Shelbi launches open-source orchestrator for terminal agents

Managing agent workflows

Managing multiple AI agents can often feel like babysitting several different terminal tabs at once. Developers frequently struggle to keep track of which agent is stalled, which one needs human attention, and which tasks are still pending. Shelbi aims to solve this by providing a personal orchestrator for the terminal.

The system is designed to function like 'Inbox Zero' for agent-based work. Instead of letting an agent get derailed by a complex request, the orchestrator agent takes an initial idea and writes it up in a standard format. This ensures that nothing small or large gets dropped during the development process.

By using an orchestrator that does not perform the actual work itself, the system maintains focus. Every idea becomes a scoped task before a worker agent ever touches it. This structure allows for small changes to flow through the system quickly while keeping the overall project organized.

Specialized worker agents

Shelbi relies on the principle that agents do their best work in focused chunks. Rather than asking one AI to do everything, the system assigns specific roles to different workers. One agent might build the core code, while others are dedicated to scrutinizing that work for specific flaws.

The platform ships with three specialized review agents: QA, Security, and Adversarial Review. These agents are designed to do one job well, providing a gate for every task. Users can tailor these agents with custom instructions or author their own specialized skills to fit specific project needs.

The Adversarial Review agent is particularly distinct in its instructions. It is told to assume there is a bug until it can prove otherwise, specifically looking for defects that the developer and the tests missed. It focuses on correctness, security, error handling, and edge cases rather than simple style issues.

Workflow boundaries and gates

Autonomy in AI coding is made safer by creating clear boundaries. Shelbi uses YAML files to define workflows, which dictate the stages a task must move through before reaching a human. These gates ensure that every piece of code meets a certain standard before it is merged into the main branch.

For example, a feature workflow might move from a backlog to an in-progress state, then through adversarial review, QA, and security gates. Only after these automated checks are passed does the task land in a review column for a human to sign off on. This prevents humans from having to manually check every minor detail of the code.

Users can also enable 'shelbi zen' to automate the process further. When this mode is active, the orchestrator automatically merges work that passes all the gates and reports back on what landed. This allows the human user to focus only on the items that actually require manual intervention.

Terminal-native features

The tool is built to be a native experience for developers who live in the command line. It includes a Kanban board TUI (Text User Interface) that displays every task as a card. This keeps the status of the project visible on the screen at all times rather than hidden in a database or a web UI.

One of the most flexible features is the ability to run workers on any machine. Any box that can be accessed via SSH and has git and an agent CLI can serve as a worker. Each worker runs in its own Shelbi session, allowing users to attach to any terminal to watch an agent work or type to it directly.

The system avoids the need for complex cloud setups or external databases. Because tasks are markdown files and workflows are YAML, everything is stored directly in the repository. This means the state of the project can be read, grepped, and committed just like any other part of the codebase.

Open source and installation

Shelbi is released as an open-source project under the MIT license. This allows users to inspect every line of code that runs on their machines, which is a critical factor for many developers who are wary of how AI agents interact with their local files. The project is hosted on GitHub for public review and contribution.

The software is distributed as a single binary, making it easy to deploy across different environments. It is designed to be lightweight and accessible for anyone who wants to move beyond simple chat interfaces into structured agentic workflows. The goal is to provide a reliable way to manage complex coding tasks without the overhead of heavy enterprise platforms.

Developers can get started by installing the tool via Homebrew. The project provides specific commands for installing the main binary as well as a desktop version. By providing these options, the creator ensures that the tool can be integrated into various types of development environments quickly.