Agentlas OS

Agentlas OS is a local-first AI agent framework that assembles specialist teams around a task. It runs inside Claude Code, Codex, Cursor and Gemini and is open source under Apache 2.0. Its roles cover software development and work such as insurance analysis, M&A diligence and travel planning.

The framework splits a brief into tasks with defined roles, handoffs and expected outputs. It finds candidate agents on your machine, in your private Agent Cloud packages and in the public Agentlas Hub. Your chosen model selects the specialists and records its reasons. Project context and memory accompany each assignment, so agents receive information about the work they're taking on.

Verification is a separate stage. An independent verifier checks the work before the system marks it complete, with records of execution and handoffs. Agent releases are pinned to their exact contents, and external agents need signed identity verification as well as matching capabilities.

You supply the model access. Agent Cloud stores and retrieves agent packages; the selected model and current host runtime execute them. Credentials, local files and machine permissions don't travel with those packages. Project memory stays in a local database, while the chosen model provider determines where model requests go.

The related desktop app for macOS, Windows and Linux adds editable automation graphs and reusable Experience Chips for reviewed solutions. Tool-capable local models can browse pages, work with files and call connected tools; vision-capable models can control the screen within the access you allow.