NEAR AI has released IronClaw 1.0, presenting the update as a ground-up reconstruction of its personal AI assistant rather than an incremental software upgrade. The stable version was published on July 27, 2026, with NEAR describing it as faster, more reliable and more consistent across different user interfaces.
The assistant is available through the command line, web, Slack and Telegram, with those channels designed to use the same memory, operational rules and security controls. NEAR said new IronClaw agents can be deployed immediately, although direct upgrades for previously created instances were not available at launch.
Rebuild Separates Decisions From Agent Actions
IronClaw’s revised architecture separates the component that decides what to do from the systems that execute actions. A coordination layer described by NEAR as a “guard” sits between them, routing operations through a common control point. The design is intended to apply the same approval and security policies across every supported channel, instead of implementing separate safeguards for each interface.
The release introduces persistent workspace memory, scheduled automations, subagents, trace capture and checkpointing intended to preserve progress when a task is interrupted. IronClaw can also create sandboxed tools when it encounters a capability gap and reuse them in later workflows. Sensitive actions require explicit approval, while credentials are designed to be issued only when needed and removed from logs and reports, according to NEAR AI.
NEAR AI also said IronClaw 1.0 leads PinchBench, ClawBench and OfficeQA when tested with DeepSeek V4 Flash as the common underlying model. The company reported scores of 93.5% on PinchBench, 88.6% on ClawBench and 76.4% on OfficeQA. Those results are claims published by NEAR AI using its own benchmark framework and comparisons, rather than findings from an external audit or independent evaluation body.
PinchBench evaluates agents across 147 tasks involving activities such as email management, scheduling, research, coding and file operations. ClawBench focuses on multi-step interactions with live websites, while OfficeQA measures document retrieval and numerical reasoning using historical US Treasury materials. The benchmarks test different aspects of agent execution, but leaderboard performance does not by itself establish reliability in every production environment.
Existing Installations Face a Separate Upgrade Process
The public GitHub release notes confirm that version 1.0 replaces the earlier runtime, storage system, extension host and web interface. The release includes installers for macOS, Linux and Windows, support for multiple model providers and built-in extensions for services including GitHub, Gmail, Google Workspace, Slack and Telegram. GitHub’s technical documentation provides independent confirmation that a stable 1.0 software package was released, beyond NEAR’s promotional posts on social media.
However, the release notes state that IronClaw 1.0 is not an in-place upgrade from the previous 0.29.x version. Existing configuration files, databases, settings and secrets are not automatically migrated, and self-hosted users are instructed to treat version 1.0 as a fresh installation. That technical limitation adds important context to NEAR’s statement that existing accounts will receive upgrade access later.
The software also launched with several documented limitations. Some command-line functions appear in the help menu but are not yet implemented, certain earlier commands have no direct equivalent and the history-import option does not currently perform a migration. The stable designation therefore marks the release of the rebuilt architecture, not the completion of every planned feature.
IronClaw strengthens NEAR’s effort to position AI agents, persistent memory and confidential execution as part of its broader infrastructure strategy. For now, however, the evidence primarily establishes that the redesigned runtime is publicly available and has produced strong results in NEAR’s published testing. Broader adoption, sustained operational performance and the practical value of its benchmark advantage remain to be demonstrated through external usage.