# IronClaw

A local, privacy-focused personal AI assistant, an 'Agent OS' that keeps all data encrypted on the user's own machine and runs untrusted tools inside a WebAssembly sandbox with capability-based permissions.

IronClaw is a personal AI assistant built as an 'Agent OS' focused on privacy, security and extensibility. Its guided onboarding command lets a user choose an LLM provider, enter its API key and pick a model, then sets up local configuration, an encrypted credential store and a WebUI login token, starting a background service automatically on macOS and Linux.

Its design keeps all data stored locally and encrypted rather than sent to a vendor's cloud, is fully open source and auditable with no hidden telemetry, and lets new tools be built on the fly instead of waiting on a vendor update. Security is handled in several layers: untrusted tools run inside an isolated WebAssembly sandbox with capability-based permissions, secrets are injected only at the host boundary with leak detection rather than exposed to tools directly, prompt-injection defenses combine pattern detection, content sanitization and policy enforcement, and outbound HTTP requests are restricted to an explicit allow list of hosts and paths.

It is installed with a shell script and reachable through a REPL, HTTP webhooks or messaging-platform channels.

Website: https://www.ironclaw.com