IronClaw is an open-source AI agent runtime built by NEAR AI, the AI arm of NEAR Protocol, positioned as a security-hardened alternative to the OpenClaw agent framework. NEAR AI launched IronClaw on February 25, 2026 at the NEARCON conference, rebuilding the runtime from scratch in Rust with cryptographic security as a core design goal rather than an add-on. NEAR co-founder and Transformer paper co-author Illia Polosukhin is listed among the project's contributors.
IronClaw's central idea is that AI agents should be able to act on a user's behalf without the underlying model ever touching raw credentials: API keys, tokens, and passwords sit in an encrypted vault and get injected directly at the tool boundary, never passed through the language model itself. Every tool an agent calls runs inside its own WebAssembly sandbox with no filesystem access and no outbound network connection beyond an explicit allowlist, and the whole runtime can run inside a NEAR AI Cloud trusted execution environment or be downloaded from GitHub and self-hosted.
IronClaw is aimed at developers and teams building always-on AI agents that need to touch real credentials and take real actions, such as background automation, scheduled tasks, and multi-channel bots, without accepting the security risk of exposing secrets to the model. A free Starter tier includes one hosted agent instance with a small inference credit allowance; paid tiers begin at $20 a month and scale to $200 a month for teams running more agents.
