agentlas-ai/Agentlas-OS

Agentlas OS – A portable “operating system” for LLM‑driven agents

What it is – Agentlas OS is an open‑source framework (the Hephaestus engine) that lets you build, borrow, and own autonomous AI agents or whole teams of agents. An agent here is a packaged method document that describes a workflow, the tools it needs, and the permissions it requires. The package can be run on any supported host (macOS, Windows, Linux) with any LLM you already have access to (Claude Code, Gemini, Ollama, etc.). Because the package is just data, you can store it in a private Agent Cloud and retrieve it on another machine, or pull public specialists from the Agentlas Hub.

Key concepts

• Build – Write a plain‑language request; Agentlas classifies it, runs an interview/research step, and emits a portable agent package (package‑contract.json, brief.json, etc.) that contains the method, input/output schemas, and routing contracts.
• Borrow – Search the public Hub for ready‑made specialist agents and import their runtime bundle without copying the original source code.
• Own – Upload your own packages to a private, owner‑scoped cloud so they can be restored on any signed‑in desktop.
• Runtime‑agnostic – The same package runs on Claude Code, Codex, Gemini, Antigravity, Cursor, or any local model you configure. The runtime supplies the capabilities (file system, web browsing, shell) while the LLM supplies the reasoning.
• Trust model – Packages are immutable contracts; verification scripts (verify-generated-package.sh) ensure every required artifact is present before the agent can be executed. Permissions and credentials stay on the host, never travel with the package.

How you get it

1. Install the OS with a single command (the README even provides a “paste‑to‑install” snippet that fetches install‑all‑runtimes.sh and writes files only under ~/.agentlas, ~/.local/bin, and LLM‑specific plugin folders).
2. After installation you get a command surface (/agentlas build, /agentlas hub, /agentlas upload, etc.) that can be invoked from supported LLM tools or the terminal.
3. Use the Agentlas Desktop UI (downloadable from agentlas.cloud/desktop) for a visual workflow: describe a task, watch the build pipeline, orchestrate teams, and run/verify locally.

Typical workflow

# 1. Describe a task in plain language (e.g. “generate a weekly sales report”)
/agentlas build

# 2. Review the generated package, optionally upload to your private cloud
/agentlas upload

# 3. On another machine, sign in and retrieve the package
/agentlas cloud

# 4. Execute it with the LLM of your choice; the host enforces its own permissions
/agentlas run

Why it matters

• Portability – Agents are not locked to a single chat or model vendor; they travel with you.
• Standardisation – By defining a clear JSON‑schema contract for inputs, outputs, and routing, Agentlas lets different graph engines or runtimes interoperate.
• Ownership – Private Agent Cloud gives you a version‑controlled, inspectable asset rather than a transient prompt.
• Extensibility – Public Hub specialists can be mixed with locally‑built agents, enabling rapid composition of complex workflows.

Current status

• Latest release badge shown in the README (semantic versioning, Apache‑2.0 license).
• Supports macOS (Apple silicon & Intel), Windows, and Linux.
• Works with a wide range of LLM providers (Claude Code, Codex, Gemini, Ollama, DeepSeek, etc.).
• Documentation is split into multiple languages and includes contracts, trust model, and detailed subsystem descriptions.

Bottom line – Agentlas OS is a genuine, open‑source platform for creating, sharing, and running LLM‑driven agents across machines and model providers, with a strong emphasis on portability, trust, and ownership.