally running AI agent hub built specifically to solve this problem. It runs on your own machine, and its target users are developers coordinating tool-oriented AI agents across multiple repositories. MIT licensed, open source.

Core Design

A few notable points:

Reuse your existing subscriptions. It doesn't make you buy another set of API keys. Log in with existing ChatGPT, Claude, Kimi, Grok, Z.AI, and Qwen accounts; multiple accounts are automatically pooled and load-balanced. For people who already subscribe to these services, it's effectively putting idle quota to work.

Roles are capability boundaries. An agent designated as reviewer can read code and run tests, but it has no tools for editing code. This restriction isn't enforced through prompt constraints—it's cut off at the tool level. All sub-agents it can delegate to inherit the same restriction.

Cross-session messaging. Two lop sessions you start, in different repositories with no parent-child relationship between them, can message each other directly. One tells another to pause a deployment, hand off a branch, or occupy a shared resource—no need for you to copy and paste between two terminals.

Close the terminal, and tasks keep running. Scheduled wake-ups still trigger after the terminal is closed. When an operation requiring approval comes up during unattended operation, it stops and waits for you to handle it from your phone or via /resume.

Terminal interface. It's not yet another web app. It's a full-screen TUI written in Textual, with everything done in the terminal.

The Real Hard Part of Multi-Agent Collaboration

Someone in the comments asked a practical question: the hardest part of managing agents isn't routing—it's shared state. When two agents touch the same repository, you need locks or separate worktrees; otherwise you wake up in the morning to a merge conflict nobody wanted.

Local Operator's answer is roles plus messaging. Role restrictions keep agents from touching anything outside their responsibilities; cross-session messaging lets agents coordinate themselves instead of relying on a human watching them. The image above shows two independent sessions negotiating a deployment slot: one claims occupancy and asks for objections, while the other checks its own work and gives the go-ahead.

Installation and Getting Started

One command, no need to install Python beforehand. macOS/Linux:

curl -LsSf https://raw.githubusercontent.com/damianvtran/local-operator/main/scripts/install.sh | sh

Windows PowerShell:

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/damianvtran/local-operator/main/scripts/install.ps1 | iex"

Then run lop to start and /login to log in. It supports OAuth login for subscription accounts such as ChatGPT, Claude, Kimi, and Grok, as well as local model servers like Ollama and LM Studio.

The security model defaults to approvals: read operations run automatically, while writing files and shell commands trigger a confirmation showing the exact command that will run. For full automation, you have to manually type /approvals auto—that action itself is an explicit choice.

Who It's For

This project treats 'existing subscriptions' as a core resource to manage, which is a different approach from most agent frameworks that require separate API keys. Cache-aware prompt layout, account rotation strategy, token-efficiency design—these are the things people who actually use agents to get work done every day care about.

It's not for demo users. It's for people who open their terminal in the morning and find five agents across three repositories waiting for them to coordinate.