OpenClaw

OpenClaw is a free and open-source autonomous AI agent framework created by Peter Steinberger and released in November 2025 (originally as Clawdbot, rebranded via Moltbot to OpenClaw after trademark complaints). Licensed under MIT and written in TypeScript/Swift, it runs entirely on the employee's device and connects to external large language models — Claude, GPT, Gemini, DeepSeek, or local models via Ollama — while using existing messaging platforms as its user interface. The architecture is local-first: configuration, session history, and memory files stay on-device. A skills system (SKILL.md directories) lets users extend the agent with tool integrations (GitHub, Notion, browser, shell, filesystem, scheduled tasks). Companion apps include a macOS menu bar utility and iOS/Android nodes with voice control. Agents can run in the main session (full host access) or be isolated in Docker/SSH sandboxes for untrusted channels. From an enterprise governance perspective, OpenClaw is the archetype of a "shadow AI agent": it installs via npm, bypasses CASB/DNS/proxy inspection, authenticates to corporate Slack, GitHub, and Notion via personal user credentials, and routes inbound data through the employee's personal LLM account — not the enterprise DPA. Traditional SSO logs, software allowlists, and network monitoring do not see it.

Storage

Local device — config, session history, skills, and memory files stored on the user's machine. No OpenClaw-operated cloud by default; optional OpenClaw Cloud tier is hosted separately.

Retention

Persistent on-device until manually deleted; no automatic expiry. Retention of prompt data at the LLM provider depends on which provider the user configures.

Training on inputs

OpenClaw itself does not train on user data. Inputs are forwarded to the user's configured LLM (OpenAI, Anthropic, DeepSeek, Gemini, local Ollama, etc.) under that provider's own training terms.