agentlas-ai/Agentlas-OS

Agent OS: keep specialist agents in a hub, spin up a temporary orchestrator per task. Local-first, works with any model.

What it solves

Agentlas OS addresses the problem of "trapped" AI agents—those tied to a single chat session, a specific model provider's workspace, or a single computer. It transforms AI agents from temporary prompt-based settings into portable, owner-scoped assets that can be moved across different LLM hosts and environments.

How it works

At its core is Hephaestus, an open-source engine that acts as a local-first runtime. It separates the agent lifecycle into three distinct operations:

  • Build: Converts plain-language requests into runnable agent or team packages containing roles, tools, memory boundaries, and verification contracts.
  • Borrow: Allows users to integrate public specialists from the Agentlas Hub into their local runtime without copying private source work.
  • Own: Stores created agents in a private Agent Cloud, enabling them to be retrieved and run on any supported host where the user is signed in.

The system functions like an operating system for agents, providing a kernel for routing and security, a process scheduler for parallel execution, and a memory management unit (MMU) to isolate local project memory from durable promotions.

Who it’s for

It is designed for users who want to manage a team of AI agents as permanent assets across multiple LLM tools (such as Claude Code, Cursor, Gemini, and DeepSeek) and local environments, rather than relying on isolated chat sessions.

Highlights

  • Portable Agent Packages: Agents are compiled into manifests and contracts that can be moved between different LLM hosts.
  • Briefing Interview Engine: A specialized system that interviews the user to resolve ambiguity in prompts before an agent is built, ensuring higher reliability.
  • Cross-Host Compatibility: Works as an adapter layer across various LLM interfaces including Claude Code, Codex, Gemini, and Cursor.
  • Multi-Agent Orchestration: Supports combining local agents and Hub specialists into coordinated teams with explicit routing and handoffs.
  • Local-First Execution: While packages are stored in the cloud, the actual execution happens on the user's local host under their own permissions and API keys.