Agentlas OS provides a local-first framework for building, owning, and orchestrating AI agents, designed for users who need robust agent management beyond single-session LLM interactions. It addresses the challenge of agent portability and team collaboration by separating agent creation, ownership, and execution, allowing agents to function as persistent assets across different models and environments. The system targets power users and developers seeking to integrate and manage AI agents as a core part of their workflow.

How It Works

Agentlas OS is powered by the open-source Hephaestus engine, which acts as a local-first runtime. It enables users to define agents from plain-language requests, package them with explicit contracts (roles, tools, memory, verification), and manage them through a private "Agent Cloud" for ownership and portability. Agents can be borrowed from a public "Agentlas Hub" or orchestrated into teams. Execution occurs on the user's chosen host, leveraging existing LLM integrations (Claude, Codex, Gemini, etc.) and local resources, ensuring data privacy and control.

Quick Start & Requirements

The primary installation method involves pasting a curl command directly into an LLM's code interpreter or shell environment: curl -fsSL https://raw.githubusercontent.com/agentlas-ai/Agentlas-OS/main/scripts/install-all-runtimes.sh | HEPHAESTUS_INSTALL_GLOBAL_ROUTER=1 bash This installs the Hephaestus engine and command adapters. Manual installation via cloning the repository and copying files is also supported. Prerequisites include a supported LLM host (e.g., Claude Code, Codex, Gemini CLI, Cursor) and Git.

Highlighted Details

• Briefing Interview Engine: Enhances prompt clarity by gating agent compilation based on quantitative ambiguity scores and lens-driven analysis, freezing resolved details into a work brief.
• Network 2.0 Routing: A sophisticated, local-first scheduler that manages agent dispatch across local, cloud, and hub sources with fallback mechanisms and explicit scope handling.
• Stormbreaker Execution Gating: Ensures agent operations are deterministically verified through a Kernel Gating Envelope, preventing autonomous completion until all outcomes are checked.
• Ontology Runtime: A local-first knowledge filesystem that ingests unstructured data into an agent-readable SQLite database using CJK trigram tokenization and Reciprocal Rank Fusion, supporting GraphRAG without external API calls.
• Agent Packaging: Agents are compiled into portable packages with detailed contracts (routing cards, memory maps, agent ontologies) enabling auditability and cross-runtime execution.

Maintenance & Community

The README does not explicitly detail community channels (like Discord/Slack), specific maintainers, sponsorships, or a public roadmap. Contributions are expected via pull requests, with a requirement to pass verification test suites.

Licensing & Compatibility

The project is licensed under the Apache-2.0 license. This license generally permits commercial use and integration into closed-source projects, with standard attribution and copyleft provisions for derivative works.

Limitations & Caveats

Agentlas OS is fundamentally local-first; Agent Cloud and Hub are package stores, not hosted LLM execution environments. Publicly published packages have strict safety boundaries, excluding sensitive data like API keys or raw transcripts. The system requires specific LLM host integrations and may involve a learning curve due to its comprehensive agent operating system architecture.